null  null
Web Scraping with Python
Collecting Data from the Modern Web
Ryan Mitchell
Web Scraping with Python
by Ryan Mitchell
Copyright © 2015 Ryan Mitchell. All rights reserved.
Printed in the United States of America.
Published by O’Reilly Media, Inc., 1005 Gravenstein Highway North, Sebastopol,
CA 95472.
O’Reilly books may be purchased for educational, business, or sales promotional
use. Online editions are also available for most titles
( For more information, contact our
corporate/institutional sales department: 800-998-9938 or [email protected]
Editors: Simon St. Laurent and Allyson MacDonald
Production Editor: Shiny Kalapurakkel
Copyeditor: Jasmine Kwityn
Proofreader: Carla Thornton
Indexer: Lucie Haskins
Interior Designer: David Futato
Cover Designer: Karen Montgomery
Illustrator: Rebecca Demarest
June 2015: First Edition
Revision History for the First Edition
2015-06-10: First Release
See for release details.
The O’Reilly logo is a registered trademark of O’Reilly Media, Inc. Web Scraping
with Python, the cover image, and related trade dress are trademarks of O’Reilly
Media, Inc.
While the publisher and the author have used good faith efforts to ensure that the
information and instructions contained in this work are accurate, the publisher and
the author disclaim all responsibility for errors or omissions, including without
limitation responsibility for damages resulting from the use of or reliance on this
work. Use of the information and instructions contained in this work is at your own
risk. If any code samples or other technology this work contains or describes is
subject to open source licenses or the intellectual property rights of others, it is
your responsibility to ensure that your use thereof complies with such licenses
and/or rights.
To those who have not developed the skill, computer programming can seem like a
kind of magic. If programming is magic, then web scraping is wizardry; that is, the
application of magic for particularly impressive and useful — yet surprisingly
effortless — feats.
In fact, in my years as a software engineer, I’ve found that very few programming
practices capture the excitement of both programmers and laymen alike quite like
web scraping. The ability to write a simple bot that collects data and streams it
down a terminal or stores it in a database, while not difficult, never fails to provide
a certain thrill and sense of possibility, no matter how many times you might have
done it before.
It’s unfortunate that when I speak to other programmers about web scraping, there’s
a lot of misunderstanding and confusion about the practice. Some people aren’t sure
if it’s legal (it is), or how to handle the modern Web, with all its JavaScript,
multimedia, and cookies. Some get confused about the distinction between APIs
and web scrapers.
This book seeks to put an end to many of these common questions and
misconceptions about web scraping, while providing a comprehensive guide to
most common web-scraping tasks.
Beginning in Chapter 1, I’ll provide code samples periodically to demonstrate
concepts. These code samples are in the public domain, and can be used with or
without attribution (although acknowledgment is always appreciated). All code
samples also will be available on the website for viewing and downloading.
What Is Web Scraping?
The automated gathering of data from the Internet is nearly as old as the Internet
itself. Although web scraping is not a new term, in years past the practice has been
more commonly known as screen scraping, data mining, web harvesting, or
similar variations. General consensus today seems to favor web scraping, so that is
the term I’ll use throughout the book, although I will occasionally refer to the webscraping programs themselves as bots.
In theory, web scraping is the practice of gathering data through any means other
than a program interacting with an API (or, obviously, through a human using a web
browser). This is most commonly accomplished by writing an automated program
that queries a web server, requests data (usually in the form of the HTML and other
files that comprise web pages), and then parses that data to extract needed
In practice, web scraping encompasses a wide variety of programming techniques
and technologies, such as data analysis and information security. This book will
cover the basics of web scraping and crawling (Part I), and delve into some of the
advanced topics in Part II.
Why Web Scraping?
If the only way you access the Internet is through a browser, you’re missing out on a
huge range of possibilities. Although browsers are handy for executing JavaScript,
displaying images, and arranging objects in a more human-readable format (among
other things), web scrapers are excellent at gathering and processing large amounts
of data (among other things). Rather than viewing one page at a time through the
narrow window of a monitor, you can view databases spanning thousands or even
millions of pages at once.
In addition, web scrapers can go places that traditional search engines cannot. A
Google search for “cheapest flights to Boston” will result in a slew of
advertisements and popular flight search sites. Google only knows what these
websites say on their content pages, not the exact results of various queries entered
into a flight search application. However, a well-developed web scraper can chart
the cost of a flight to Boston over time, across a variety of websites, and tell you
the best time to buy your ticket.
You might be asking: “Isn’t data gathering what APIs are for?” (If you’re unfamiliar
with APIs, see Chapter 4.) Well, APIs can be fantastic, if you find one that suits
your purposes. They can provide a convenient stream of well-formatted data from
one server to another. You can find an API for many different types of data you
might want to use such as Twitter posts or Wikipedia pages. In general, it is
preferable to use an API (if one exists), rather than build a bot to get the same data.
However, there are several reasons why an API might not exist:
You are gathering data across a collection of sites that do not have a cohesive
The data you want is a fairly small, finite set that the webmaster did not think
warranted an API.
The source does not have the infrastructure or technical ability to create an API.
Even when an API does exist, request volume and rate limits, the types of data, or
the format of data that it provides might be insufficient for your purposes.
This is where web scraping steps in. With few exceptions, if you can view it in
your browser, you can access it via a Python script. If you can access it in a script,
you can store it in a database. And if you can store it in a database, you can do
virtually anything with that data.
There are obviously many extremely practical applications of having access to
nearly unlimited data: market forecasting, machine language translation, and even
medical diagnostics have benefited tremendously from the ability to retrieve and
analyze data from news sites, translated texts, and health forums, respectively.
Even in the art world, web scraping has opened up new frontiers for creation. The
2006 project “We Feel Fine” by Jonathan Harris and Sep Kamvar, scraped a
variety of English-language blog sites for phrases starting with “I feel” or “I am
feeling.” This led to a popular data visualization, describing how the world was
feeling day by day and minute by minute.
Regardless of your field, there is almost always a way web scraping can guide
business practices more effectively, improve productivity, or even branch off into a
brand-new field entirely.
About This Book
This book is designed to serve not only as an introduction to web scraping, but as a
comprehensive guide to scraping almost every type of data from the modern Web.
Although it uses the Python programming language, and covers many Python basics,
it should not be used as an introduction to the language.
If you are not an expert programmer and don’t know any Python at all, this book
might be a bit of a challenge. If, however, you are an experienced programmer, you
should find the material easy to pick up. Appendix A covers installing and working
with Python 3.x, which is used throughout this book. If you have only used Python
2.x, or do not have 3.x installed, you might want to review Appendix A.
If you’re looking for a more comprehensive Python resource, the book Introducing
Python by Bill Lubanovic is a very good, if lengthy, guide. For those with shorter
attention spans, the video series Introduction to Python by Jessika McKeller is an
excellent resource.
Appendix C includes case studies, as well as a breakdown of key issues that might
affect how you can legally run scrapers in the United States and use the data that
they produce.
Technical books are often able to focus on a single language or technology, but web
scraping is a relatively disparate subject, with practices that require the use of
databases, web servers, HTTP, HTML, Internet security, image processing, data
science, and other tools. This book attempts to cover all of these to an extent for the
purpose of gathering data from remote sources across the Internet.
Part I covers the subject of web scraping and web crawling in depth, with a strong
focus on a small handful of libraries used throughout the book. Part I can easily be
used as a comprehensive reference for these libraries and techniques (with certain
exceptions, where additional references will be provided).
Part II covers additional subjects that the reader might find useful when writing
web scrapers. These subjects are, unfortunately, too broad to be neatly wrapped up
in a single chapter. Because of this, frequent references will be made to other
resources for additional information.
The structure of this book is arranged to be easy to jump around among chapters to
find only the web-scraping technique or information that you are looking for. When
a concept or piece of code builds on another mentioned in a previous chapter, I will
explicitly reference the section that it was addressed in.
Conventions Used in This Book
The following typographical conventions are used in this book:
Indicates new terms, URLs, email addresses, filenames, and file extensions.
Constant width
Used for program listings, as well as within paragraphs to refer to program
elements such as variable or function names, databases, data types,
environment variables, statements, and keywords.
Constant width bold
Shows commands or other text that should be typed by the user.
Constant width italic
Shows text that should be replaced with user-supplied values or by values
determined by context.
This element signifies a tip or suggestion.
This element signifies a general note.
This element indicates a warning or caution.
Using Code Examples
Supplemental material (code examples, exercises, etc.) is available for download
This book is here to help you get your job done. In general, if example code is
offered with this book, you may use it in your programs and documentation. You do
not need to contact us for permission unless you’re reproducing a significant
portion of the code. For example, writing a program that uses several chunks of
code from this book does not require permission. Selling or distributing a CDROM of examples from O’Reilly books does require permission. Answering a
question by citing this book and quoting example code does not require permission.
Incorporating a significant amount of example code from this book into your
product’s documentation does require permission.
We appreciate, but do not require, attribution. An attribution usually includes the
title, author, publisher, and ISBN. For example: “Web Scraping with Python by
Ryan Mitchell (O’Reilly). Copyright 2015 Ryan Mitchell, 978-1-491-91029-0.”
If you feel your use of code examples falls outside fair use or the permission given
here, feel free to contact us at [email protected]
Safari® Books Online
Safari Books Online is an on-demand digital library that delivers expert content in
both book and video form from the world’s leading authors in technology and
Technology professionals, software developers, web designers, and business and
creative professionals use Safari Books Online as their primary resource for
research, problem solving, learning, and certification training.
Safari Books Online offers a range of product mixes and pricing programs for
organizations, government agencies, and individuals. Subscribers have access to
thousands of books, training videos, and prepublication manuscripts in one fully
searchable database from publishers like O’Reilly Media, Prentice Hall
Professional, Addison-Wesley Professional, Microsoft Press, Sams, Que, Peachpit
Press, Focal Press, Cisco Press, John Wiley & Sons, Syngress, Morgan Kaufmann,
IBM Redbooks, Packt, Adobe Press, FT Press, Apress, Manning, New Riders,
McGraw-Hill, Jones & Bartlett, Course Technology, and dozens more. For more
information about Safari Books Online, please visit us online.
How to Contact Us
Please address comments and questions concerning this book to the publisher:
O’Reilly Media, Inc.
1005 Gravenstein Highway North
Sebastopol, CA 95472
800-998-9938 (in the United States or Canada)
707-829-0515 (international or local)
707-829-0104 (fax)
We have a web page for this book, where we list errata, examples, and any
additional information. You can access this page at
To comment or ask technical questions about this book, send email to
[email protected]
For more information about our books, courses, conferences, and news, see our
website at
Find us on Facebook:
Follow us on Twitter:
Watch us on YouTube:
Just like some of the best products arise out of a sea of user feedback, this book
could have never existed in any useful form without the help of many collaborators,
cheerleaders, and editors. Thank you to the O’Reilly staff and their amazing support
for this somewhat unconventional subject, to my friends and family who have
offered advice and put up with impromptu readings, and to my coworkers at
LinkeDrive who I now likely owe many hours of work to.
Thank you, in particular, to Allyson MacDonald, Brian Anderson, Miguel Grinberg,
and Eric VanWyk for their feedback, guidance, and occasional tough love. Quite a
few sections and code samples were written as a direct result of their inspirational
Thank you to Yale Specht for his limitless patience throughout the past nine months,
providing the initial encouragement to pursue this project, and stylistic feedback
during the writing process. Without him, this book would have been written in half
the time but would not be nearly as useful.
Finally, thanks to Jim Waldo, who really started this whole thing many years ago
when he mailed a Linux box and The Art and Science of C to a young and
impressionable teenager.
Part I. Building Scrapers
This section focuses on the basic mechanics of web scraping: how to use Python to
request information from a web server, how to perform basic handling of the
server’s response, and how to begin interacting with a website in an automated
fashion. By the end, you’ll be cruising around the Internet with ease, building
scrapers that can hop from one domain to another, gather information, and store that
information for later use.
To be honest, web scraping is a fantastic field to get into if you want a huge payout
for relatively little upfront investment. In all likelihood, 90% of web scraping
projects you’ll encounter will draw on techniques used in just the next six chapters.
This section covers what the general (albeit technically savvy) public tends to think
of when they think of “web scrapers”:
Retrieving HTML data from a domain name
Parsing that data for target information
Storing the target information
Optionally, moving to another page to repeat the process
This will give you a solid foundation before moving on to more complex projects
in part II. Don’t be fooled into thinking that this first section isn’t as important as
some of the more advanced projects in the second half. You will use nearly all the
information in the first half of this book on a daily basis while writing web
Chapter 1. Your First Web Scraper
Once you start web scraping, you start to appreciate all the little things that
browsers do for us. The Web, without a layer of HTML formatting, CSS styling,
JavaScript execution, and image rendering, can look a little intimidating at first, but
in this chapter, as well as the next one, we’ll cover how to format and interpret data
without the help of a browser.
This chapter will start with the basics of sending a GET request to a web server for
a specific page, reading the HTML output from that page, and doing some simple
data extraction in order to isolate the content that we are looking for.
If you haven’t spent much time in networking, or network security, the mechanics of
the Internet might seem a little mysterious. We don’t want to think about what,
exactly, the network is doing every time we open a browser and go to, and, these days, we don’t have to. In fact, I would argue that it’s
fantastic that computer interfaces have advanced to the point where most people
who use the Internet don’t have the faintest idea about how it works.
However, web scraping requires stripping away some of this shroud of interface,
not just at the browser level (how it interprets all of this HTML, CSS, and
JavaScript), but occasionally at the level of the network connection.
To give you some idea of the infrastructure required to get information to your
browser, let’s use the following example. Alice owns a web server. Bob uses a
desktop computer, which is trying to connect to Alice’s server. When one machine
wants to talk to another machine, something like the following exchange takes
1. Bob’s computer sends along a stream of 1 and 0 bits, indicated by high and
low voltages on a wire. These bits form some information, containing a
header and body. The header contains an immediate destination of his local
router’s MAC address, with a final destination of Alice’s IP address. The
body contains his request for Alice’s server application.
2. Bob’s local router receives all these 1’s and 0’s and interprets them as a
packet, from Bob’s own MAC address, and destined for Alice’s IP address.
His router stamps its own IP address on the packet as the “from” IP address,
and sends it off across the Internet.
3. Bob’s packet traverses several intermediary servers, which direct his packet
toward the correct physical/wired path, on to Alice’s server.
4. Alice’s server receives the packet, at her IP address.
5. Alice’s server reads the packet port destination (almost always port 80 for
web applications, this can be thought of as something like an “apartment
number” for packet data, where the IP address is the “street address”), in the
header, and passes it off to the appropriate application – the web server
6. The web server application receives a stream of data from the server
processor. This data says something like:
- This is a GET request
- The following file is requested: index.html
7. The web server locates the correct HTML file, bundles it up into a new
packet to send to Bob, and sends it through to its local router, for transport
back to Bob’s machine, through the same process.
And voilà! We have The Internet.
So, where in this exchange did the web browser come into play? Absolutely
nowhere. In fact, browsers are a relatively recent invention in the history of the
Internet, when Nexus was released in 1990.
Yes, the web browser is a very useful application for creating these packets of
information, sending them off, and interpreting the data you get back as pretty
pictures, sounds, videos, and text. However, a web browser is just code, and code
can be taken apart, broken into its basic components, re-written, re-used, and made
to do anything we want. A web browser can tell the processor to send some data to
the application that handles your wireless (or wired) interface, but many languages
have libraries that can do that just as well.
Let’s take a look at how this is done in Python:
from urllib.request import urlopen
html = urlopen("")
You can save this code as and run it in your terminal using the
Note that if you also have Python 2.x installed on your machine, you may need to
explicitly call Python 3.x by running the command this way:
This will output the complete HTML code for the page at
More accurately, this outputs the HTML file page1.html, found in the directory
<web root>/pages, on the server located at the domain name
What’s the difference? Most modern web pages have many resource files
associated with them. These could be image files, JavaScript files, CSS files, or
any other content that the page you are requesting is linked to. When a web browser
hits a tag such as <img src="cuteKitten.jpg">, the browser knows that it needs
to make another request to the server to get the data at the file cuteKitten.jpg in
order to fully render the page for the user. Keep in mind that our Python script
doesn’t have the logic to go back and request multiple files (yet);it can only read
the single HTML file that we’ve requested.
So how does it do this? Thanks to the plain-English nature of Python, the line
from urllib.request import urlopen
means what it looks like it means: it looks at the Python module request (found
within the urllib library) and imports only the function urlopen.
urllib or urllib2?
If you’ve used the urllib2 library in Python 2.x, you might have noticed that
things have changed somewhat between urllib2 and urllib. In Python 3.x, urllib2
was renamed urllib and was split into several submodules:
urllib.request, urllib.parse, and urllib.error. Although function names
mostly remain the same, you might want to note which functions have moved to
submodules when using the new urllib.
is a standard Python library (meaning you don’t have to install anything
extra to run this example) and contains functions for requesting data across the web,
handling cookies, and even changing metadata such as headers and your user agent.
We will be using urllib extensively throughout the book, so we recommend you
read the Python documentation for the library (
urlopen is used to open a remote object across a network and read it. Because it is
a fairly generic library (it can read HTML files, image files, or any other file
stream with ease), we will be using it quite frequently throughout the book.
An Introduction to BeautifulSoup
“Beautiful Soup, so rich and green,
Waiting in a hot tureen!
Who for such dainties would not stoop?
Soup of the evening, beautiful Soup!”
The BeautifulSoup library was named after a Lewis Carroll poem of the same name
in Alice’s Adventures in Wonderland. In the story, this poem is sung by a character
called the Mock Turtle (itself a pun on the popular Victorian dish Mock Turtle Soup
made not of turtle but of cow).
Like its Wonderland namesake, BeautifulSoup tries to make sense of the
nonsensical; it helps format and organize the messy web by fixing bad HTML and
presenting us with easily-traversible Python objects representing XML structures.
Installing BeautifulSoup
Because the BeautifulSoup library is not a default Python library, it must be
installed. We will be using the BeautifulSoup 4 library (also known as BS4)
throughout this book. The complete instructions for installing BeautifulSoup 4 can
be found at; however, the basic method for Linux is:
$sudo apt-get install python-bs4
and for Macs:
$sudo easy_install pip
This installs the Python package manager pip. Then run the following:
$pip install beautifulsoup4
to install the library.
Again, note that if you have both Python 2.x and 3.x installed on your machine, you
might need to call python3 explicitly:
Make sure to also use this when installing packages, or the packages might be
installed under Python 2.x, but not Python 3.x:
$sudo python3 install
If using pip, you can also call pip3 to install the Python 3.x versions of packages:
$pip3 install beautifulsoup4
Installing packages in Windows is nearly identical to the process for the Mac and
Linux. Download the most recent BeautifulSoup 4 release from the download URL
above, navigate to the directory you unzipped it to, and run:
>python install
And that’s it! BeautifulSoup will now be recognized as a Python library on your
machine. You can test this out by opening a Python terminal and importing it:
> from bs4 import BeautifulSoup
The import should complete without errors.
In addition, there is an .exe installer for pip on Windows, so you can easily install
and manage packages:
>pip install beautifulsoup4
If you intend to work on multiple Python projects or you need a way to easily bundle
projects with all associated libraries, or you’re worried about potential conflicts
between installed libraries, you can install a Python virtual environment to keep
everything separated and easy to manage.
When you install a Python library without a virtual environment, you are installing
it globally. This usually requires that you be an administrator, or run as root, and that
Python library exists for every user and every project on the machine. Fortunately,
creating a virtual environment is easy:
$ virtualenv scrapingEnv
This creates a new environment called scrapingEnv, which you must activate in order
to use:
$ cd scrapingEnv/
$ source bin/activate
After you have activated the environment, you will see that environment’s name in
your command line prompt, reminding you that you’re currently working with it. Any
libraries you install or scripts you run will be under that virtual environment only.
Working in the newly-created scrapingEnv environment, I can install and use
BeautifulSoup, for instance:
(scrapingEnv)ryan$ pip install beautifulsoup4
(scrapingEnv)ryan$ python
> from bs4 import BeautifulSoup
I can leave the environment with the deactivate command, after which I can no longer
access any libraries that were installed inside the virtual environment:
(scrapingEnv)ryan$ deactivate
ryan$ python
> from bs4 import BeautifulSoup
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
ImportError: No module named 'bs4'
Keeping all your libraries separated by project also makes it easy to zip up the entire
environment folder and send it to someone else. As long as they have the same version
of Python installed on their machine, your code will work from the virtual environment
without requiring them to install any libraries themselves.
Although we won’t explicitly instruct you to use a virtual environment in all of this
book’s examples, keep in mind that you can apply virtual environment any time simply
by activating it beforehand.
Running BeautifulSoup
The most commonly used object in the BeautifulSoup library is, appropriately, the
BeautifulSoup object. Let’s take a look at it in action, modifying the example found
in the beginning of this chapter:
from urllib.request import urlopen
from bs4 import BeautifulSoup
html = urlopen("")
bsObj = BeautifulSoup(;
The output is:
<h1>An Interesting Title</h1>
As in the example before, we are importing the urlopen library and calling in order to get the HTML content of the page. This HTML content is
then transformed into a BeautifulSoup object, with the following structure:
html → <html><head>...</head><body>...</body></html>
head → <head><title>A Useful Page<title></head>
title → <title>A Useful Page</title>
body → <body><h1>An Int...</h1><div>Lorem ip...</div></body>
h1 → <h1>An Interesting Title</h1>
div → <div>Lorem Ipsum dolor...</div>
Note that the <h1> tag that we extracted from the page was nested two layers deep
into our BeautifulSoup object structure (html → body → h1). However, when we
actually fetched it from the object, we called the h1 tag directly:
In fact, any of the following function calls would produce the same output:
We hope this small taste of BeautifulSoup has given you an idea of the power and
simplicity of this library. Virtually any information can be extracted from any
HTML (or XML) file, as long as it has some identifying tag surrounding it, or near
it. In chapter 3, we’ll delve more deeply into some more-complex BeautifulSoup
function calls, as well as take a look at regular expressions and how they can be
used with BeautifulSoup in order to extract information from websites.
Connecting Reliably
The web is messy. Data is poorly formatted, websites go down, and closing tags go
missing. One of the most frustrating experiences in web scraping is to go to sleep
with a scraper running, dreaming of all the data you’ll have in your database the
next day — only to find out that the scraper hit an error on some unexpected data
format and stopped execution shortly after you stopped looking at the screen. In
situations like these, you might be tempted to curse the name of the developer who
created the website (and the oddly formatted data), but the person you should really
be kicking is yourself, for not anticipating the exception in the first place!
Let’s take a look at the first line of our scraper, after the import statements, and
figure out how to handle any exceptions this might throw:
html = urlopen("")
There are two main things that can go wrong in this line:
The page is not found on the server (or there was some error in retrieving it)
The server is not found
In the first situation, an HTTP error will be returned. This HTTP error may be “404
Page Not Found,” “500 Internal Server Error,” etc. In all of these cases, the
urlopen function will throw the generic exception “HTTPError” We can handle
this exception in the following way:
html = urlopen("")
except HTTPError as e:
#return null, break, or do some other "Plan B"
#program continues. Note: If you return or break in the
#exception catch, you do not need to use the "else" statement
If an HTTP error code is returned, the program now prints the error, and does not
execute the rest of the program under the else statement.
If the server is not found at all (if, say, was down,
or the URL was mistyped), urlopen returns a None object. This object is analogous
to null in other programming languages. We can add a check to see if the returned
html is None:
if html is None:
print("URL is not found")
#program continues
Of course, if the page is retrieved successfully from the server, there is still the
issue of the content on the page not quite being what we expected. Every time you
access a tag in a BeautifulSoup object, it’s smart to add a check to make sure the tag
actually exists. If you attempt to access a tag that does not exist, BeautifulSoup will
return a None object. The problem is, attempting to access a tag on a None object
itself will result in an AttributeError being thrown.
The following line (where nonExistentTag is a made-up tag, not the name of a
real BeautifulSoup function):
returns a None object. This object is perfectly reasonable to handle and check for.
The trouble comes if you don’t check for it, but instead go on and try to call some
other function on the None object, as illustrated in the following:
which returns the exception:
AttributeError: 'NoneType' object has no attribute 'someTag'
So how can we guard against these two situations? The easiest way is to explicitly
check for both situations:
badContent = bsObj.nonExistingTag.anotherTag
except AttributeError as e:
print("Tag was not found")
if badContent == None:
print ("Tag was not found")
This checking and handling of every error does seem laborious at first, but it’s easy
to add a little reorganization to this code to make it less difficult to write (and,
more importantly, much less difficult to read). This code, for example, is our same
scraper written in a slightly different way:
from urllib.request import urlopen
from urllib.error import HTTPError
from bs4 import BeautifulSoup
def getTitle(url):
html = urlopen(url)
except HTTPError as e:
return None
bsObj = BeautifulSoup(
title = bsObj.body.h1
except AttributeError as e:
return None
return title
title = getTitle("")
if title == None:
print("Title could not be found")
In this example, we’re creating a function getTitle, which returns either the title
of the page, or a None object if there was some problem with retrieving it. Inside
getTitle, we check for an HTTPError, as in the previous example, and also
encapsulate two of the BeautifulSoup lines inside one try statement. An
AttributeError might be thrown from either of these lines (if the server did not
exist, html would be a None object, and would throw an
AttributeError). We could, in fact, encompass as many lines as we wanted
inside one try statement, or call another function entirely, which can throw an
AttributeError at any point.
When writing scrapers, it’s important to think about the overall pattern of your code
in order to handle exceptions and make it readable at the same time. You’ll also
likely want to heavily reuse code. Having generic functions such as
getSiteHTML and getTitle (complete with thorough exception handling) makes it
easy to quickly — and reliably — scrape the web.
Chapter 2. Advanced HTML Parsing
When Michelangelo was asked how he could sculpt a work of art as masterful as
his David, he is famously reported to have said: “It is easy. You just chip away the
stone that doesn’t look like David.”
Although web scraping is unlike marble sculpting in most other respects, we must
take a similar attitude when it comes to extracting the information we’re seeking
from complicated web pages. There are many techniques to chip away the content
that doesn’t look like the content that we’re searching for, until we arrive at the
information we’re seeking. In this chapter, we’ll take look at parsing complicated
HTML pages in order to extract only the information we’re looking for.
You Don’t Always Need a Hammer
It can be tempting, when faced with a Gordian Knot of tags, to dive right in and use
multiline statements to try to extract your information. However, keep in mind that
layering the techniques used in this section with reckless abandon can lead to code
that is difficult to debug, fragile, or both. Before getting started, let’s take a look at
some of the ways you can avoid altogether the need for advanced HTML parsing!
Let’s say you have some target content. Maybe it’s a name, statistic, or block of
text. Maybe it’s buried 20 tags deep in an HTML mush with no helpful tags or
HTML attributes to be found. Let’s say you dive right in and write something like
the following line to attempt extraction:
That doesn’t look so great. In addition to the aesthetics of the line, even the slightest
change to the website by a site administrator might break your web scraper
altogether. So what are your options?
Look for a “print this page” link, or perhaps a mobile version of the site that has
better-formatted HTML (more on presenting yourself as a mobile device — and
receiving mobile site versions — in Chapter 12).
Look for the information hidden in a JavaScript file. Remember, you might need
to examine the imported JavaScript files in order to do this. For example, I once
collected street addresses (along with latitude and longitude) off a website in a
neatly formatted array by looking at the JavaScript for the embedded Google
Map that displayed a pinpoint over each address.
This is more common for page titles, but the information might be available in
the URL of the page itself.
If the information you are looking for is unique to this website for some reason,
you’re out of luck. If not, try to think of other sources you could get this
information from. Is there another website with the same data? Is this website
displaying data that it scraped or aggregated from another website?
Especially when faced with buried or poorly formatted data, it’s important not to
just start digging. Take a deep breath and think of alternatives. If you’re certain no
alternatives exist, the rest of this chapter is for you.
Another Serving of BeautifulSoup
In Chapter 1, we took a quick look at installing and running BeautifulSoup, as well
as selecting objects one at a time. In this section, we’ll discuss searching for tags
by attributes, working with lists of tags, and parse tree navigation.
Nearly every website you encounter contains stylesheets. Although you might think
that a layer of styling on websites that is designed specifically for browser and
human interpretation might be a bad thing, the advent of CSS is actually a boon for
web scrapers. CSS relies on the differentiation of HTML elements that might
otherwise have the exact same markup in order to style them differently. That is,
some tags might look like this:
<span class="green"></span>
while others look like this:
<span class="red"></span>
Web scrapers can easily separate these two different tags based on their class; for
example, they might use BeautifulSoup to grab all of the red text but none of the
green text. Because CSS relies on these identifying attributes to style sites
appropriately, you are almost guaranteed that these class and ID attributes will be
plentiful on most modern websites.
Let’s create an example web scraper that scrapes the page located at
On this page, the lines spoken by characters in the story are in red, whereas the
names of characters themselves are in green. You can see the span tags, which
reference the appropriate CSS classes, in the following sample of the page’s source
"<span class="red">Heavens! what a virulent attack!</span>" replied <span class=
"green">the prince</span>, not in the least disconcerted by this reception.
We can grab the entire page and create a BeautifulSoup object with it using a
program similar to the one used in Chapter 1:
from urllib.request import urlopen
from bs4 import BeautifulSoup
html = urlopen("")
bsObj = BeautifulSoup(html)
Using this BeautifulSoup object, we can use the findAll function to extract a
Python list of proper nouns found by selecting only the text within <span
tags (findAll is an extremely flexible function we’ll be
using a lot later in this book):
nameList = bsObj.findAll("span", {"class":"green"})
for name in nameList:
When run, it should list all the proper nouns in the text, in the order they appear in
War and Peace. So what’s going on here? Previously, we’ve called
bsObj.tagName in order to get the first occurrence of that tag on the page. Now,
we’re calling bsObj.findAll(tagName, tagAttributes) in order to get a list
of all of the tags on the page, rather than just the first.
After getting a list of names, the program iterates through all names in the list, and
prints name.get_text() in order to separate the content from the tags.
When to get_text() and When to Preserve Tags
strips all tags from the document you are working with and returns
a string containing the text only. For example, if you are working with a large
block of text that contains many hyperlinks, paragraphs, and other tags, all those
will be stripped away and you’ll be left with a tagless block of text.
Keep in mind that it’s much easier to find what you’re looking for in a
BeautifulSoup object than in a block of text. Calling .get_text() should always
be the last thing you do, immediately before you print, store, or manipulate your
final data. In general, you should try to preserve the tag structure of a document
as long as possible.
find() and findAll() with BeautifulSoup
BeautifulSoup’s find() and findAll() are the two functions you will likely use
the most. With them, you can easily filter HTML pages to find lists of desired tags,
or a single tag, based on their various attributes.
The two functions are extremely similar, as evidenced by their definitions in the
BeautifulSoup documentation:
findAll(tag, attributes, recursive, text, limit, keywords)
find(tag, attributes, recursive, text, keywords)
In all likelihood, 95% of the time you will find yourself only needing to use the first
two arguments: tag and attributes. However, let’s take a look at all of the
arguments in greater detail.
The tag argument is one that we’ve seen before — you can pass a string name of a
tag or even a Python list of string tag names. For example, the following will return
a list of all the header tags in a document:1
The attributes argument takes a Python dictionary of attributes and matches tags
that contain any one of those attributes. For example, the following function would
return both the green and red span tags in the HTML document:
.findAll("span", {"class":"green", "class":"red"})
The recursive argument is a boolean. How deeply into the document do you want
to go? If recursion is set to True, the findAll function looks into children, and
children’s children, for tags that match your parameters. If it is false, it will look
only at the top-level tags in your document. By default, findAll works recursively
(recursive is set to True); it’s generally a good idea to leave this as is, unless you
really know what you need to do and performance is an issue.
The text argument is unusual in that it matches based on the text content of the tags,
rather than properties of the tags themselves. For instance, if we want to find the
number of times “the prince” was surrounded by tags on the example page, we
could replace our .findAll() function in the previous example with the following
nameList = bsObj.findAll(text="the prince")
The output of this is “7.”
The limit argument, of course, is only used in the findAll method; find is
equivalent to the same findAll call, with a limit of 1. You might set this if you’re
only interested in retrieving the first x items from the page. Be aware, however, that
this gives you the first items on the page in the order that they occur, not necessarily
the first ones that you want.
The keyword argument allows you to select tags that contain a particular attribute.
For example:
allText = bsObj.findAll(id="text")
A Caveat to the keyword Argument
The keyword argument can be very helpful in some situations. However, it is
technically redundant as a BeautifulSoup feature. Keep in mind that anything
that can be done with keyword can also be accomplished using techniques we will
discuss later in this chapter (see Regular Expressions and Lambda Expressions).
For instance, the following two lines are identical:
bsObj.findAll("", {"id":"text"})
In addition, you might occasionally run into problems using keyword, most
notably when searching for elements by their class attribute, because class is a
protected keyword in Python. That is, class is a reserved word in Python that
cannot be used as a variable or argument name (no relation to the
BeautifulSoup.findAll() keyword argument, previously discussed).2 For example,
if you try the following call, you’ll get a syntax error due to the nonstandard use
of class:
Instead, you can use BeautifulSoup’s somewhat clumsy solution, which involves
adding an underscore:
Alternatively, you can enclose class in quotes:
bsObj.findAll("", {"class":"green"}
At this point, you might be asking yourself, “But wait, don’t I already know how to
get a list of tags by attribute — by passing attributes to the function in a dictionary
Recall that passing a list of tags to .findAll() via the attributes list acts as an
“or” filter (i.e., it selects a list of all tags that have tag1 or tag2 or tag3...). If you
have a lengthy list of tags, you can end up with a lot of stuff you don’t want. The
keyword argument allows you to add an additional “and” filter to this.
Other BeautifulSoup Objects
So far in the book, you’ve seen two types of objects in the BeautifulSoup library:
Seen in previous code examples as bsObj
Retrieved in lists or individually by calling find and findAll on a
BeautifulSoup object, or drilling down, as in:
However, there are two more objects in the library that, although less commonly
used, are still important to know about:
Used to represent text within tags, rather than the tags themselves (some
functions operate on, and produce, NavigableStrings, rather than tag
The Comment object
Used to find HTML comments in comment tags, <!--like
this one-->
These four objects are the only objects you will ever encounter (as of the time of
this writing) in the BeautifulSoup library.
Navigating Trees
The findAll function is responsible for finding tags based on their name and
attribute. But what if you need to find a tag based on its location in a document?
That’s where tree navigation comes in handy. In Chapter 1, we looked at navigating
a BeautifulSoup tree in a single direction:
Now let’s look at navigating up, across, and diagonally through HTML trees using
our highly questionable online shopping site as an example
page for scraping (see Figure 2-1):
Figure 2-1. Screenshot from
The HTML for this page, mapped out as a tree (with some tags omitted for brevity),
looks like:
...table rows continue...
We will use this same HTML structure as an example in the next few sections.
Dealing with children and other descendants
In computer science and some branches of mathematics, you often hear about
horrible things done to children: moving them, storing them, removing them, and
even killing them. Fortunately, in BeautifulSoup, children are treated differently.
In the BeautifulSoup library, as well as many other libraries, there is a distinction
drawn between children and descendants: much like in a human family tree,
children are always exactly one tag below a parent, whereas descendants can be at
any level in the tree below a parent. For example, the tr tags are children of the
tag, whereas tr, th, td, img, and span are all descendants of the table tag
(at least in our example page). All children are descendants, but not all descendants
are children.
In general, BeautifulSoup functions will always deal with the descendants of the
current tag selected. For instance, bsObj.body.h1 selects the first h1 tag that is a
descendant of the body tag. It will not find tags located outside of the body.
Similarly, bsObj.div.findAll("img") will find the first div tag in the document,
then retrieve a list of all img tags that are descendants of that div tag.
If you want to find only descendants that are children, you can use the .children
from urllib.request import urlopen
from bs4 import BeautifulSoup
html = urlopen("")
bsObj = BeautifulSoup(html)
for child in bsObj.find("table",{"id":"giftList"}).children:
This code prints out the list of product rows in the giftList table. If you were to
write it using the descendants() function instead of the children() function,
about two dozen tags would be found within the table and printed, including img
tags, span tags, and individual td tags. It’s definitely important to differentiate
between children and descendants!
Dealing with siblings
The BeautifulSoup next_siblings() function makes it trivial to collect data from
tables, especially ones with title rows:
from urllib.request import urlopen
from bs4 import BeautifulSoup
html = urlopen("")
bsObj = BeautifulSoup(html)
for sibling in bsObj.find("table",{"id":"giftList"}).tr.next_siblings:
The output of this code is to print all rows of products from the product table,
except for the first title row. Why does the title row get skipped? Two reasons:
first, objects cannot be siblings with themselves. Any time you get siblings of an
object, the object itself will not be included in the list. Second, this function calls
next siblings only. If we were to select a row in the middle of the list, for example,
and call next_siblings on it, only the subsequent (next) siblings would be
returned. So, by selecting the title row and calling next_siblings, we can select
all the rows in the table, without selecting the title row itself.
Make Selections Specific
Note that the preceding code will work just as well, if we select
or even just in order to select the first row of the table. However, in
the code, I go through all of the trouble of writing everything out in a longer
Even if it looks like there’s just one table (or other target tag) on the page, it’s
easy to miss things. In addition, page layouts change all the time. What was once
the first of its kind on the page, might someday be the second or third tag of that
type found on the page. To make your scrapers more robust, it’s best to be as
specific as possible when making tag selections. Take advantage of tag attributes
when they are available.
As a complement to next_siblings, the previous_siblings function can often
be helpful if there is an easily selectable tag at the end of a list of sibling tags that
you would like to get.
And, of course, there are the next_sibling and previous_sibling functions,
which perform nearly the same function as next_siblings and
previous_siblings, except they return a single tag rather than a list of them.
Dealing with your parents
When scraping pages, you will likely discover that you need to find parents of tags
less frequently than you need to find their children or siblings. Typically, when we
look at HTML pages with the goal of crawling them, we start by looking at the top
layer of tags, and then figure out how to drill our way down into the exact piece of
data that we want. Occasionally, however, you can find yourself in odd situations
that require BeautifulSoup’s parent-finding functions, .parent and .parents. For
from urllib.request import urlopen
from bs4 import BeautifulSoup
html = urlopen("")
bsObj = BeautifulSoup(html)
This code will print out the price of the object represented by the image at the
location ../img/gifts/img1.jpg (in this case, the price is “$15.00”).
How does this work? The following diagram represents the tree structure of the
portion of the HTML page we are working with, with numbered steps:
“$15.00” (4)
s<td> (2)
<img src=”../img/gifts/img1.jpg">(1)
1. The image tag where src="../img/gifts/img1.jpg" is first selected
2. We select the parent of that tag (in this case, the <td> tag).
3. We select the previous_sibling of the <td> tag (in this case, the <td> tag
that contains the dollar value of the product).
4. We select the text within that tag, “$15.00”
Regular Expressions
As the old computer-science joke goes: “Let’s say you have a problem, and you
decide to solve it with regular expressions. Well, now you have two problems.”
Unfortunately, regular expressions (often shortened to regex) are often taught using
large tables of random symbols, strung together to look like a lot of nonsense. This
tends to drive people away, and later they get out into the workforce and write
needlessly complicated searching and filtering functions, when all they needed was
a one-line regular expression in the first place!
Fortunately for you, regular expressions are not all that difficult to get up and
running with quickly, and can easily be learned by looking at and experimenting
with a few simple examples.
Regular expressions are so called because they are used to identify regular strings;
that is, they can definitively say, “Yes, this string you’ve given me follows the rules,
and I’ll return it,” or “This string does not follow the rules, and I’ll discard it.”
This can be exceptionally handy for quickly scanning large documents to look for
strings that look like phone numbers or email addresses.
Notice that I used the phrase regular string. What is a regular string? It’s any string
that can be generated by a series of linear rules,3 such as:
1. Write the letter “a” at least once.
2. Append to this the letter “b” exactly five times.
3. Append to this the letter “c” any even number of times.
4. Optionally, write the letter “d” at the end.
Strings that follow these rules are: “aaaabbbbbccccd,” “aabbbbbcc,” and so on
(there are an infinite number of variations).
Regular expressions are merely a shorthand way of expressing these sets of rules.
For instance, here’s the regular expression for the series of steps just described:
aa*bbbbb(cc)*(d | )
This string might seem a little daunting at first, but it becomes clearer when we
break it down into its components:
The letter a is written, followed by a* (read as a star) which means “any
number of a’s, including 0 of them.” In this way, we can guarantee that the
letter a is written at least once.
No special effects here — just five b’s in a row.
Any even number of things can be grouped into pairs, so in order to enforce
this rule about even things, you can write two c’s, surround them in
parentheses, and write an asterisk after it, meaning that you can have any
number of pairs of c’s (note that this can mean 0 pairs, as well).
(d | )
Adding a bar in the middle of two expressions means that it can be “this thing
or that thing.” In this case, we are saying “add a d followed by a space or just
add a space without a d.” In this way we can guarantee that there is, at most,
one d, followed by a space, completing the string.
Experimenting with RegEx
When learning how to write regular expressions, it’s critical to play around with
them and get a feel for how they work.
If you don’t feel like firing up a code editor, writing a few lines, and running
your program in order to see if a regular expression works as expected, you can
go to a website such as RegexPal and test your regular expressions on the fly.
One classic example of regular expressions can be found in the practice of
identifying email addresses. Although the exact rules governing email addresses
vary slightly from mail server to mail server, we can create a few general rules.
The corresponding regular expression for each of these rules is shown in the
second column:
Rule 1
The first part of
The regular expression shorthand is pretty smart.
an email
For example, it knows that “A-Z” means “any
uppercase letter, A through Z.” By putting all these
contains at least
possible sequences and symbols in brackets (as
one of the
opposed to parentheses) we are saying “this symbol
can be any one of these things we’ve listed in the
brackets.” Note also that the + sign means “these
characters can occur as many times as they want
to, but must occur at least once.”
letters, the
numbers 0-9,
periods (.), plus
signs (+), or
Rule 2
After this, the
email address
contains the @
Rule 3
Rule 4
This is fairly straightforward: the @ symbol must
occur in the middle, and it must occur exactly once.
The email
We may use only letters in the first part of the
address then
domain name, after the @ symbol. Also, there must
must contain at
be at least one character.
least one
uppercase or
This is followed
by a period (.).
Rule 5
Finally, the
email address
ends with com,
org, edu, or net
(in reality, there
are many
possible toplevel domains,
but, these four
should suffice
for the sake of
You must include a period (.) before the domain
This lists the possible sequences of letters that can
occur after the period in the second part of an email
By concatenating all of the rules, we arrive at the regular expression:
[A-Za-z0-9\._+][email protected][A-Za-z]+\.(com|org|edu|net)
When attempting to write any regular expression from scratch, it’s best to first make
a list of steps that concretely outlines what your target string looks like. Pay
attention to edge cases. For instance, if you’re identifying phone numbers, are you
considering country codes and extensions?.
Table 2-1 lists some commonly used regular expression symbols, with a brief
explanation and example. This list is by no means complete, and as mentioned
before, you might encounter slight variations from language to language. However,
these 12 symbols are the most commonly used regular expressions in Python, and
can be used to find and collect most any string type.
Table 2-1. Commonly used regular expression symbols
Symbol(s) Meaning
Matches the preceding character,
subexpression, or bracketed character, 0 or
more times
Matches the preceding character,
subexpression, or bracketed character, 1 or
more times
Matches any character within the brackets [A-Z]*
(i.e., “Pick any one of these things”)
A grouped subexpression (these are
evaluated first, in the “order of operations”
of regular expressions)
{m, n}
Matches the preceding character,
subexpression, or bracketed character
between m and n times (inclusive)
Matches any single character that is not in [^A-Z]*
the brackets
Matches any character, string of
characters, or subexpression, separated by
the “I” (note that this is a vertical bar, or
“pipe,” not a capital “i”)
bad, bid, bed
Matches any single character (including
symbols, numbers, a space, etc.)
bad, bzd,
b$d, b d
Indicates that a character or subexpression ^a
occurs at the beginning of a string
apple, asdf,
An escape character (this allows you to
use “special” characters as their literal
Often used at the end of a regular
[A-Z]*[aexpression, it means “match this up to the z]*$
end of the string.” Without it, every regular
expression has a defacto “.*” at the end of
it, accepting strings where only the first
part of the string matches. This can be
thougt of as analogous to the ^ symbol.
zzzyx, Bob
“Does not contain.” This odd pairing of
^((?![Asymbols, immediately preceding a character Z]).)*$
(or regular expression), indicates that that
character should not be found in that
specific place in the larger string. This can
be tricky to use; after all, the character
might be found in a different part of the
string. If trying to eliminate a character
entirely, use in conjunction with a ^ and $ at
either end.
$ymb0ls a4e
a{2,3}b{2,3} aabbb,
\. \| \\
Regular Expressions: Not Always Regular!
The standard version of regular expressions (the one we are covering in this
book, and that is used by Python and BeautifulSoup) is based on syntax used by
Perl. Most modern programming languages use this or one very similar to it. Be
aware, however, that if you are using regular expressions in another language,
you might encounter problems. Even some modern languages, such as Java,
have slight differences in the way they handle regular expressions. When in
doubt, read the docs!
Regular Expressions and BeautifulSoup
If the previous section on regular expressions seemed a little disjointed from the
mission of this book, here’s where it all ties together. BeautifulSoup and regular
expressions go hand in hand when it comes to scraping the Web. In fact, most
functions that take in a string argument (e.g., find(id="aTagIdHere")) will also
take in a regular expression just as well.
Let’s take a look at some examples, scraping the page found at
Notice that there are many product images on the site — they take the following
<img src="../img/gifts/img3.jpg">
If we wanted to grab URLs to all of the product images, it might seem fairly
straightforward at first: just grab all the image tags using .findAll("img"), right?
But there’s a problem. In addition to the obvious “extra” images (e.g., logos),
modern websites often have hidden images, blank images used for spacing and
aligning elements, and other random image tags you might not be aware of.
Certainly, you can’t count on the only images on the page being product images.
Let’s also assume that the layout of the page might change, or that, for whatever
reason, we don’t want to depend on the position of the image in the page in order to
find the correct tag. This might be the case when you are trying to grab specific
elements or pieces of data that are scattered randomly throughout a website. For
instance, there might be a featured product image in a special layout at the top of
some pages, but not others.
The solution is to look for something identifying about the tag itself. In this case, we
can look at the file path of the product images:
from urllib.request
import urlopenfrom bs4
import BeautifulSoupimport re
html = urlopen("")
bsObj = BeautifulSoup(html)
images = bsObj.findAll("img", {"src":re.compile("\.\.\/img\/gifts/img.*\.jpg")})
for image in images:
This prints out only the relative image paths that start with ../img/gifts/img and end
in .jpg, the output of which is the following:
A regular expression can be inserted as any argument in a BeautifulSoup
expression, allowing you a great deal of flexibility in finding target elements.
Accessing Attributes
So far, we’ve looked at how to access and filter tags and access content within
them. However, very often in web scraping you’re not looking for the content of a
tag; you’re looking for its attributes. This becomes especially useful for tags such
as <a>, where the URL it is pointing to is contained within the href attribute, or the
<img> tag, where the target image is contained within the src attribute.
With tag objects, a Python list of attributes can be automatically accessed by
Keep in mind that this literally returns a Python dictionary object, which makes
retrieval and manipulation of these attributes trivial. The source location for an
image, for example, can be found using the following line:
Lambda Expressions
If you have a formal education in computer science, you probably learned about
lambda expressions once in school and then never used them again. If you don’t,
they might be unfamiliar to you (or familiar only as “that thing I’ve been meaning to
learn at some point”). In this section, we won’t go deeply into these extremely
useful functions, but we will look at a few examples of how they can be useful in
web scraping.
Essentially, a lambda expression is a function that is passed into another function as
a variable; that is, instead of defining a function as f(x, y), you may define a
function as f(g(x), y), or even f(g(x), h(x)).
BeautifulSoup allows us to pass certain types of functions as parameters into the
findAll function. The only restriction is that these functions must take a tag object
as an argument and return a boolean. Every tag object that BeautifulSoup encounters
is evaluated in this function, and tags that evaluate to “true” are returned while the
rest are discarded.
For example, the following retrieves all tags that have exactly two attributes:
soup.findAll(lambda tag: len(tag.attrs) == 2)
That is, it will find tags such as the following:
<div class="body" id="content"></div>
<span style="color:red" class="title"></span>
Using lambda functions in BeautifulSoup, selectors can act as a great substitute for
writing a regular expression, if you’re comfortable with writing a little code.
Beyond BeautifulSoup
Although BeautifulSoup is used throughout this book (and is one of the most
popular HTML libraries available for Python), keep in mind that it’s not the only
option. If BeautifulSoup does not meet your needs, check out these other widely
used libraries:
This library is used for parsing both HTML and XML documents, and is
known for being very low level and heavily based on C. Although it takes a
while to learn (a steep learning curve actually means you learn it very fast), it
is very fast at parsing most HTML documents.
HTML Parser
This is Python’s built-in parsing library. Because it requires no installation
(other than, obviously, having Python installed in the first place), it can be
extremely convenient to use.
If you’re looking to get a list of all h<some_level> tags in the document, there
are more succinct ways of writing this code to accomplish the same thing. We’ll
take a look at other ways of approaching these types of problems in the section
BeautifulSoup and regular expressions.
The Python Language Reference provides a complete list of protected keywords.
You might be asking yourself, “Are there ‘irregular’ expressions?” Nonregular
expressions are beyond the scope of this book, but they encompass strings such as
“write a prime number of a’s, followed by exactly twice that number of b’s” or
“write a palindrome.” It’s impossible to identify strings of this type with a regular
expression. Fortunately, I’ve never been in a situation where my web scraper
needed to identify these kinds of strings.
Chapter 3. Starting to Crawl
So far, the examples in the book have covered single static pages, with somewhat
artificial canned examples. In this chapter, we’ll start looking at some real-world
problems, with scrapers traversing multiple pages and even multiple sites.
Web crawlers are called such because they crawl across the Web. At their core is
an element of recursion. They must retrieve page contents for a URL, examine that
page for another URL, and retrieve that page, ad infinitum.
Beware, however: just because you can crawl the Web doesn’t mean that you
always should. The scrapers used in previous examples work great in situations
where all the data you need is on a single page. With web crawlers, you must be
extremely conscientious of how much bandwidth you are using and make every
effort to determine if there’s a way to make the target server’s load easier.
Traversing a Single Domain
Even if you haven’t heard of “Six Degrees of Wikipedia,” you’ve almost certainly
heard of its namesake, “Six Degrees of Kevin Bacon.” In both games, the goal is to
link two unlikely subjects (in the first case, Wikipedia articles that link to each
other, in the second case, actors appearing in the same film) by a chain containing
no more than six total (including the two original subjects).
For example, Eric Idle appeared in Dudley Do-Right with Brendan Fraser, who
appeared in The Air I Breathe with Kevin Bacon.1 In this case, the chain from Eric
Idle to Kevin Bacon is only three subjects long.
In this section, we’ll begin a project that will become a “Six Degrees of
Wikipedia" solution finder. That is, we’ll be able to take the Eric Idle page and find
the fewest number of link clicks that will take us to the Kevin Bacon page.
According to the Wikimedia Foundation (the parent organization behind Wikipedia), the
site’s web properties receive approximately 2,500 hits per second, with more than 99%
of them to the Wikipedia domain (see the “Traffic Volume” section of the “Wikimedia
in Figures” page). Because of the sheer volume of traffic, your web scrapers are
unlikely to have any noticeable impact on Wikipedia’s server load. However, if you run
the code samples in this book extensively, or create your own projects that scrape
Wikipedia, I encourage you to make a tax-deductible donation to the Wikimedia
Foundation — even a few dollars will offset your server load and help make education
resources available for everyone else.
You should already know how to write a Python script that retrieves an arbitrary
Wikipedia page and produces a list of links on that page:
from urllib.request import urlopen
from bs4 import BeautifulSoup
html = urlopen("")
bsObj = BeautifulSoup(html)
for link in bsObj.findAll("a"):
if 'href' in link.attrs:
If you look at the list of links produced, you’ll notice that all the articles you’d
expect are there: “Apollo 13,” “Philadelphia,” “Primetime Emmy Award,” and so
on. However, there are some things that we don’t want as well:
In fact, Wikipedia is full of sidebar, footer, and header links that appear on every
page, along with links to the category pages, talk pages, and other pages that do not
contain different articles:
Recently a friend of mine, while working on a similar Wikipedia-scraping project,
mentioned he had written a very large filtering function, with over 100 lines of
code, in order to determine whether an internal Wikipedia link was an article page
or not. Unfortunately, he had not spent much time up front trying to find patterns
between “article links” and “other links,” or he might have discovered the trick. If
you examine the links that point to article pages (as opposed to other internal
pages), they all have three things in common:
They reside within the div with the id set to bodyContent
The URLs do not contain semicolons
The URLs begin with /wiki/
We can use these rules to revise the code slightly to retrieve only the desired article
from urllib.request import urlopen
from bs4 import BeautifulSoup
import re
html = urlopen("")
bsObj = BeautifulSoup(html)
for link in bsObj.find("div", {"id":"bodyContent"}).findAll("a",
if 'href' in link.attrs:
If you run this, you should see a list of all article URLs that the Wikipedia article on
Kevin Bacon links to.
Of course, having a script that finds all article links in one, hardcoded Wikipedia
article, while interesting, is fairly useless in practice. We need to be able to take
this code and transform it into something more like the following:
A single function, getLinks, that takes in a Wikipedia article URL of the form
/wiki/<Article_Name> and returns a list of all linked article URLs in the
same form.
A main function that calls getLinks with some starting article, chooses a
random article link from the returned list, and calls getLinks again, until we
stop the program or until there are no article links found on the new page.
Here is the complete code that accomplishes this:
from urllib.request import urlopen
from bs4 import BeautifulSoup
import datetime
import random
import re
def getLinks(articleUrl):
html = urlopen(""+articleUrl)
bsObj = BeautifulSoup(html)
return bsObj.find("div", {"id":"bodyContent"}).findAll("a",
links = getLinks("/wiki/Kevin_Bacon")
while len(links) > 0:
newArticle = links[random.randint(0, len(links)-1)].attrs["href"]
links = getLinks(newArticle)
The first thing the program does, after importing the needed libraries, is set the
random number generator seed with the current system time. This practically
ensures a new and interesting random path through Wikipedia articles every time
the program is run.
In the previous example, I used Python’s random number generator to select an article
at random on each page in order to continue a random traversal of Wikipedia.
However, random numbers should be used with caution.
While computers are great at calculating correct answers, they’re terrible at just making
things up. For this reason, random numbers can be a challenge. Most random number
algorithms strive to produce an evenly distributed and hard-to-predict sequence of
numbers, but a “seed” number is needed to give these algorithms something to work
with initially. The exact same seed will produce the exact same sequence of “random”
numbers every time, so for this reason I’ve used the system clock as a starter for
producing new sequences of random numbers, and, thus, new sequences of random
articles. This makes the program a little more exciting to run.
For the curious, the Python pseudorandom number generator is powered by the
Mersenne Twister algorithm. While it produces random numbers that are difficult to
predict and uniformly distributed, it is slightly processor intensive. Random numbers
this good don’t come cheap!
Next, it defines the getLinks function, which takes in an article URL of the form
/wiki/..., prepends the Wikipedia domain name,,
and retrieves the BeautifulSoup object for the HTML at that domain. It then extracts
a list of article link tags, based on the parameters discussed previously, and returns
The main body of the program begins with setting a list of article link tags (the
links variable) to the list of links in the initial page: It then
goes into a loop, finding a random article link tag in the page, extracting the href
attribute from it, printing the page, and getting a new list of links from the extracted
Of course, there’s a bit more to solving a “Six Degrees of Wikipedia” problem than
simply building a scraper that goes from page to page. We must also be able to
store and analyze the resulting data. For a continuation of the solution to this
problem, see Chapter 5.
Handle Your Exceptions!
Although we are omitting most exception handling in the code examples for the
sake of brevity in these examples, be aware that there are many potential pitfalls
that could arise: What if Wikipedia changed the name of the bodyContent tag, for
example? (Hint: the code would crash.)
So although these scripts might be fine to run as closely watched examples,
autonomous production code requires far more exception handling than we can
fit into this book. Look back to Chapter 1 for more information about this.
Crawling an Entire Site
In the previous section, we took a random walk through a website, going from link
to link. But what if you need to systematically catalog or search every page on a
site? Crawling an entire site, especially a large one, is a memory-intensive process
that is best suited to applications where a database to store crawling results is
readily available. However, we can explore the behavior of these types of
applications without actually running them full-scale. To learn more about running
these applications using a database, see Chapter 5.
You’ve likely heard the terms deep Web, dark Web, or hidden Web being thrown
around a lot, especially in the media lately. What do they mean?
The deep Web is simply any part of the Web that’s not part of the surface Web The
surface is part of the Internet that is indexed by search engines. Estimates vary widely,
but the deep Web almost certainly makes up about 90% of the Internet. Because
Google can’t do things like submit forms, find pages that haven’t been linked to by a
top-level domain, or investigate sites where the robots.txt prohibits it, the surface Web
stays relatively small.
The dark Web, also known as the Darknet or dark Internet, is another beast entirely. It
is run over the existing network infrastructure but uses a Tor client with an application
protocol that runs on top of HTTP, providing a secure channel to exchange
information. Although it is possible to scrape the Darknet, just like you’d scrape any
other website, doing so is outside the scope of this book.
Unlike the dark Web, the deep Web is relatively easy to scrape. In fact, there are many
tools in this book that will teach you how to crawl and scrape information from many
places that Google bots can’t go.
So when might crawling an entire website be useful and when might it actually be
harmful? Web scrapers that traverse an entire site are good for many things,
Generating a site map
A few years ago, I was faced with a problem: an important client wanted an
estimate for a website redesign, but did not want to provide my company with
access to the internals of their current content management system, and did not
have a publicly available site map. I was able to use a crawler to cover their
entire site, gather all internal links, and organize the pages into the actual
folder structure they had on their site. This allowed me to quickly find sections
of the site I wasn’t even aware existed, and accurately count how many page
designs would be required, and how much content would need to be migrated.
Gathering data
Another client of mine wanted to gather articles (stories, blog posts, news
articles, etc.) in order to create a working prototype of a specialized search
platform. Although these website crawls didn’t need to be exhaustive, they did
need to be fairly expansive (there were only a few sites we were interested in
getting data from). I was able to create crawlers that recursively traversed
each site and collected only data it found on article pages.
The general approach to an exhaustive site crawl is to start with a top-level page
(such as the home page), and search for a list of all internal links on that page.
Every one of those links is then crawled, and additional lists of links are found on
each one of them, triggering another round of crawling.
Clearly this is a situation that can blow up very quickly. If every page has 10
internal links, and a website is five pages deep (a fairly typical depth for a
medium-size website), then the number of pages you need to crawl is
, or
100,000 pages, before you can be sure that you’ve exhaustively covered the
website. Strangely enough, while “5 pages deep and 10 internal links per page” are
fairly typical dimensions for a website, there are very few websites with 100,000
or more pages. The reason, of course, is that the vast majority of internal links are
In order to avoid crawling the same page twice, it is extremely important that all
internal links discovered are formatted consistently, and kept in a running list for
easy lookups, while the program is running. Only links that are “new” should be
crawled and searched for additional links:
from urllib.request import urlopen
from bs4 import BeautifulSoup
import re
pages = set()
def getLinks(pageUrl):
global pages
html = urlopen(""+pageUrl)
bsObj = BeautifulSoup(html)
for link in bsObj.findAll("a", href=re.compile("^(/wiki/)")):
if 'href' in link.attrs:
if link.attrs['href'] not in pages:
#We have encountered a new page
newPage = link.attrs['href']
In order to get the full effect of how this web-crawling business works, I’ve
relaxed the standards of what constitutes an “internal link we are looking for,” from
previous examples. Rather than limit the scraper to article pages, it looks for all
links that begin with /wiki/ regardless of where they are on the page, and regardless
of whether they contain colons. Remember: article pages do not contain colons, but
file upload pages, talk pages, and the like do contain colons in the URL).
Initially, getLinks is called with an empty URL. This is translated as “the front
page of Wikipedia” as soon as the empty URL is prepended with inside the function. Then, each link on the first page is
iterated through and a check is made to see if it is in the global set of pages (a set of
pages that the script has encountered already). If not, it is added to the list, printed
to the screen, and the getLinks function is called recursively on it.
A Warning Regarding Recursion
This is a warning rarely seen in software books, but I thought you should be
aware: if left running long enough, the preceding program will almost certainly
Python has a default recursion limit (how many times programs can recursively
call themselves) of 1,000. Because Wikipedia’s network of links is extremely
large, this program will eventually hit that recursion limit and stop, unless you
put in a recursion counter or something to prevent that from happening.
For “flat” sites that are fewer than 1,000 links deep, this method usually works
very well, with some unusual exceptions. For example, I once encountered a
website that had a rule for generating internal links to blog posts. The rule was
“take the URL for the current page we are on, and append
/blog/title_of_blog.php to it.”
The problem was that they would append /blog/title_of_blog.php to URLs that
were already on a page that had /blog/ in the URL. So the site would simply add
another /blog/ on. Eventually, my crawler was going to URLs like:
Eventually, I had to add a check to make sure that URLs weren’t overly
ridiculous, containing repeating segments that might indicate an infinite loop.
However, if I had left this running unchecked overnight, it could have easily
Collecting Data Across an Entire Site
Of course, web crawlers would be fairly boring if all they did was hop from one
page to the other. In order to make them useful, we need to be able to do something
on the page while we’re there. Let’s look at how to build a scraper that collects the
title, the first paragraph of content, and the link to edit the page (if available).
As always, the first step to determine how best to do this is to look at a few pages
from the site and determine a pattern. By looking at a handful of Wikipedia pages
both articles and non-article pages such as the privacy policy page, the following
things should be clear:
All titles (on all pages, regardless of their status as an article page, an edit
history page, or any other page) have titles under h1→span tags, and these are
the only h1 tags on the page.
As mentioned before, all body text lives under the div#bodyContent tag.
However, if we want to get more specific and access just the first paragraph of
text, we might be better off using div#mw-content-text →p (selecting the first
paragraph tag only). This is true for all content pages except file pages (for
example:, which do not have sections of content text.
Edit links occur only on article pages. If they occur, they will be found in the
li#ca-edit tag, under li#ca-edit →​ span →​ a.
By modifying our basic crawling code, we can create a combination crawler/datagathering (or, at least, data printing) program:
from urllib.request import urlopen
from bs4 import BeautifulSoup
import re
pages = set()
def getLinks(pageUrl):
global pages
html = urlopen(""+pageUrl)
bsObj = BeautifulSoup(html)
print(bsObj.find(id ="mw-content-text").findAll("p")[0])
except AttributeError:
print("This page is missing something! No worries though!")
for link in bsObj.findAll("a", href=re.compile("^(/wiki/)")):
if 'href' in link.attrs:
if link.attrs['href'] not in pages:
#We have encountered a new page
newPage = link.attrs['href']
The for loop in this program is essentially the same as it was in the original
crawling program (with the addition of some printed dashes for clarity, separating
the printed content).
Because we can never be entirely sure that all the data is on each page, each print
statement is arranged in the order that it is likeliest to appear on the site. That is, the
<h1> title tag appears on every page (as far as I can tell, at any rate) so we attempt
to get that data first. The text content appears on most pages (except for file pages),
so that is the second piece of data retrieved. The “edit” button only appears on
pages where both titles and text content already exist, but it does not appear on all
of those pages.
Different Patterns for Different Needs
There are obviously some dangers involved with wrapping multiple lines in an
exception handler. You cannot tell which line threw the exception, for one thing.
Also, if for some reason a page contained an “edit” button but no title, the “edit”
button would never be logged. However, it suffices for many instances in which
there is an order of likeliness of items appearing on the site, and inadvertently
missing a few data points or keeping detailed logs is not a problem.
You might notice that in this and all the previous examples, we haven’t been
“collecting” data so much as “printing” it. Obviously, data in your terminal is hard
to manipulate. We’ll look more at storing information and creating databases in
Chapter 5.
Crawling Across the Internet
Whenever I give a talk on web scraping, someone inevitably asks: “How do you
build Google?” My answer is always twofold: “First, you get many billions of
dollars so that you can buy the world’s largest data warehouses and place them in
hidden locations all around the world. Second, you build a web crawler.”
When Google started in 1994, it was just two Stanford graduate students with an
old server and a Python web crawler. Now that you know this, you officially have
the tools you need to become the next tech multi-billionaire!
In all seriousness, web crawlers are at the heart of what drives many modern web
technologies, and you don’t necessarily need a large data warehouse to use them. In
order to do any cross-domain data analysis, you do need to build crawlers that can
interpret and store data across the myriad of different pages on the Internet.
Just like in the previous example, the web crawlers we are going to build will
follow links from page to page, building out a map of the Web. But this time, they
will not ignore external links; they will follow them. For an extra challenge, we’ll
see if we can record some kind of information about each page as we move through
it. This will be harder than working with a single domain as we did before —
different websites have completely different layouts. This means we will have to
be very flexible in the kind of information we’re looking for and how we’re
looking for it.
Unknown Waters Ahead
Keep in mind that the code from the next section can go anywhere on the
Internet. If we’ve learned anything from “Six Degrees of Wikipedia,” it’s that it’s
entirely possible to go from a site like to something
less savory in just a few hops.
Kids, ask your parents before running this code. For those with sensitive
constitutions or with religious restrictions that might prohibit seeing text from a
prurient site, follow along by reading the code examples but be careful when
actually running them.
Before you start writing a crawler that simply follows all outbound links willy
nilly, you should ask yourself a few questions:
What data am I trying to gather? Can this be accomplished by scraping just a few
predefined websites (almost always the easier option), or does my crawler need
to be able to discover new websites I might not know about?
When my crawler reaches a particular website, will it immediately follow the
next outbound link to a new website, or will it stick around for a while and drill
down into the current website?
Are there any conditions under which I would not want to scrape a particular
site? Am I interested in non-English content?
How am I protecting myself against legal action if my web crawler catches the
attention of a webmaster on one of the sites it runs across? (Check out
Appendix C for more information on this subject.)
A flexible set of Python functions that can be combined to perform a variety of
different types of web scraping can be easily written in fewer than 50 lines of code:
from urllib.request import urlopen
from bs4 import BeautifulSoup
import re
import datetime
import random
pages = set()
#Retrieves a list of all Internal links found on a page
def getInternalLinks(bsObj, includeUrl):
internalLinks = []
#Finds all links that begin with a "/"
for link in bsObj.findAll("a", href=re.compile("^(/|.*"+includeUrl+")")):
if link.attrs['href'] is not None:
if link.attrs['href'] not in internalLinks:
return internalLinks
#Retrieves a list of all external links found on a page
def getExternalLinks(bsObj, excludeUrl):
externalLinks = []
#Finds all links that start with "http" or "www" that do
#not contain the current URL
for link in bsObj.findAll("a",
if link.attrs['href'] is not None:
if link.attrs['href'] not in externalLinks:
return externalLinks
def splitAddress(address):
addressParts = address.replace("http://", "").split("/")
return addressParts
def getRandomExternalLink(startingPage):
html = urlopen(startingPage)
bsObj = BeautifulSoup(html)
externalLinks = getExternalLinks(bsObj, splitAddress(startingPage)[0])
if len(externalLinks) == 0:
internalLinks = getInternalLinks(startingPage)
return getNextExternalLink(internalLinks[random.randint(0,
return externalLinks[random.randint(0, len(externalLinks)-1)]
def followExternalOnly(startingSite):
externalLink = getRandomExternalLink("")
print("Random external link is: "+externalLink)
The preceding program starts at and randomly hops from external
link to external link. Here’s an example of the output it produces:
External links are not always guaranteed to be found on the first page of a website.
In order to find external links in this case, a method similar to the one used in the
previous crawling example is employed to recursively drill down into a website
until it finds an external link.
Figure 3-1 visualizes the operation as a flowchart:
Figure 3-1. Flowchart for script that crawls through different sites on the Internet
Don’t Put Example Programs into Production
I keep bringing this up, but it’s important for space and readability, the example
programs in this book do not always contain the necessary checks and exception
handling required for production-ready code.
For example, if an external link is not found anywhere on a site that this crawler
encounters (unlikely, but it’s bound to happen at some point if you run it for long
enough), this program will keep running until it hits Python’s recursion limit.
Before running this code for any serious purpose, make sure that you are putting
checks in place to handle potential pitfalls.
The nice thing about breaking up tasks into simple functions such as “find all
external links on this page” is that the code can later be easily refactored to perform
a different crawling task. For example, if our goal is to crawl an entire site for
external links, and make a note of each one, we can add the following function:
#Collects a list of all external URLs found on the site
allExtLinks = set()
allIntLinks = set()
def getAllExternalLinks(siteUrl):
html = urlopen(siteUrl)
bsObj = BeautifulSoup(html)
internalLinks = getInternalLinks(bsObj,splitAddress(siteUrl)[0])
externalLinks = getExternalLinks(bsObj,splitAddress(siteUrl)[0])
for link in externalLinks:
if link not in allExtLinks:
for link in internalLinks:
if link not in allIntLinks:
print("About to get link: "+link)
This code can be thought of as two loops — one gathering internal links, one
gathering external links — working in conjunction with each other. The flowchart
looks something like Figure 3-2:
Figure 3-2. Flow diagram for the website crawler that collects all external links
Jotting down or making diagrams of what the code should do before you write the
code itself is a fantastic habit to get into, and one that can save you a lot of time and
frustration as your crawlers get more complicated.
Redirects allow the same web page to be viewable under different domain names.
Redirects come in two flavors:
Server-side redirects, where the URL is changed before the page is loaded
Client-side redirects, sometimes seen with a “You will be directed in 10
seconds...” type of message, where the page loads before redirecting to the new
This section will handle server-side redirects. For more information on client-side
redirects, which are performed using JavaScript or HTML, see Chapter 10.
For server-side redirects, you usually don’t have to worry. If you’re using the urllib
library with Python 3.x, it handles redirects automatically! Just be aware that,
occasionally, the URL of the page you’re crawling might not be exactly the URL that
you entered the page on.
Crawling with Scrapy
One of the challenges of writing web crawlers is that you’re often performing the
same tasks again and again: find all links on a page, evaluate the difference
between internal and external links, go to new pages. These basic patterns are
useful to know about and to be able to write from scratch, but there are options if
you want something else to handle the details for you.
Scrapy is a Python library that handles much of the complexity of finding and
evaluating links on a website, crawling domains or lists of domains with ease.
Unfortunately, Scrapy has not yet been released for Python 3.x, though it is
compatible with Python 2.7.
The good news is that multiple versions of Python (e.g., Python 2.7 and 3.4) usually
work well when installed on the same machine. If you want to use Scrapy for a
project but also want to use various other Python 3.4 scripts, you shouldn’t have a
problem doing both.
The Scrapy website offers the tool for download from its website, as well as
instructions for installing Scrapy with third-party installation managers such as pip.
Keep in mind that you will need to install Scrapy using Python 2.7 (it is not
compatible with 2.6 or 3.x) and run all programs using Scrapy with Python 2.7 as
Although writing Scrapy crawlers is relatively easy, there is a small amount of
setup that needs to be done for each crawler. To create a new Scrapy project in the
current directory, run from the command line:
$scrapy startproject wikiSpider
is the name of our new project. This creates a new directory in the
directory the project was created in, with the title wikiSpider. Inside this
directory is the following file structure:
In order to create a crawler, we will add a new file to
wikiSpider/wikiSpider/spiders/ called In addition, we
will define a new item called Article inside the file.
Your file should be edited to look like this (with Scrapy-generated
comments left in place, although you can feel free to remove them):
-*- coding: utf-8 -*Define here the models for your scraped items
See documentation in:
from scrapy import Item, Field
class Article(Item):
# define the fields for your item here like:
# name = scrapy.Field()
title = Field()
Each Scrapy Item object represents a single page on the website. Obviously, you
can define as many fields as you’d like (url, content, header image, etc.), but
I’m simply collecting the title field from each page, for now.
In your newly created file, write the following:
from scrapy.selector import Selector
from scrapy import Spider
from wikiSpider.items import Article
class ArticleSpider(Spider):
allowed_domains = [""]
start_urls = ["",
def parse(self, response):
item = Article()
title = response.xpath('//h1/text()')[0].extract()
print("Title is: "+title)
item['title'] = title
return item
The name of this object (ArticleSpider) is different from the name of the
directory (WikiSpider), indicating that this class in particular is responsible for
spidering only through article pages, under the broader category of WikiSpider. For
large sites with many types of content, you might have separate Scrapy items for
each type (blog posts, press releases, articles, etc.), each with different fields, but
all running under the same Scrapy project.
You can run this ArticleSpider from inside the main WikiSpider directory by
$ scrapy crawl article
This calls the scraper by the item name of article (not the class or file name, but
the name defined on the line: name="article" in the ArticleSpider).
Along with some debugging information, this should print out the lines:
Title is: Main Page
Title is: Python (programming language)
The scraper goes to the two pages listed as the start_urls, gathers information,
and then terminates. Not much of a crawler, but using Scrapy in this way can be
useful if you have a list of URLs you need scrape. To turn it into a fully fledged
crawler, you need to define a set of rules that Scrapy can use to seek out new URLs
on each page it encounters:
from scrapy.contrib.spiders import CrawlSpider, Rule
from wikiSpider.items import Article
from scrapy.contrib.linkextractors.sgml import SgmlLinkExtractor
class ArticleSpider(CrawlSpider):
allowed_domains = [""]
start_urls = ["
rules = [Rule(SgmlLinkExtractor(allow=('(/wiki/)((?!:).)*$'),),
callback="parse_item", follow=True)]
def parse_item(self, response):
item = Article()
title = response.xpath('//h1/text()')[0].extract()
print("Title is: "+title)
item['title'] = title
return item
This crawler is run from the command line in the same way as the previous one, but
it will not terminate (at least not for a very, very long time) until you halt execution
using Ctrl+C or by closing the terminal.
Logging with Scrapy
The debug information generated by Scrapy can be useful, but it is often too
verbose. You can easily adjust the level of logging by adding a line to the file in your Scrapy project:
There are five levels of logging in Scrapy, listed in order here:
If logging is set to ERROR, only CRITICAL and ERROR logs will be displayed. If
logging is set to INFO, all logs will be displayed, and so on.
To output logs to a separate logfile instead of the terminal, simply define a logfile
when running from the command line:
$ scrapy crawl article -s LOG_FILE=wiki.log
This will create a new logfile, if one does not exist, in your current directory and
output all logs and print statements to it.
Scrapy uses the Item objects to determine which pieces of information it should
save from the pages it visits. This information can be saved by Scrapy in a variety
of ways, such as a CSV, JSON, or XML files, using the following commands:
$ scrapy crawl article -o articles.csv -t csv
$ scrapy crawl article -o articles.json -t json
$ scrapy crawl article -o articles.xml -t xml
Of course, you can use the Item objects yourself and write them to a file or a
database in whatever way you want, simply by adding the appropriate code to the
parsing function in the crawler.
Scrapy is a powerful tool that handles many problems associated with crawling the
Web. It automatically gathers all URLs and compares them against predefined rules;
makes sure all URLs are unique; normalizes relative URLs where needed; and
recurses to go more deeply into pages.
Although this section hardly scratches the surface of what Scrapy is capable of, I
encourage you to check out the Scrapy documentation or many of the other
resources available online. Scrapy is an extremely large and sprawling library with
many features. If there’s something you’d like to do with Scrapy that has not been
mentioned here, there is likely a way (or several) to do it!
Thanks to The Oracle of Bacon for satisfying my curiosity about this particular
Chapter 4. Using APIs
Like many programmers who have worked on large projects, I have my share of
horror stories when it comes to working with other people’s code. From namespace
issues to type issues to misunderstandings of function output, simply trying to get
information from point A to method B can be a nightmare.
This is where application programming interfaces come in handy: they provide
nice, convenient interfaces between multiple disparate applications. It doesn’t
matter if the applications are written by different programmers, with different
architectures, or even in different languages — APIs are designed to serve as a
lingua franca between different pieces of software that need to share information
with each other.
Although various APIs exist for a variety of different software applications, in
recent times “API” has been commonly understood as meaning “web application
API.” Typically, a programmer will make a request to an API via HTTP for some
type of data, and the API will return this data in the form of XML or JSON.
Although most APIs still support XML, JSON is quickly becoming the encoding
protocol of choice.
If taking advantage of a ready-to-use program to get information prepackaged in a
useful format seems like a bit of a departure from the rest of this book, well, it is
and it isn’t. Although using APIs isn’t generally considered web scraping by most
people, both practices use many of the same techniques (sending HTTP requests)
and produce similar results (getting information); they often can be very
complementary to each other.
For instance, you might want to combine information gleaned from a web scraper
with information from a published API in order to make the information more useful
to you. In an example later in this chapter, we will look at combining Wikipedia
edit histories (which contain IP addresses) with an IP address resolver API in
order to get the geographic location of Wikipedia edits around the world.
In this chapter, we’ll offer a general overview of APIs and how they work, look at
a few popular APIs available today, and look at how you might use an API in your
own web scrapers.
How APIs Work
Although APIs are not nearly as ubiquitous as they should be (a large motivation for
writing this book, because if you can’t find an API, you can still get the data through
scraping), you can find APIs for many types of information. Interested in music?
There are a few different APIs that can give you songs, artists, albums, and even
information about musical styles and related artists. Need sports data? ESPN
provides APIs for athlete information, game scores, and more. Google has dozens
of APIs in its Developers section for language translations, analytics, geolocation,
and more.
APIs are extremely easy to use. In fact, you can try out a simple API request just by
entering the following in your browser:1
This should produce the following response:
{"ip":"","country_code":"US","country_name":"United States,"region_
So, wait, you navigate to a web address in your browser window and it produces
some information (albeit, very well-formatted)? What’s the difference between an
API and a regular website? Despite the hype around APIs, the answer is often: not
much. APIs function via HTTP, the same protocol used to fetch data for websites,
download a file, and do almost anything else on the Internet. The only things that
makes an API an API is the extremely regulated syntax it uses, and the fact that APIs
present their data as JSON or XML, rather than HTML.
Common Conventions
Unlike the subjects of most web scraping, APIs follow an extremely standardized
set of rules to produce information, and they produce that information in an
extremely standardized way as well. Because of this, it is easy to learn a few
simple ground rules that will help you to quickly get up and running with any given
API, as long as it’s fairly well written.
That being said, keep in mind that some APIs do deviate slightly from these rules,
so it’s important to read an API’s documentation the first time you use it, regardless
of how familiar you are with APIs in general.
There are four ways to request information from a web server via HTTP:
is what you use when you visit a website through the address bar in your
browser. GET is the method you are using when you make a call
to You can think of GET as saying, “Hey,
web server, please get me this information.”
is what you use when you fill out a form, or submit information, presumably
to a backend script on the server. Every time you log into a website, you are making
a POST request with your username and (hopefully) encrypted password. If you are
making a POST request with an API, you are saying, “Please store this information
in your database.”
is less commonly used when interacting with websites, but is used from time to
time in APIs. A PUT request is used to update an object or information. An API
might require a POST request to create a new user, for example, but it might need a
PUT request if you want to update that user’s email address.2
is straightforward; it is used to delete an object. For instance, if I send a
DELETE request to, it will delete the user with the ID 23.
DELETE methods are not often encountered in public APIs, which are primarily
created to disseminate information rather than allow random users to remove that
information from their databases. However, like the PUT method, it’s a good one to
know about.
Although a handful of other HTTP methods are defined under the specifications for
HTTP, these four constitute the entirety of what is used in just about any API you
will ever encounter.
Although some APIs do not use any authentication to operate (meaning anyone can
make an API call for free, without registering with the application first), many
modern APIs require some type of authentication before they can be used.
Some APIs require authentication in order to charge money per API call, or they
might offer their service on some sort of a monthly subscription basis. Others
authenticate in order to “rate limit” users (restrict them to a certain number of calls
per second, hour, or day), or to restrict the access of certain kinds of information or
types of API calls for some users. Other APIs might not place restrictions, but they
might want to keep track of which users are making which calls for marketing
All methods of API authentication generally revolve around the use of a token of
some sort, which is passed to the web server with each API call made. This token
is either provided to the user when the user registers and is a permanent fixture of
the user’s calls (generally in lower-security applications), or it can frequently
change, and is retrieved from the server using a username and password
For example, to make a call to the Echo Nest API in order to retrieve a list of songs
by the band Guns N’ Roses we would use:<your api key here>
This provides the server with an api_key value of what was provided to me on
registration, allowing the server to identify the requester as Ryan Mitchell, and
provide the requester with the JSON data.
In addition to passing tokens in the URL of the request itself, tokens might also be
passed to the server via a cookie in the request header. We will discuss headers in
greater detail later in this chapter as well as in Chapter 12, but by way of brief
example, they can be sent using the urllib package from previous chapters:
token = "<your api key>"
webRequest = urllib.request.Request("", headers={"token":token})
html = urlopen(webRequest)
As you saw in the FreeGeoIP example at the beginning of the chapter, an important
feature of APIs is that they have well-formatted responses. The most common types
of response formatting are eXtensible Markup Language (XML) and JavaScript
Object Notation (JSON).
In recent years, JSON has become vastly more popular than XML for a couple of
major reasons. First, JSON files are generally smaller than well-designed XML
files. Compare, for example, the XML data:
which clocks in at 98 characters, and the same data in JSON:
which is only 73 characters, or a whopping 36% smaller than the equivalent XML.
Of course, one could argue that the XML could be formatted like this:
<user firstname="ryan" lastname="mitchell" username="Kludgist"></user>
but this is considered bad practice because it doesn’t support deep nesting of data.
Regardless, it still requires 71 characters, about the same length as the equivalent
Another reason JSON is quickly becoming more popular than XML is simply due to
a shift in web technologies. In the past, it was more common for a server-side
script such as PHP or .NET to be on the receiving end of an API. Nowadays, it is
likely that a framework, such as Angular or Backbone, will be sending and
receiving API calls. Server-side technologies are somewhat agnostic as to the form
in which their data comes. But JavaScript libraries like Backbone find JSON easier
to handle.
Although most APIs still support XML output, we will be using JSON examples in
this book. Regardless, it is a good idea to familiarize yourself with both if you
haven’t already — they are unlikely to go away any time soon.
API Calls
The syntax of an API call can vary wildly from API to API, but there are a few
standard practices they often have in common. When retrieving data through a GET
request, the URL path describes how you would like to drill down into the
data, while the query parameters serve as filters or additional requests tacked onto
the search.
For example, in a hypothetical API, you might request the following to retrieve all
posts by the user with the ID 1234 during the month of August 2014:
Many other APIs use the path in order to specify an API version, the format you
would like your data in, and other attributes. For example, the following would
return the same data, using API version 4, formatted in JSON:
Other APIs require that you pass the formatting and API version information in as a
request parameter, like:
Echo Nest
The Echo Nest is a fantastic example of a company that is built on web scrapers.
Although some music-based companies, such as Pandora, depend on human
intervention to categorize and annotate music, The Echo Nest relies on automated
intelligence and information scraped from blogs and news articles in order to
categorize musical artists, songs, and albums.
Even better, this API is freely available for noncommercial use.3 You cannot use the
API without a key, but you can obtain a key by going to The Echo Nest “Create an
Account” page and registering with a name, email address, and username.
A Few Examples
The Echo Nest API is built around several basic content types: artists, songs,
tracks, and genres. Except for genres, these content types all have unique IDs,
which are used to retrieve information about them in various forms, through API
calls. For example, if I wanted to retrieve a list of songs performed by Monty
Python, I would make the following call to retrieve their ID (remember to replace
<your api key> with your own API key):<your api
This produces the following result:
{"response": {"status": {"version": "4.2", "code": 0, "message": "Suc
cess"}, "artists": [{"id": "AR5HF791187B9ABAF4", "name": "Monty Pytho
n"}, {"id": "ARWCIDE13925F19A33", "name": "Monty Python's SPAMALOT"},
{"id": "ARVPRCC12FE0862033", "name": "Monty Python's Graham Chapman"
I could also use that ID to query for a list of songs:<your api key>&id=
Which provides some of Monty Python’s hits, along with lesser-known recordings:
{"response": {"status": {"version": "4.2", "code": 0, "message": "Success"},
"start": 0, "total": 476, "songs": [{"id": "SORDAUE12AF72AC547", "title":
"Neville Shunt"}, {"id": "SORBMPW13129A9174D", "title": "Classic (Silbury Hill)
(Part 2)"}, {"id": "SOQXAYQ1316771628E", "title": "Famous Person Quiz (The
Final Rip Off Remix)"}, {"id": "SOUMAYZ133EB4E17E8", "title": "Always Look On
The Bright Side Of Life - Monty Python"}, ...]}}
Alternatively, I could make a single call using the name monty%20python in place
of the unique ID and retrieve the same information:<your api key>2&name=
Using that same ID, I can request a list of similar artists:<your api key>&id=
The results include other comedy artists such as Eric Idle, who was in the Monty
Python group:
{"response": {"status": {"version": "4.2", "code": 0, "message": "Suc
cess"}, "artists": [{"name": "Life of Brian", "id": "ARNZYOS1272BA7FF
38"}, {"name": "Eric Idle", "id": "ARELDIS1187B9ABC79"}, {"name": "Th
e Simpsons", "id": "ARNR4B91187FB5027C"}, {"name": "Tom Lehrer", "id"
: "ARJMYTZ1187FB54669"}, ...]}}
Note that while this list of similar artists contains some genuinely interesting
information (e.g., “Tom Lehrer”), the first result is “The Life of Brian,” from the
film soundtrack that Monty Python released. One of the hazards of using a database
culled from many sources with minimal human intervention is that you can get some
slightly strange results. This is something to keep in mind when creating your own
app using data from third-party APIs.
I’ve covered just a few examples of how The Echo Nest API can be used. For
complete documentation check out Echo Nest API Overview.
The Echo Nest sponsors many hackathons and programming projects focused on the
intersection of technology and music. If you need some inspiration, The Echo Nest
demo page is a good place to start.
Twitter is notoriously protective of its API and rightfully so. With over 230 million
active users and a revenue of over $100 million a month, the company is hesitant to
let just anyone come along and have any data they want.
Twitter’s rate limits (the number of calls it allows each user to make) fall into two
categories: 15 calls per 15-minute period, and 180 calls per 15-minute period,
depending on the type of call. For instance, you can make up to 12 calls a minute to
retrieve basic information about Twitter users, but only one call a minute to retrieve
lists of those users’ Twitter followers.4
Getting Started
In addition to rate limiting, Twitter has a more complicated authorization system
than do other APIs such as The Echo Nest, both for obtaining API keys and for
using those keys. In order to obtain an API key, you will, of course, need a Twitter
account; you can create one relatively painlessly on the signup page. In addition,
you will need to register a new “application” with Twitter on its developer site.
After this registration, you will be taken to a page containing your basic application
information, including your Consumer Key (Figure 4-1):
Figure 4-1. Twitter’s Application Settings page provides basic information about your
new application
If you click on the “manage keys and access tokens” page, you will be directed to a
page containing further information (Figure 4-2):
Figure 4-2. You’ll need your secret key to use the Twitter API
This page also contains a button to automatically regenerate your keys should your
secret key somehow become public (e.g., if you accidentally publish it as an
illustration in a technical book).
A Few Examples
Twitter’s authentication system, which uses OAuth, is fairly complicated; it’s best
to wrap it in one of the readily available libraries rather than try to keep track of it
yourself. Because of the relative complexity of working with Twitter’s API “by
hand,” this section’s samples will focus on using Python code to interact with the
API, rather than working with the API itself.
As of this writing, there are many Python 2.x libraries used to interact with Twitter
but relatively few Python 3.x libraries available. Fortunately, one of the best
Twitter libraries (appropriately called Twitter) is available for Python 3.x. You can
download it from the Python Twitter Tools page and install it in the usual way:
$cd twitter-x.xx.x
$python install
Twitter Credential Permissions
By default, application access tokens are given read-only permission. This is fine
for most purposes, unless you want your application to actually make a tweet on
your behalf.
To change the permissions of your tokens to read/write, go to the Permissions
tab on Twitter’s Application Management panel. The tokens will then need to be
regenerated for the permission update to take place.
Similarly, you can update the token permissions to be able to access direct
messages to your Twitter account if necessary for your application. Be careful,
however; you should give tokens only the permissions they absolutely need. In
general, it’s good practice to generate multiple sets of tokens for multiple
applications, rather than reuse too-powerful tokens for applications that don’t
need all their permissions.
In our first exercise, we search for specific tweets. The following code connects to
the Twitter API and prints a JSON list of tweets containing the hashtag #python.
Remember to replace the strings in the OAuth line with your actual credentials:
from twitter import Twitter
t = Twitter(auth=OAuth(<Access Token>,<Access Token Secret>,
<Consumer Key>,<Consumer Secret>))
pythonTweets = = "#python")
Although the output of this script might be overwhelming, keep in mind that you’re
getting a lot of information per tweet: the date and time the tweet was made, details
about any retweets or favorites, and details about the user’s account and profile
image. Although you might be looking for just a subset of this data, the Twitter API
was designed for web developers who want to display the tweets received through
the API on their own websites, so there’s a lot of extra stuff involved!
You can see an example output of a single tweet when making a status update
through the API:
from twitter import *
t = Twitter(auth=OAuth("Access Token", "Access Token Secret",
"Consumer Key", "Consumer Secret"))
statusUpdate = t.statuses.update(status='Hello, world!')
Here is JSON’s description of the tweet:
{'created_at': 'Sun Nov 30 07:23:39 +0000 2014', 'place': None, 'in_reply_to_scr
een_name': None, 'id_str': '538956506478428160', 'in_reply_to_user_id: None,'lan
g': 'en', 'in_reply_to_user_id_str': None, 'user': {'profile_sidebar_border_colo
r': '000000', 'profile_background_image_url': '
ground_images/497094351076347904/RXn8MUlD.png', 'description':'Software Engine
[email protected], Masters student @HarvardEXT, @OlinCollege graduate, writer @OReil
lyMedia. Really tall. Has pink hair. Female, despite the name.','time_zone': 'Ea
stern Time (US & Canada)', 'location': 'Boston, MA', 'lang': 'en', 'url': 'http:
//', 'profile_location': None, 'name': 'Ryan Mitchell', 'screen_n
ame': 'Kludgist', 'protected': False, 'default_profile_image': False, 'id_str':
'14983299', 'favourites_count': 140, 'contributors_enabled': False, 'profile_use
_background_image': True, 'profile_background_image_url_https': 'https://pbs.twi', 'profile_side
bar_fill_color': '889654', 'profile_link_color': '0021B3', 'default_profile': Fa
lse, 'statuses_count': 3344, 'profile_background_color': 'FFFFFF', 'profile_imag
e_url': '
jpeg', 'profile_background_tile': True, 'id': 14983299, 'friends_count': 409, 'p
rofile_image_url_https': '
8/XJh_d5f5_normal.jpeg', 'following': False, 'created_at': 'Mon Jun 02 18:35:1
8 +0000 2008', 'is_translator': False, 'geo_enabled': True,
ed': False, 'follow_request_sent': False, 'followers_count': 2085, 'utc_offset'
: -18000, 'verified': False, 'profile_text_color': '383838', 'notifications': F
alse, 'entities': {'description': {'urls': []}, 'url': {'urls': [{'indices': [
0, 22], 'url': '', 'expanded_url': 'http://ryanemitchell.
com', 'display_url': ''}]}}, 'listed_count': 22, 'profile_banne
r_url': ''}, 'retweeted
': False, 'in_reply_to_status_id_str': None, 'source': '<a href="http://ryanemit" rel="nofollow">PythonScraping</a>', 'favorite_count': 0, 'text':
o,world!', 'truncated': False, 'id': 538956506478428160, 'retweet_count': 0, 'fa
vorited': False, 'in_reply_to_status_id': None, 'geo': None, 'entities':
entions': [], 'hashtags': [], 'urls': [], 'symbols': []}, 'coordinates': None, '
contributors': None}
Yes, this is the result of sending a single tweet. Sometimes I think Twitter limits
access to its API because of the bandwidth required to respond to every request!
For any request that retrieves a list of tweets, you can limit the number of tweets
received by specifying a count:
pythonStatuses = t.statuses.user_timeline(screen_name="montypython", count=5)
In this case, we are asking for the last five tweets that were posted to
@montypython’s timeline (this includes any retweets they might have made).
Although these three examples (searching for tweets, retrieving the tweets of a
specific user, and posting your own tweets) cover a great deal of what most people
do with the Twitter API, the capabilities of the Twitter Python library are far more
numerous. You can search and manipulate Twitter lists, follow and unfollow users,
look up profile information for users, and more. The complete documentation can
be found on GitHub.
Google APIs
Google has one of the most comprehensive, easy-to-use collections of APIs on the
Web today. Whenever you’re dealing with some sort of basic subject, such as
language translation, geolocation, calendars, or even genomics, Google has an API
for it. Google also has APIs for many of its popular apps, such as Gmail, YouTube,
and Blogger.
There are two main reference pages for browsing Google APIs. The first is the
Products page, which serves as an organized repository of its APIs, software
development kits, and other projects that might be of interest to software
developers. The other is the APIs console, which provides a convenient interface
to turn API services on and off, view rate limits and usage at a glance, and even
spin up a Google-powered cloud computing instance if you feel like it.
Most of Google’s APIs are free although some, such as its search API, require a
paid license. Google is fairly liberal with its collection of free APIs allowing from
250 requests per day to 20,000,000 requests per day with a basic account. There is
also the option to raise the rate limits on some of the APIs by verifying your identity
with a credit card (the card is not charged). For example, the Google Places API
has a basic rate limit of 1,000 requests per 24-hour period, but this can be raised to
150,000 requests by verifying your identity. For more information, see the Usage
Limits and Billing page.
Getting Started
If you have a Google account, you can view the list of available APIs and create an
API key using the Google Developers Console. If you don’t have a Google account,
signing up is easy to do via the Create Your Google Account page.
Once you’ve logged in or created your account, you can view your account
credentials, including API keys, on the API console page; click “Credentials” in the
left menu (Figure 4-3):
Figure 4-3. Google’s API Credentials page
On the Credentials page, you can click the “Create new Key” button to create an
API key, with the option to limit the referring IP address or URL that can use it. To
create an API key that can be used from any URL or IP address, simply leave the
“Accept Request From These Server IP Addresses” box blank when prompted.
Keep in mind, however, that it is important to keep your key secret if you do not
limit the originating IP address — any calls made using your API key will count
against your rate limit, even if they are unauthorized.
You can create multiple API keys as well. For example, you might have a separate
API key for each project, or for each web domain that you own. However,
Google’s API rate limits are per account, not per key, so although this might be a
handy way to manage your API permissions, it will not work to get around rate
A Few Examples
Google’s most popular (and in my opinion most interesting) APIs can be found in
its collection of Maps APIs. You might be familiar with this feature through the
embeddable Google Maps found on many websites. However, the Maps APIs go
far beyond embedding maps — you can resolve street addresses to
latitude/longitude coordinates, get the elevation of any point on Earth, create a
variety of location-based visualizations, and get time zone information for an
arbitrary location, among other bits of information.
When trying these examples out on your own, remember to activate each needed
API from Google’s API console before sending a request. Google uses these API
activations for its metrics (“how many users does this API have?”), so you’re
required to explicitly activate the API before using.
Using Google’s Geocode API, you can make a simple GET request from your
browser to resolve any street address (in this case, the Boston Museum of Science)
to a latitude and longitude:
+MA+02114&key=<your API key>
"results" : [ { "address_components" : [ { "long_name" : "Museum Of Science
way", "short_name" : "Museum Of Science Driveway", "types" : [ "route" ] }, { "l
ong_name" : "Boston", "short_name" : "Boston", "types" : [ "locality", "politica
l" ] }, { "long_name" : "Massachusetts", "short_name" : "MA", "types" : [ "admin
istrative_area_level_1", "political" ] }, { "long_name" : "United States", "shor
t_name" : "US", "types" : [ "country", "political" ] }, { "long_name" : "0211
4", "short_name" : "02114", "types" : [ "postal_code" ] } ], "formatted_address"
: "Museum Of Science Driveway, Boston, MA 02114, USA", "geometry" : { "bounds" :
{ "northeast" : { "lat" : 42.368454, "lng" : -71.06961339999999 }, "southwest" :
{ "lat" : 42.3672568, "lng" : -71.0719624 } }, "location" : { "lat" : 42.3677994
, "lng" : -71.0708078 }, "location_type" : "GEOMETRIC_CENTER", "viewport" : { "n
ortheast" : { "lat" : 42.3692043802915, "lng" : -71.06943891970849 }, "southwest
" : { "lat" : 42.3665064197085, "lng" : -71.0721368802915 } } }, "types" : [ "ro
ute" ] } ], "status" : "OK" }
Note that the address we send the API doesn’t have to be especially well formatted.
Google being Google, the Geocode API is excellent at taking addresses that are
missing postal codes or state information (or even misspelled addresses), and
giving you back its best guess at what the correct address is. For instance, the
deplorably spelled request argument 1+Skience+Park+Bostton+MA (without even
a postal code) returns the same result.
I have used the Geocode API on several occasions, not only to format user-entered
addresses on a website, but to crawl the Web looking for things that look like
addresses, and using the API to reformat them into something easier to store and
To get the time zone information for our newly found latitude and longitude, you can
use the Time zone API:,-71.0708
078&timestamp=1412649030&key=<your API key>
Which provides the response:
{ "dstOffset" : 3600, "rawOffset" : -18000, "status" : "OK", "timeZon
eId" : "America/New_York", "timeZoneName" : "Eastern Daylight Time" }
A Unix timestamp is required to make a request to the Time Zone API. This allows
Google to provide you with a time zone appropriately adjusted for daylight saving
time. Even in areas where the time zone is not affected by the time of year (such as
Phoenix, which does not use daylight saving time), the timestamp is still required in
an API request.
To round out this all-too-brief tour of the Google Maps API collection, you can also
fetch the elevation for this latitude and longitude:,-71.070
8078&key=<your API key>
This returns the elevation in meters above sea level, along with the “resolution,”
which indicates the distance in meters of the farthest data point this elevation was
interpolated from. A smaller value for resolution indicates a higher degree of
accuracy in the given elevation:
{ "results" : [ { "elevation" : 5.127755641937256, "location" : { "la
t" : 42.3677994, "lng" : -71.0708078 }, "resolution" : 9.543951988220
215 } ], "status" : "OK" }
Parsing JSON
In this chapter, we’ve looked at various types of APIs and how they function, and
we’ve looked at some sample JSON responses from these APIs. Now let’s look at
how we can parse and use this information.
At the beginning of the chapter I used the example of the IP, which
resolves IP addresses to physical addresses:
I can take the output of this request and use Python’s JSON parsing functions to
decode it:
import json
from urllib.request import urlopen
def getCountry(ipAddress):
response = urlopen(""+ipAddress).read()
responseJson = json.loads(response)
return responseJson.get("country_code")
This prints out the country code for the IP address:
The JSON parsing library used is part of Python’s core library. Just type in import
json at the top, and you’re all set! Unlike many languages that might parse JSON
into a special JSON object or JSON node, Python uses a more flexible approach
and turns JSON objects into dictionaries, JSON arrays into lists, JSON strings into
strings, and so forth. In this way, it makes it extremely easy to access and
manipulate values stored in JSON.
The following gives a quick demonstration of how Python’s JSON library handles
the different values that might be encountered in a JSON string:
import json
jsonString = '{"arrayOfNums":[{"number":0},{"number":1},{"number":2}],
jsonObj = json.loads(jsonString)
The output of which is:
[{'number': 0}, {'number': 1}, {'number': 2}]
{'number': 1}
Line 1 is a list of dictionary objects, line 2 is a dictionary object, line 3 is an
integer (the sum of the integers accessed in the dictionaries), and line 4 is a string.
Bringing It All Back Home
Although the raison d'être of many modern web applications is to take existing data
and format it in a more appealing way, I would argue that this isn’t very interesting
thing to do in most instances. If you’re using an API as your only data source, the
best you can do is merely copy someone else’s database that already exists, and
which is, essentially, already published. What can be far more interesting is to take
two or more data sources and combine them in a novel way, or use an API as a tool
to look at scraped data from a new perspective.
Let’s look at one example of how data from APIs can be used in conjunction with
web scraping: to see which parts of the world contribute the most to Wikipedia.
If you’ve spent much time on Wikipedia, you’ve likely come across an article’s
revision history page, which displays a list of recent edits. If users are logged into
Wikipedia when they make the edit, their username is displayed. If they are not
logged in, their IP address is recorded, as shown in Figure 4-4.
Figure 4-4. The IP address of an anonymous editor on the revision history page for
Wikipedia’s Python entry
The IP address outlined on the history page is By using the API, as of this writing (IP address can occasionally shift
geographically) that IP address is from Quezon, Phillipines.
This information isn’t all that interesting on its own, but what if we could gather
many, many, points of geographic data about Wikipedia edits and where they occur?
A few years ago I did just that and used Google’s Geochart library to create an
interesting chart ( that shows where edits on the Englishlanguage Wikipedia, along with the Wikipedias written in other languages, originate
from (Figure 4-5).
Figure 4-5. Visualization of Wikipedia edits created using Google’s Geochart library
Creating a basic script that crawls Wikipedia, looks for revision history pages, and
then looks for IP addresses on those revision history pages isn’t difficult. Using
modified code from Chapter 3, the following script does just that:
from urllib.request import urlopen
from bs4 import BeautifulSoup
import datetime
import random
import re
def getLinks(articleUrl):
html = urlopen(""+articleUrl)
bsObj = BeautifulSoup(html)
return bsObj.find("div", {"id":"bodyContent"}).findAll("a",
def getHistoryIPs(pageUrl):
#Format of revision history pages is:
pageUrl = pageUrl.replace("/wiki/", "")
historyUrl = ""
print("history url is: "+historyUrl)
html = urlopen(historyUrl)
bsObj = BeautifulSoup(html)
#finds only the links with class "mw-anonuserlink" which has IP addresses
#instead of usernames
ipAddresses = bsObj.findAll("a", {"class":"mw-anonuserlink"})
addressList = set()
for ipAddress in ipAddresses:
return addressList
links = getLinks("/wiki/Python_(programming_language)")
while(len(links) > 0):
for link in links:
historyIPs = getHistoryIPs(link.attrs["href"])
for historyIP in historyIPs:
newLink = links[random.randint(0, len(links)-1)].attrs["href"]
links = getLinks(newLink)
This program uses two main functions: getLinks (which was also used in
Chapter 3), and the new getHistoryIPs, which searches for the contents of all
links with the class mw-anonuserlink (indicating an anonymous user with an IP
address, rather than a username) and returns it as a set.
Up until this point in the book, I’ve relied almost exclusively on two Python data
structures to store multiple pieces of data: lists and dictionaries. With both of these
options, why use a set?
Python sets are unordered, meaning you shouldn’t reference a specific position in the
set and expect to get the value you’re looking for. The order in which you add items to
the set is not necessarily the order in which you’ll get them back. One nice property of
sets that I’m taking advantage of in the code sample is that they won’t hold multiples
of the same item. If you add a string to a set when that string already exists, it will not
be duplicated. In this way, I can quickly get a list of only the unique IP addresses on
the revision history page, disregarding multiple edits by the same user.
A couple of things to keep in mind when deciding between sets and lists in code that
needs to scale: although lists are slightly faster to iterate over, sets are slightly faster for
doing lookups (determining whether an object exists in the set or not).
This code also uses a somewhat arbitrary (yet effective for the purposes of this
example) search pattern to look for articles from which to retrieve revision
histories. It starts by retrieving the histories of all Wikipedia articles linked to by
the starting page (in this case, the article on the Python programming language).
Afterward, it selects a new starting page randomly, and retrieves all revision
history pages of articles linked to by that page. It will continue until it hits a page
with no links.
Now that we have code that retrieves IP addresses as a string, we can combine this
with the getCountry function from the previous section in order to resolve these IP
addresses to countries. I modified getCountry slightly, in order to account for
invalid or malformed IP addresses that will result in a “404 Not Found” error (as
of this writing, does not resolve IPv6, for instance, which might
trigger such an error):
def getCountry(ipAddress):
response = urlopen(""
except HTTPError:
return None
responseJson = json.loads(response)
return responseJson.get("country_code")
links = getLinks("/wiki/Python_(programming_language)")
while(len(links) > 0):
for link in links:
historyIPs = getHistoryIPs(link.attrs["href"])
for historyIP in historyIPs:
country = getCountry(historyIP)
if country is not None:
print(historyIP+" is from "+country)
newLink = links[random.randint(0, len(links)-1)].attrs["href"]
links = getLinks(newLink)
The complete executable code can be found at Here’s a sample output:
------------------history url is:
paradigm&action=history is from US is from GB is from SA is from US is from CA is from US is from PK is from GB is from JM is from PH is from US is from US
More About APIs
In this chapter, we’ve looked at a few ways that modern APIs are commonly used
to access data on the Web, in particular uses of APIs that you might find useful in
web scraping. However, I’m afraid that this does not do justice to the broad scope
that “software sharing data with disparate software” entails.
Because this book is about web scraping and is not intended as a general guide on
data collection, I can only point you in the direction of some excellent resources for
further research on the subject, if you need it.
Leonard Richardson, Mike Amundsen, and Sam Ruby’s RESTful Web APIs
provides a strong overview of the theory and practice of using APIs on the Web. In
addition, Mike Amundsen has a fascinating video series, Designing APIs for the
Web, that teaches you how to create your own APIs, a useful thing to know if you
decide to make your scraped data available to the public in a convenient format.
Web scraping and web APIs might seem like very different subjects at first glance.
However, I hope that this chapter has shown that they are complementary skills on
the same continuum of data collection. In some sense, using a web API can even be
thought of as a subset of the subject of web scraping. After all, you are ultimately
writing a script that collects data from a remote web server, and parsing it into a
usable format, as you would do with any web scraper.
This API resolves IP addresses to geographic locations and is one I’ll be using
later in the chapter as well. You can learn more about it at
In reality, many APIs use POST requests in lieu of PUT requests when updating
information. Whether a new entity is created or an old one is merely updated is
often left to how the API request itself is structured. However, it’s still good to
know the difference, and you will often encounter PUT requests in commonly used
See The Echo Nest Licensing page restriction requirements details.
For a complete list of rate limits, see
Chapter 5. Storing Data
Although printing out to the terminal is a lot of fun, it’s not incredibly useful when it
comes to data aggregation and analysis. In order to make the majority of web
scrapers remotely useful, you need to be able to save the information that they
In this chapter, we will look at three main methods of data management that are
sufficient for almost any imaginable application. Do you need to power the backend
of a website or create your own API? You’ll probably want your scrapers to write
to a database. Need a fast and easy way to collect some documents off the Internet
and put them on your hard drive? You’ll probably want to create a file stream for
that. Need occasional alerts, or aggregated data once a day? Send yourself an
Above and beyond web scraping, the ability to store and interact with large
amounts of data is incredibly important for just about any modern programming
application. In fact, the information in this chapter is necessary for implementing
many of the examples in later sections of the book. I highly recommend that you at
least skim it if you’re unfamiliar with automated data storage.
Media Files
There are two main ways to store media files: by reference, and by downloading
the file itself. You can store a file by reference simply by storing the URL that the
file is located at. This has several advantages:
Scrapers run much faster, and require much less bandwidth, when they don’t
have to download files.
You save space on your own machine by storing only the URLs.
It is easier to write code that only stores URLs and doesn’t need to deal with
additional file downloads.
You can lessen the load on the host server by avoiding large file downloads.
Here are the disadvantages:
Embedding these URLs in your own website or application is known
as hotlinking and doing it is a very quick way to get you in hot water on the
You do not want to use someone else’s server cycles to host media for your own
The file hosted at any particular URL is subject to change. This might lead to
embarrassing effects if, say, you’re embedding a hotlinked image on a public
blog. If you’re storing the URLs with the intent to store the file later, for further
research, it might eventually go missing or be changed to something completely
irrelevant at a later date.
Real web browsers do not just request a page’s HTML and move on — they
download all of the assets required by the page as well. Downloading files can
help make your scraper look like an actual human is browsing the site, which
can be an advantage.
If you’re debating over whether to store a file or simply a URL to a file, you should
ask yourself whether you’re likely to actually view or read that file more than once
or twice, or if this database of files is just going to be sitting around gathering
electronic dust for most of its life. If the answer is the latter, it’s probably best to
simply store the URL. If it’s the former, read on!
In Python 3.x, urllib.request.urlretrieve can be used to download files from
any remote URL:
from urllib.request import urlretrieve
from urllib.request import urlopen
from bs4 import BeautifulSoup
html = urlopen("")
bsObj = BeautifulSoup(html)
imageLocation = bsObj.find("a", {"id": "logo"}).find("img")["src"]
urlretrieve (imageLocation, "logo.jpg")
This downloads the logo from and stores it as logo.jpg
in the same directory that the script is running from.
This works well if you only need to download a single file and know what to call it
and what the file extension is. But most scrapers don’t download a single file and
call it a day. The following will download all internal files, linked to by any tag’s
src attribute, from the home page of
import os
from urllib.request import urlretrieve
from urllib.request import urlopen
from bs4 import BeautifulSoup
downloadDirectory = "downloaded"
baseUrl = ""
def getAbsoluteURL(baseUrl, source):
if source.startswith("http://www."):
url = "http://"+source[11:]
elif source.startswith("http://"):
url = source
elif source.startswith("www."):
url = source[4:]
url = "http://"+source
url = baseUrl+"/"+source
if baseUrl not in url:
return None
return url
def getDownloadPath(baseUrl, absoluteUrl, downloadDirectory):
path = absoluteUrl.replace("www.", "")
path = path.replace(baseUrl, "")
path = downloadDirectory+path
directory = os.path.dirname(path)
if not os.path.exists(directory):
return path
html = urlopen("")
bsObj = BeautifulSoup(html)
downloadList = bsObj.findAll(src=True)
for download in downloadList:
fileUrl = getAbsoluteURL(baseUrl, download["src"])
if fileUrl is not None:
urlretrieve(fileUrl, getDownloadPath(baseUrl, fileUrl, downloadDirectory))
Run with Caution
You know all those warnings you hear about downloading unknown files off the
Internet? This script downloads everything it comes across to your computer’s
hard drive. This includes random bash scripts, .exe files, and other potential
Think you’re safe because you’d never actually execute anything sent to your
downloads folder? Especially if you run this program as an administrator, you’re
asking for trouble. What happens if you run across a file on a website that sends
itself to ../../../../usr/bin/python? The next time you run a Python script from the
command line, you could actually be deploying malware on your machine!
This program is written for illustrative purposes only; it should not be randomly
deployed without more extensive filename checking, and it should only be run in
an account with limited permissions. As always, backing up your files, not
storing sensitive information on your hard drive, and a little common sense go a
long way.
This script uses a lambda function (introduced in Chapter 2) to select all tags on the
front page that have the src attribute, then cleans and normalizes the URLs to get an
absolute path for each download (making sure to discard external links). Then, each
file is downloaded to its own path in the local folder downloaded on your own
Notice that Python’s os module is used briefly to retrieve the target directory for
each download and create missing directories along the path if needed. The os
module acts as an interface between Python and the operating system, allowing it to
manipulate file paths, create directories, get information about running processes
and environment variables, and many other useful things.
Storing Data to CSV
CSV, or comma-separated values, is one of the most popular file formats in which
to store spreadsheet data. It is supported by Microsoft Excel and many other
applications because of its simplicity. The following is an example of a perfectly
valid CSV file:
As with Python, whitespace is important here: Each row is separated by a newline
character, while columns within the row are separated by commas (hence the
“comma-separated”). Other forms of CSV files (sometimes called “characterseparated value” files) use tabs or other characters to separate rows, but these file
formats are less common and less widely supported.
If you’re looking to download CSV files directly off the Web and store them
locally, without any parsing or modification, you don’t need this section. Simply
download them like you would any other file and save them with the CSV file
format using the methods described in the previous section.
Modifying a CSV file, or even creating one entirely from scratch, is extremely easy
with Python’s csv library:
import csv
csvFile = open("../files/test.csv", 'wt')
writer = csv.writer(csvFile)
writer.writerow(('number', 'number plus 2', 'number times 2'))
for i in range(10):
writer.writerow( (i, i+2, i*2))
A precautionary reminder: file creation in Python is fairly bullet-proof. If
../files/test.csv does not already exist, Python will create the directory and the file
automatically. If it already exists, Python will overwrite test.csv with the new data.
After running, you should see a CSV file:
number,number plus 2,number times 2
One common web-scraping task is to retrieve an HTML table and write it as a CSV
file. Wikipedia’s Comparison of Text Editors provides a fairly complex HTML
table, complete with color coding, links, sorting, and other HTML garbage that
needs to be discarded before it can be written to CSV. Using BeautifulSoup and the
get_text() function copiously, you can do that in fewer than 20 lines:
import csv
from urllib.request import urlopen
from bs4 import BeautifulSoup
html = urlopen("")
bsObj = BeautifulSoup(html)
#The main comparison table is currently the first table on the page
table = bsObj.findAll("table",{"class":"wikitable"})[0]
rows = table.findAll("tr")
csvFile = open("../files/editors.csv", 'wt')
writer = csv.writer(csvFile)
for row in rows:
csvRow = []
for cell in row.findAll(['td', 'th']):
This script is great to integrate into scrapers if you encounter many HTML tables that
need to be converted to CSV files, or many HTML tables that need to be collected into
a single CSV file. However, if you just need to do it once, there’s a better tool for that:
copying and pasting. Selecting and copying all of the content of an HTML table and
pasting it into Excel will get you the CSV file you’re looking for without running a
The result should be a well-formatted CSV file saved locally, under
../files/editors.csv — perfect for sending and sharing with folks who haven’t quite
gotten the hang of MySQL yet!
MySQL (officially pronounced “My es-kew-el,” although many say, “My Sequel”)
is the most popular open source relational database management system today.
Somewhat unusually for an open source project with large competitors, its
popularity has historically been neck and neck with the two other major closed
source database systems: Microsoft’s SQL Server and Oracle’s DBMS.
Its popularity is not without cause. For most applications, it’s very hard to go
wrong with MySQL. It’s a very scaleable, robust, and full-featured DBMS, used by
top websites: YouTube,1 Twitter,2 and Facebook,3 among many others.
Because of its ubiquity, price (“free” is a pretty great price), and out-of-box
usability, it makes a fantastic database for web-scraping projects, and we will use
it throughout the remainder of this book.
“Relational data” is data that has relations. Glad we cleared that up!
Just kidding! When computer scientists talk about relational data, they’re referring to
data that doesn’t exist in a vacuum — it has properties that relate it to other pieces of
data. For example, “User A goes to school at Institution B,” where User A can be
found in the “users” table in the database and Institution B can be found in the
“institutions” table in the database.
Later in this chapter, we’ll take a look at modeling different types of relations and how
to store data in MySQL (or any other relational database) effectively.
Installing MySQL
If you’re new to MySQL, installing a database might sound a little intimidating (if
you’re an old hat at it, feel free to skip this section). In reality, it’s as simple as
installing just about any other kind of software. At its core, MySQL is powered by
a set of data files, stored on your server or local machine, that contain all the
information stored in your database. The MySQL software layer on top of that
provides a convenient way of interacting with the data, via a command-line
interface. For example, the following command will dig through the data files and
return a list of all users in your database whose first name is “Ryan”:
SELECT * FROM users WHERE firstname = "Ryan"
If you’re on Linux, installing MySQL is as easy as:
$sudo apt-get install mysql-server
Just keep an eye on the installation process, approve the memory requirements, and
enter a new password for your new root user when prompted.
For Mac OS X and Windows, things are a little trickier. If you haven’t already,
you’ll need to create an Oracle account before downloading the package.
If you’re on Mac OS X, you’ll need to first get the installation package.
Select the .dmg package, and log in with or create your Oracle account to
download the file. After opening, you should be guided through a fairly
straightforward installation wizard (Figure 5-1):
Figure 5-1. The Mac OS X MySQL installer
The default installation steps should suffice, and for the purposes of this book, I
will assume you have a default MySQL installation.
If downloading and running an installer seems a little tedious, you can always
install the package manager, Homebrew. With Homebrew installed, you can also
install MySQL by running:
$brew install mysql
Homebrew is a great open source project with very good Python package
integration. In fact, most of the third-party Python modules used in this book can be
installed very easily with Homebrew. If you don’t have it already, I highly
recommend checking it out!
Once MySQL is installed on Mac OS X, you can start the MySQL server as
$cd /usr/local/mysql
$sudo ./bin/mysqld_safe
On Windows, installing and running MySQL is slightly more complicated, but the
good news is that there’s a convenient installer that simplifies the process. Once
downloaded, it will guide you through the steps you’ll need to take (see Figure 52):
Figure 5-2. The MySQL Windows Installer
You should be able to install MySQL using the default selections, with one
exception: on the Setup Type page, I recommend you choose “Server Only” to
avoid installing a lot of additional Microsoft software and libraries.
From there, you should be able to use the default installation settings and follow the
prompts to start your MySQL server.
Some Basic Commands
Once your MySQL server is running, you have many options for interacting with the
database. There are plenty of software tools that act as an intermediary so that you
don’t have to deal with MySQL commands (or at least deal with them less often).
Tools like phpMyAdmin and MySQL Workbench can make it easy to quickly view,
sort, and insert data. However, it’s still important to know your way around the
command line.
Except for variable names, MySQL is case insensitive; for example, SELECT is the
same as sElEcT. However, by convention, all MySQL keywords are in all caps
when you are writing a MySQL statement. Conversely, most developers prefer to
name their tables and databases in lowercase, although this standard is often
When you first log into MySQL, there are no databases to add data to. You can
create one:
Because every MySQL instance can have multiple databases, before we can start
interacting with a database we need to specify to MySQL which database we want
to use:
>USE scraping;
From this point on (at least until we close the MySQL connection or switch to
another database), all commands entered will be run against the new “scraping”
That all seems pretty straightforward. It must be similarly easy to create a table in
the database, right? Let’s try to create a table to store a collection of scraped web
This results in the error:
ERROR 1113 (42000): A table must have at least 1 column
Unlike a database, which can exist without any tables, a table in MySQL cannot
exist without columns. In order to define columns in MySQL, you must enter them in
a comma-delimited list, within parentheses, after the CREATE TABLE <tablename>
Each column definition has three parts:
The name (id, title, created, etc.)
The variable type (BIGINT(7), VARCHAR, TIMESTAMP)
Optionally, any additional attributes (NOT
At the end of the list of columns, you must define a table’s “key.” MySQL uses keys
to organize the content in the table for fast lookups. Later in this chapter, I’ll
describe how to use these keys to your advantage for speedier databases, but for
now, using a table’s id column as the key is generally the best way to go.
After the query executes, you can see what the structure of the table looks like at
any time by using DESCRIBE:
> DESCRIBE pages;
| Field
| Type
| Null | Key | Default
| Extra
| id
| bigint(7)
| NO
| auto_increment |
| title
| varchar(200)
| YES |
| content | varchar(10000) | YES |
| created | timestamp
| NO
4 rows in set (0.01 sec)
Of course, this is still an empty table. You can insert some test data into the pages
table, by using the following line:
> INSERT INTO pages (title, content) VALUES ("Test page title", "This is some te
st page content. It can be up to 10,000 characters long.");
Notice that although the table has four columns (id, title, content, created),
you need to define only two of them (title and content) in order to insert a row.
That’s because the id column is autoincremented (MySQL automatically adds a 1
each time a new row is inserted) and generally can take care of itself. In addition,
the timestamp column is set to contain the current time as a default.
Of course, we can override these defaults:
INSERT INTO pages (id, title, content, created) VALUES (3, "Test page title", "
This is some test page content. It can be up to 10,000 characters long.", "201409-21 10:25:32");
As long as the integer you provide for the id column doesn’t already exist in the
database, this override will work perfectly fine. However, it is generally bad
practice to do this; it’s best to let MySQL handle the id and timestamp columns
unless there is a compelling reason to do it differently.
Now that we have some data in the table, you can use a wide variety of methods to
select this data. Here are a few examples of SELECT statements:
>SELECT * FROM pages WHERE id = 2;
This statement tells MySQL, “Select all from pages where id equals 2.” The
asterisk (*) acts as a wildcard, returning all of the rows where the clause (where
id equals 2) is true. It returns the second row in the table, or an empty result if
there is no row with an id of 2. For example, the following case-insensitive query
returns all the rows where the title field contains “test” (the % symbol acts as a
wildcard in MySQL strings):
>SELECT * FROM pages WHERE title LIKE "%test%";
But what if you have a table with many columns, and you only want a particular
piece of data returned? Rather than selecting all, you can do something like this:
>SELECT id, title FROM pages WHERE content LIKE "%page content%";
This returns just the id and title where the content contains the phrase “page
statements have much the same syntax as SELECT statements:
>DELETE FROM pages WHERE id = 1;
For this reason it is a good idea, especially when working on important databases
that can’t be easily restored, to write any DELETE statements as a SELECT statement
first (in this case, SELECT * FROM pages WHERE id = 1), test to make sure only
the rows you want to delete are returned, and then replace SELECT * with DELETE.
Many programmers have horror stories of miscoding the clause on a DELETE
statement, or worse, leaving it off entirely when they were in a hurry, and ruining
customer data. Don’t let it happen to you!
Similar precautions should be taken with UPDATE statements:
>UPDATE pages SET title="A new title", content="Some new content" WHERE id=2;
For the purposes of this book, we will be working only with simple MySQL
statements, doing basic selecting, inserting, and updating. If you’re interested in
learning more commands and techniques with this powerful database tool, I
recommend Paul DuBois’s MySQL Cookbook.
Integrating with Python
Unfortunately, Python support for MySQL is not built in. However, there are a good
number of open source libraries you can use, both with Python 2.x and Python 3.x,
that allow you to interact with a MySQL database. One of the most popular of these
is PyMySQL.
As of this writing, the current version of PyMySQL is 0.6.2, and you can download
and install this version with the following commands:
$ curl -L | tar xz
$ cd PyMySQL-PyMySQL-f953785/
$ python install
Remember to check for the latest version of PyMySQL on the website and update
the version number in the first line of this command, as needed.
After installation, you should have access to the PyMySQL package automatically,
and while your local MySQL server is running, you should be able to execute the
following script successfully (remember to add the root password for your
import pymysql
conn = pymysql.connect(host='', unix_socket='/tmp/mysql.sock',
user='root', passwd=None, db='mysql')
cur = conn.cursor()
cur.execute("USE scraping")
cur.execute("SELECT * FROM pages WHERE id=1")
There are two new types of objects at work in this example: the Connection object
(conn), and the Cursor object (cur).
The connection/cursor model is commonly used in database programming, although
some users might find it tricky to differentiate the two at first. The connection is
responsible for, well, connecting to the database of course, but also sending the
database information, handling rollbacks (when a query or set of queries needs to
be aborted and the database needs to be returned to its previous state), and creating
new cursor objects.
A connection can have many cursors. A cursor keeps track of certain state
information, such as which database it is using. If you have multiple databases and
need to write information across all of them, you might have multiple cursors to
handle this. A cursor also contains the results of the latest query it has executed. By
calling functions on the cursor, such as cur.fetchone(), you can access this
It is important that both the cursor and the connection are closed after you are done
using them. Not doing this might result in connection leaks, a buildup of unclosed
connections that are no longer being used but the software isn’t able to close
because it’s under the impression that you might still use them. This is the sort of
thing that brings databases down all the time, so remember to close your
The most common thing you’ll probably want to do, starting out, is to be able to
store your scraping results in a database. Let’s take a look at how this could be
done, using a previous example: the Wikipedia scraper.
Dealing with Unicode text can be tough when web scraping. By default, MySQL
does not handle Unicode. Fortunately, you can turn on this feature (just keep in mind
that doing so will increase the size of your database). Because we’re bound to run
into a variety of colorful characters on Wikipedia, now is a good time to tell your
database to expect some Unicode:
ALTER DATABASE scraping CHARACTER SET = utf8mb4 COLLATE = utf8mb4_unicode_ci;
ALTER TABLE pages CONVERT TO CHARACTER SET utf8mb4 COLLATE utf8mb4_unicode_ci;
ALTER TABLE pages CHANGE content content VARCHAR(10000) CHARACTER SET utf8mb4 CO
LLATE utf8mb4_unicode_ci;
These four lines change the following: the default character set for the database, for
the table, and for both of the two columns, from utf8mb4 (still technically Unicode,
but with notoriously terrible support for most Unicode characters) to
You’ll know that you’re successful if you try inserting a few umlauts or Mandarin
characters into the title or content field in the database and it succeeds with no
Now that the database is prepared to accept a wide variety of all that Wikipedia
can throw at it, you can run the following:
from urllib.request import urlopen
from bs4 import BeautifulSoup
import datetime
import random
import pymysql
conn = pymysql.connect(host='', unix_socket='/tmp/mysql.sock',
user='root', passwd=None, db='mysql', charset='utf8')
cur = conn.cursor()
cur.execute("USE scraping")
def store(title, content):
cur.execute("INSERT INTO pages (title, content) VALUES (\"%s\",
\"%s\")", (title, content))
def getLinks(articleUrl):
html = urlopen(""+articleUrl)
bsObj = BeautifulSoup(html)
title = bsObj.find("h1").find("span").get_text()
content = bsObj.find("div", {"id":"mw-content-text"}).find("p").get_text()
store(title, content)
return bsObj.find("div", {"id":"bodyContent"}).findAll("a",
links = getLinks("/wiki/Kevin_Bacon")
while len(links) > 0:
newArticle = links[random.randint(0, len(links)-1)].attrs["href"]
links = getLinks(newArticle)
There are a few things to note here: first, "charset='utf8'" is added to the
database connection string. This tells the connection that it should send all
information to the database as UTF-8 (and, of course, the database should already
be configured to handle this).
Second, note the addition of a store function. This takes in two string variables,
title and content, and adds them to an INSERT statement that is executed by the
cursor and then committed by the cursor’s connection. This is an excellent example
of the separation of the cursor and the connection; while the cursor has some stored
information about the database and its own context, it needs to operate through the
connection in order to send information back to the database and insert some
Last, you’ll see that a finally statement was added to the program’s main loop, at
the bottom of the code. This will ensure that, regardless of how the program is
interrupted or the exceptions that might be thrown during its execution (and, of
course, because the Web is messy, you should always assume exceptions will be
thrown), the cursor and the connection will both be closed immediately before the
program ends. It is a good idea to include a try...finally statement whenever
you are scraping the Web and have an open database connection.
Although PyMySQL is not a huge package, there are a fair number of useful
functions that this book simply can’t accommodate. Check out this documentation
Database Techniques and Good Practice
There are people who spend their entire careers studying, tuning, and inventing
databases. I am not one of them, and this is not that kind of book. However, as with
many subjects in computer science, there are a few tricks you can learn quickly to
at least make your databases sufficient, and sufficiently speedy, for most
First, with very few exceptions, always add id columns to your tables. All tables
in MySQL must have at least one primary key (the key column that MySQL sorts
on), so that MySQL knows how to order it, and it can often be difficult to choose
these keys intelligently. The debate over whether to use an artificially created id
column for this key or some unique attribute such as username has raged among
data scientists and software engineers for years, although I tend to lean on the side
of creating id columns. The reasons for doing it one way or the other are
complicated but for nonenterprise systems, you should always be using an id
column as an autoincremented primary key.
Second, use intelligent indexing. A dictionary (like the book, not the Python
object) is a list of words indexed alphabetically. This allows quick lookups
whenever you need a word, as long as you know how it’s spelled. You could also
imagine a dictionary that was organized alphabetically by the word’s definition.
This wouldn’t be nearly as useful unless you were playing some strange game of
Jeopardy! where a definition was presented and you needed to come up with the
word. But in the world of database lookups, these sorts of situations happen. For
example, you might have a field in your database that you will often be querying
>SELECT * FROM dictionary WHERE definition="A small furry animal that says
| id
| word | definition
| 200 | cat
| A small furry animal that says meow |
1 row in set (0.00 sec)
You might very well want to add an additional key to this table (in addition to the
key presumably already in place on the id) in order to make lookups on the
definition column faster. Unfortunately, adding additional indexing requires more
space for the new index, as well as some additional processing time when inserting
new rows. To make this a little easier, you can tell MySQL to index only the first
few characters in the column value. This command creates an index on the first 16
characters in the definition field:
CREATE INDEX definition ON dictionary (id, definition(16));
This index will make your lookups much faster when searching for words by their
full definition, and also not add too much in the way of extra space and upfront
processing time.
On the subject of query time versus database size (one of the fundamental balancing
acts in database engineering), one of the common mistakes made, especially with
web scraping where you often have large amounts of natural text data, is to store
lots of repeating data. For example, say you want to measure the frequency of
certain phrases that crop up across websites. These phrases might be found from a
given list, or automatically generated via some text-analysis algorithm. You might
be tempted to store the data as something like this:
| Field | Type
| Null | Key | Default | Extra
| id
| int(11)
| NO
| auto_increment |
| url
| varchar(200) | YES |
| phrase | varchar(200) | YES |
This adds a row to the database each time you find a phrase on a site and records
the URL it was found at. However, by splitting the data up into three separate
tables, you can shrink your dataset enormously:
>DESCRIBE phrases
| Field | Type
| Null | Key | Default | Extra
| id
| int(11)
| NO
| auto_increment |
| phrase | varchar(200) | YES |
| Field | Type
| Null | Key | Default | Extra
| id
| int(11)
| NO
| auto_increment |
| url
| varchar(200) | YES |
>DESCRIBE foundInstances
| Field
| Type
| Null | Key | Default | Extra
| id
| int(11) | NO
| auto_increment |
| urlId
| int(11) | YES |
| phraseId
| int(11) | YES |
| occurrences | int(11) | YES |
Although the table definitions are larger, you can see that the majority of the
columns are just integer id fields. These take up far less space. In addition, the full
text of each URL and phrase is stored exactly once.
Unless you install some third-party package or keep meticulous logs, it can be
impossible to tell when a piece of data was added, updated, or removed from your
database. Depending on the available space for your data, the frequency of changes,
and the importance of determining when those changes happened, you might want to
consider keeping several timestamps in place: “created,” “updated,” and “deleted.”
“Six Degrees” in MySQL
In Chapter 3, I introduced the “Six Degrees of Wikipedia” problem, in which the
goal is to find the connection between any two Wikipedia articles through a series
of links (i.e., find a way to get from one Wikipedia article to the next just by
clicking on links from one page to the next).
In order to solve this problem, it is necessary to not only build bots that can crawl
the site (which we have already done), but store the information in an
architecturally sound way to make data analysis easy later on.
Autoincremented id columns, timestamps, and multiple tables: they all come into
play here. In order to figure out how to best store this information, you need to think
abstractly. A link is simply something that connects Page A to Page B. It could just
as easily connect Page B to Page A, but this would be a separate link. We can
uniquely identify a link by saying, “There exists a link on page A, which connects to
page B. That is, INSERT INTO links (fromPageId, toPageId) VALUES (A, B);
(where “A” and “B” are the unique IDs for the two pages).
A two-table system designed to store pages and links, along with creation dates and
unique IDs, can be constructed as follows:
CREATE TABLE `wikipedia`.`pages` (
`url` VARCHAR(255) NOT NULL,
PRIMARY KEY (`id`));
CREATE TABLE `wikipedia`.`links` (
`fromPageId` INT NULL,
`toPageId` INT NULL,
PRIMARY KEY (`id`));
Notice that, unlike with previous crawlers that print the title of the page, I’m not
even storing the title of the page in the pages table. Why is that? Well, recording the
title of the page requires that you actually visit the page to retrieve it. If we want to
build an efficient web crawler to fill out these tables, we want to be able to store
the page, as well as links to it, even if we haven’t necessarily visited the page yet.
Of course, although this doesn’t hold true for all sites, the nice thing about
Wikipedia links and page titles is that one can be turned into the other through
simple manipulation. For
example, indicates that the title of the
page is “Monty Python.”
The following will store all pages on Wikipedia that have a “Bacon number” (the
number of links between it and the page for Kevin Bacon, inclusive) of 6 or
less: from
urllib.request import urlopen.
from bs4 import BeautifulSoup
import re
import pymysql
conn = pymysql.connect(host='', unix_socket='/tmp/mysql.sock', user=
'root', passwd=None, db='mysql', charset='utf8')
cur = conn.cursor()
cur.execute("USE wikipedia")
def insertPageIfNotExists(url):
cur.execute("SELECT * FROM pages WHERE url = %s", (url))
if cur.rowcount == 0:
cur.execute("INSERT INTO pages (url) VALUES (%s)", (url))
return cur.lastrowid
return cur.fetchone()[0]
def insertLink(fromPageId, toPageId):
cur.execute("SELECT * FROM links WHERE fromPageId = %s AND toPageId = %s",
(int(fromPageId), int(toPageId)))
if cur.rowcount == 0:
cur.execute("INSERT INTO links (fromPageId, toPageId) VALUES (%s, %s)",
(int(fromPageId), int(toPageId)))
pages = set()
def getLinks(pageUrl, recursionLevel):
global pages
if recursionLevel > 4:
pageId = insertPageIfNotExists(pageUrl)
html = urlopen(""+pageUrl)
bsObj = BeautifulSoup(html)
for link in bsObj.findAll("a",
if link.attrs['href'] not in pages:
#We have encountered a new page, add it and search it for links
newPage = link.attrs['href']
getLinks(newPage, recursionLevel+1)
getLinks("/wiki/Kevin_Bacon", 0)
Recursion is always a tricky thing to implement in code that is designed to run for a
long time. In this case, a recursionLevel variable is passed to the getLinks
function, which tracks how many times that function has been recursed on (each
time it is called, recursionLevel is incremented). When recursionLevel
reaches 5, the function automatically returns without searching further. This limit
ensures that a stack overflow never occurs.
Keep in mind that this program would likely take days to complete. Although I have
indeed run it, my database contains a mere fraction of the pages with a Kevin
Bacon number of 6 or less, for the sake of Wikipedia’s servers. However, this is
sufficient for our analysis of finding paths between linked Wikipedia articles.
For the continuation of this problem and the final solution, see Chapter 8 on solving
directed graph problems .
Just like web pages are sent over HTTP, email is sent over SMTP (Simple Mail
Transfer Protocol). And, just like you use a web server client to handle sending out
web pages over HTTP, servers use various email clients, such as Sendmail,
Postfix, or Mailman, to send and receive email.
Although sending email with Python is relatively easy to do, it does require that you
have access to a server running SMTP. Setting up an SMTP client on your server or
local machine is tricky, and outside the scope of this book, but there are many
excellent resources to help with this task, particularly if you are running Linux or
Mac OS X.
In the following code examples, I will assume that you are running an SMTP client
locally. (To modify this code for a remote SMTP client, just change localhost to
your remote server’s address.)
Sending an email with Python requires just nine lines of code:
import smtplib
from email.mime.text import MIMEText
msg = MIMEText("The body of the email is here")
msg['Subject'] = "An Email Alert"
msg['From'] = "[email protected]"
msg['To'] = "[email protected]"
s = smtplib.SMTP('localhost')
Python contains two important packages for sending emails: smtplib and email.
Python’s email module contains useful formatting functions for creating email
“packets” to send. The MIMEText object, used here, creates an empty email
formatted for transfer with the low-level MIME (Multipurpose Internet Mail
Extensions) protocol, across which the higher-level SMTP connections are made.
The MIMEText object, msg, contains to/from email addresses, as well as a body and
a header, which Python uses to create a properly formatted email.
The smtplib package contains information for handling the connection to the server.
Just like a connection to a MySQL server, this connection must be torn down every
time it is created, to avoid creating too many connections.
This basic email function can be extended and made more useful by enclosing it in
a function:
import smtplib
from email.mime.text import MIMEText
from bs4 import BeautifulSoup
from urllib.request import urlopen
import time
def sendMail(subject, body):
msg = MIMEText(body)
msg['Subject'] = subject
msg['From'] = "[email protected]"
msg['To'] = "[email protected]"
s = smtplib.SMTP('localhost')
bsObj = BeautifulSoup(urlopen(""))
while(bsObj.find("a", {"id":"answer"}).attrs['title'] == "NO"):
print("It is not Christmas yet.")
bsObj = BeautifulSoup(urlopen(""))
sendMail("It's Christmas!",
"According to, it is Christmas!")
This particular script checks the website (the main
feature of which is a giant “YES” or “NO,” depending on the day of the year) once
an hour. If it sees anything other than a “NO,” it will send you an email alerting you
that it’s Christmas.
Although this particular program might not seem much more useful than a calendar
hanging on your wall, it can be slightly tweaked to do a variety of extremely useful
things. It can email you alerts in response to site outages, test failures, or even the
appearance of an out-of-stock product you’re waiting for on Amazon — none of
which your wall calendar can do.
Joab Jackson, “YouTube Scales MySQL with Go Code,” PCWorld, December 15,
2012 (
Jeremy Cole and Davi Arnaut, “MySQL at Twitter,” The Twitter Engineering
Blog, April 9, 2012 (
“MySQL and Database Engineering: Mark Callaghan,” March 4, 2012
Chapter 6. Reading Documents
It is tempting to think of the Internet primarily as a collection of text-based websites
interspersed with newfangled web 2.0 multimedia content that can mostly be
ignored for the purposes of web scraping. However, this ignores what the Internet
most fundamentally is: a content-agnostic vehicle for transmitting files.
Although the Internet has been around in some form or another since the late 1960s,
HTML didn’t debut until 1992. Until then, the Internet consisted mostly of email and
file transmission; the concept of web pages as we know them today didn’t really
exist. In other words, the Internet is not a collection of HTML files. It is a
collection of information, with HTML files often being used as a frame to showcase
it. Without being able to read a variety of document types, including text, PDF,
images, video, email, and more, we are missing out on a huge part of the available
This chapter covers dealing with documents, whether you’re downloading them to a
local folder or reading them and extracting data. We’ll also take a look at dealing
with various types of text encoding, which can make it possible to even read
foreign-language HTML pages.
Document Encoding
A document’s encoding tells applications — whether they are your computer’s
operating system or your own Python code — how to read it. This encoding can
usually be deduced from its file extension, although this file extension is not
mandated by its encoding. I could, for example, save myImage.jpg as myImage.txt
with no problems — at least until my text editor tried to open it. Fortunately, this
situation is rare, and a document’s file extension is usually all you need to know in
order to read it correctly.
On a fundamental level, all documents are encoded in 0s and 1s. On top of that,
there are encoding algorithms that define things such as “how many bits per
character” or “how many bits represent the color for each pixel” (in the case of
image files). On top of that, you might have a layer of compression, or some spacereducing algorithm, as is the case with PNG files.
Although dealing with non-HTML files might seem intimidating at first, rest assured
that with the right library, Python will be properly equipped to deal with any format
of information you want to throw at it. The only difference between a text file, a
video file, and an image file is how their 0s and 1s are interpreted. In this chapter,
I’ll cover several commonly encountered types of files: text, PDFs, PNGs, and
It is somewhat unusual to have files stored as plain text online, but it is popular
among bare-bones or “old-school” sites to have large repositories of text files. For
example, the Internet Engineering Task Force (IETF) stores all of its published
documents as HTML, PDF, and text files (see as an example).
Most browsers will display these text files just fine and you should be able to
scrape them with no problem.
For most basic text documents, such as the practice file located at, you can use the
following method:
from urllib.request import urlopen
textPage = urlopen(
Normally, when we retrieve a page using urlopen, we turn it into a BeautifulSoup
object in order to parse the HTML. In this case, we can read the page directly.
Turning it into a BeautifulSoup object, while perfectly possible, would be
counterproductive — there’s no HTML to parse, so the library would be useless.
Once the text file is read in as a string, you merely have to analyze it like you would
any other string read into Python. The disadvantage here, of course, is that you don’t
have the ability to use HTML tags as context clues, pointing you in the direction of
the text you actually need, versus the text you don’t want. This can present a
challenge when you’re trying to extract certain information from text files.
Text Encoding and the Global Internet
Remember earlier when I said a file extension was all you needed to read a file
correctly? Well, strangely enough, that rule doesn’t apply to the most basic of all
documents: the .txt file.
Nine times out of 10, reading in text using the previously described methods will
work just fine. However, dealing with text on the Internet can be a tricky business.
Next we’ll cover the basics of English and foreign-language encoding, from ASCII
to Unicode to ISO, and how to deal with them.
A brief overview of encoding types
In the early 1990s, a nonprofit named The Unicode Consortium attempted to bring
about a universal text encoder by establishing encodings for every character that
needs to be used in any text document, in any language. The goal was to include
everything from the Latin alphabet this book is written in, to Cyrillic (кириллица),
Chinese pictograms (
), math and logic symbols (,≥), and even emoticons and
“miscellaneous” symbols, such as the biohazard sign ( ) and peace symbol ( ).
The resulting encoder, as you might already know, was dubbed UTF-8, which
stands for, confusingly, "Universal Character Set - Transformation Format 8 bit”. It
is a common misconception that UTF-8 stores all characters in 8 bits. However, the
“8 bit” refers to the smallest size that a character requires to be displayed, not the
maximum size. (If every character in UTF-8 were really stored in 8 bits, this would
allow only , or 256 possible characters. Clearly not enough room for everything
from a Chinese pictogram to a peace symbol.)
In reality, each character in UTF-8 starts with an indicator that says “only one byte
is used to encode this character” or “the next two bytes encodes a single character,”
with up to four bytes being possible. Because these four bytes also contain the
information about how many bytes are being used to encode the character, the full
32 bits (32 = 4 bytes x 8 bits per byte) are not used; rather, up to 21 bits of
information can be used, for a total of 2,097,152 possible characters, of
which, 1,114,112 are currently used.
Although Unicode was a godsend for many applications, some habits are hard to
break, and ASCII is still a popular choice for many users.
ASCII, a text-encoding standard used since the 1960s, uses 7 bits to encode each of
its characters, for a total of , or 128 characters. This is enough for the Latin
alphabet (both uppercase and lowercase), punctuation, and, essentially all of the
characters found on the average English speaker’s keyboard.
In 1960, of course, the difference between storing a text file with 7 bits per
character and 8 bits per character was significant because storage was so
expensive. Computer scientists of the day fought over whether an extra bit should
be added for the convenience of having a nice round number versus the practicality
of files requiring less storage space. In the end, 7 bits won. However, in modern
computing, each 7-bit sequence is padded with an extra “0” at the beginning,1
leaving us with the worst of both worlds — 14% larger files, and the lack of
flexibility of only 128 characters.
When UTF-8 was designed, the creators decided to use this “padding bit” in ASCII
documents to their advantage by declaring all bytes starting with a “0” to indicate
that only one byte is used in the character, and making the two encoding schemes for
ASCII and UTF-8 identical. Therefore, the following characters are valid in both
UTF-8 and ASCII:
01000001 - A
01000010 - B
01000011 - C
And the following characters are valid only in UTF-8, and will be rendered as
“non-printable” if the document is interpreted as an ASCII document:
11000011 10000000 - À
11000011 10011111 - ß
11000011 10100111 - ç
In addition to UTF-8, there are other UTF standards such as UTF-16, UTF-24, and
UTF-32, although documents encoded in these formats are rarely encountered
except in unusual circumstances, which are outside the scope of this book.
The problem with all Unicode standards, of course, is that any document written in
a single foreign language is much larger than it has to be. Although your language
might only require 100 or so characters to be used, you will need at least 16 bits
for each character rather than just 8 bits, as is the case for the English-specific
ASCII. This makes foreign-language text documents about twice the size of Englishlanguage text documents, at least for foreign languages that don’t use the Latin
character set.
ISO solves this problem by creating specific encodings for each language. Like
Unicode, it has the same encodings that ASCII does, but uses the “padding” 0 bit at
the beginning of every character to allow it to create special characters for all of
the languages that require them. This works best for European languages that also
rely heavily on the Latin alphabet (which remain in positions 0-127 in the
encoding), but require some additional special characters. This allows ISO-8859-1
(designed for the Latin alphabet) to have symbols such as fractions (e.g., ½) or the
copyright sign (©).
Other ISO character sets, such as ISO-8859-9 (Turkish), ISO-8859-2 (German,
among other languages), and ISO-8859-15 (French, among other languages) can
also be found with some regularity.
Although the popularity of ISO-encoded documents has been declining in recent
years, about 9% of websites on the Internet are still encoded with some flavor of
ISO,2 making it essential to know about and check for encodings before scraping a
Encodings in action
In the previous section, we’ve used the default settings for urlopen to read in .txt
documents you might encounter on the Internet. This works great for most English
text. However, the second you encounter Russian, Arabic, or even a word like
résumé, you might run into problems.
Take the following code, for example:
from urllib.request import urlopen
textPage = urlopen(
This reads in the first chapter of the original War and Peace (written in Russian
and French) and prints it to the screen. This screen text reads, in part:
b"\xd0\xa7\xd0\x90\xd0\xa1\xd0\xa2\xd0\xac \xd0\x9f\xd0\x95\xd0\xa0\xd0\x92\xd0\
x90\xd0\xaf\n\nI\n\n\xe2\x80\x94 Eh bien, mon prince.
In addition, visiting this page in most browsers results in gibberish (see Figure 61).
Figure 6-1. French and Cyrillic text encoded in ISO-8859-1, the default text document
encoding in many browsers
Even for native Russian speakers, that might be a bit difficult to make sense of. The
problem is that Python is attempting to read the document as an ASCII document,
whereas the browser is attempting to read it as an ISO-8859-1 encoded document.
Neither one, of course, realizes it’s a UTF-8 document.
We can explicitly define the string to be UTF-8, which correctly formats the output
into Cyrillic characters:
from urllib.request import urlopen
textPage = urlopen(
print(str(, 'utf-8'))
Using this concept in BeautifulSoup and Python 3.x looks like this:
html = urlopen("")
bsObj = BeautifulSoup(html)
content = bsObj.find("div", {"id":"mw-content-text"}).get_text()
content = bytes(content, "UTF-8")
content = content.decode("UTF-8")
You might be tempted to use UTF-8 encoding for every web scraper you write.
After all, UTF-8 will also handle ASCII characters smoothly. However, it’s
important to remember the 9% of websites out there that use some version of ISO
encoding. Unfortunately, in the case of text documents, it’s impossible to concretely
determine what encoding a document has. There are some libraries that can
examine the document and make a best guess (using a little logic to realize that “Ñ
€Ð°Ñ​Ñ​казє is probably not a word), but many times it’s wrong.
Fortunately, in the case of HTML pages, the encoding is usually contained in a tag
found in the <head> section of the site. Most sites, particularly English-language
sites, have the tag:
<meta charset="utf-8" />
Whereas the European Computer Manufacturers Association’s website has this
<META HTTP-EQUIV="Content-Type" CONTENT="text/html; charset=iso-8859-1">
If you plan on doing a lot of web scraping, particularly of international sites, it
might be wise to look for this meta tag and use the encoding it recommends when
reading the contents of the page.
When web scraping you are likely to encounter either a CSV file or a coworker
who likes data formatted in this way. Fortunately, Python has a fantastic library for
both reading and writing CSV files. Although this library is capable of handling
many variations of CSV, I will focus primarily on the standard format. If you have a
special case you need to handle, consult the documentation!
Reading CSV Files
Python’s csv library is geared primarily toward working with local files, on the
assumption that the CSV data you need is stored on your machine. Unfortunately,
this isn’t always the case, especially when you’re web scraping. There are several
ways to work around this:
Download the file locally by hand and point Python at the local file location
Write a Python script to download the file, read it, and (optionally) delete it
after retrieval
Retrieve the file as a string from the Web, and wrap the string in a StringIO
object so that it behaves like a file
Although the first two options are workable, taking up hard drive space with files
when you could very easily keep them in memory is bad practice. It’s much better
to read the file in as a string and wrap it in an object that allows Python to treat it
like a file, without ever actually saving the file. The following script retrieves a
CSV file from the Internet (in this case, a list of Monty Python albums
at and prints it, row by row, to the terminal:
from urllib.request import urlopen
from io import StringIO
import csv
data = urlopen("")
.read().decode('ascii', 'ignore')
dataFile = StringIO(data)
csvReader = csv.reader(dataFile)
for row in csvReader:
The output is too long to print in full but should look something like:
['Name', 'Year']
["Monty Python's Flying Circus", '1970']
['Another Monty Python Record', '1971']
["Monty Python's Previous Record", '1972']
As you can see from the code sample, the reader object returned by csv.reader is
iterable, and composed of Python list objects. Because of this, each row in the
csvReader object is accessible in the following way:
for row in csvReader:
print("The album \""+row[0]+"\" was released in "+str(row[1]))
Which gives the output:
"Name" was released in Year
"Monty Python's Flying Circus" was released in 1970
"Another Monty Python Record" was released in 1971
"Monty Python's Previous Record" was released in 1972
Notice the first line: The album ‘Name’ was released in Year. Although this might
be an easy-to-ignore result when writing example code, you don’t want this getting
into your data in the real world. A lesser programmer might simply skip the first
row in the csvReader object, or write in some special case to handle it.
Fortunately, there’s an alternative to the csv.reader function that takes care of all
of this for you automatically. Enter DictReader:
from urllib.request import urlopen
from io import StringIO
import csv
data = urlopen("")
.read().decode('ascii', 'ignore')
dataFile = StringIO(data)
dictReader = csv.DictReader(dataFile)
for row in dictReader:
returns the values of each row in the CSV file as dictionary
objects rather than list objects, with fieldnames stored in the variable
dictReader.fieldnames and as keys in each Dictionary object:
"Monty Python's Flying Circus", 'Year': '1970'}
'Another Monty Python Record', 'Year': '1971'}
"Monty Python's Previous Record", 'Year': '1972'}
The downside, of course, is that it takes slightly longer to create, process, and print
these DictReaders as opposed to csvReaders, but the convenience and usability
is often worth the additional overhead.
As a Linux user, I know the pain of being sent a .docx file that my non-Microsoft
software mangles and struggling trying to find the codecs to interpret some new
Apple media format. In some ways, Adobe was revolutionary in creating its
Portable Document Format in 1993. PDFs allowed users on different platforms to
view image and text documents in exactly the same way, regardless of the platform
they were viewing it on.
Although storing PDFs on the Web is somewhat passé (why store content in a static,
slow-loading format when you could write it up as HTML?), PDFs remain
ubiquitous, particularly when dealing with official forms and filings.
In 2009, a Briton named Nick Innes made the news when he requested public
student test result information from the Buckinghamshire City Council, which was
available under the United Kindom’s version of the Freedom of Information Act.
After some repeated requests and denials, he finally received the information he
was looking for — in the form of 184 PDF documents.
Although Innes persisted and eventually received a more properly formatted
database, had he been an expert web scraper he likely could have saved himself a
lot of time in the courts and used the PDF documents directly, with one of Python’s
many PDF parsing modules.
Unfortunately, many of the PDF parsing libraries built for Python 2.x were not
upgraded with the launch of Python 3.x. However, because the PDF is a relatively
simple and open source document format, there are many decent Python libraries,
even in Python 3.x, that can read them.
PDFMiner3K is one such relatively easy-to-use library. It is very flexible, allowing
for command-line usage or integration into existing code. It can also handle a
variety of language encodings — again, something that often comes in handy on the
You can download this Python module and install it by unzipping the folder and
$python install
The documentation is located at /pdfminer3k-1.3.0/docs/index.html within the
extracted folder, although the current documentation tends to be geared more
toward the command-line interface than integration with Python code.
Here is a basic implementation that allows you to read arbitrary PDFs to a string,
given a local file object:
pdfminer.pdfinterp import PDFResourceManager, process_pdf
pdfminer.converter import TextConverter
pdfminer.layout import LAParams
io import StringIO
io import open
def readPDF(pdfFile):
rsrcmgr = PDFResourceManager()
retstr = StringIO()
laparams = LAParams()
device = TextConverter(rsrcmgr, retstr, laparams=laparams)
process_pdf(rsrcmgr, device, pdfFile)
content = retstr.getvalue()
return content
pdfFile = urlopen("");
outputString = readPDF(pdfFile)
The nice thing about this function is that if you’re working with files locally, you
can simply substitute a regular Python file object for the one returned by urlopen,
and use the line:
pdfFile = open("../pages/warandpeace/chapter1.pdf", 'rb')
The output might not be perfect, especially for PDFs with images, oddly formatted
text, or text arranged in tables or charts. However, for most text-only PDFs the
output should be no different than if the PDF were a text file.
Microsoft Word and .docx
At the risk of offending my friends at Microsoft: I do not like Microsoft Word. Not
because it’s necessarily a bad piece of software, but because of the way its users
misuse it. It has a particular talent for turning what should be simple text documents
or PDFs into large, slow, difficult-to-open beasts that often lose all formatting from
machine to machine, and are, for whatever reason, editable when the content is
often meant to be static. Word files were never meant for frequent transmission.
Nevertheless, they are ubiquitous on certain sites, containing important documents,
information, and even charts and multimedia; in short, everything that can and
should be created with HTML.
Before about 2008, Microsoft Office products used the proprietary .docfile format.
This binary-file format was difficult to read and poorly supported by other word
processors. In an effort to get with the times and adopt a standard that was used by
many other pieces of software, Microsoft decided to use the Open Office XMLbased standard, which made the files compatible with open source and other
Unfortunately, Python’s support for this file format, used by Google Docs, Open
Office, and Microsoft Office, still isn’t great. There is the python-docx library, but
this only gives users the ability to create documents and read only basic file data
such as the size and title of the file, not the actual contents. To read the contents of a
Microsoft Office file, we’ll need to roll our own solution.
The first step is to read the XML from the file:
from zipfile import ZipFile
from urllib.request import urlopen
from io import BytesIO
wordFile = urlopen("").read()
wordFile = BytesIO(wordFile)
document = ZipFile(wordFile)
xml_content ='word/document.xml')
This reads a remote Word document as a binary file object (BytesIO is analogous
to StringIO, used earlier in this chapter), unzips it using Python’s core zipfile
library (all .docx files are zipped to save space), and then reads the unzipped file,
which is XML.
The Word document at is shown in Figure 6-2.
Figure 6-2. This is a Word document that’s full of content you might want very much but
it’s difficult to access because I’m putting it on my website as a .docx file instead of
publishing it as HTML
The output of the Python script reading my simple Word document is the following:
<!--?xml version="1.0" encoding="UTF-8" standalone="yes"?-->
<w:document mc:ignorable="w14 w15 wp14" xmlns:m="http://schemas.openx" xmlns:mc="" xmlns:o="urn:schemas-micros
oft-com:office:office" xmlns:r="
iceDocument/2006/relationships" xmlns:v="urn:schemas-microsoft-com:vm
l" xmlns:w="
ain" xmlns:w10="urn:schemas-microsoft-com:office:word" xmlns:w14="htt
p://" xmlns:w15="http://" xmlns:wne="http://sche" xmlns:wp="http://schemas.o" xmlns:wp14="h
ttp://" x
gCanvas" xmlns:wpg="
dprocessingGroup" xmlns:wpi="
/2010/wordprocessingInk" xmlns:wps="
ce/word/2010/wordprocessingShape"><w:body><w:p w:rsidp="00764658" w:r
sidr="00764658" w:rsidrdefault="00764658"><w:ppr><w:pstyle w:val="Tit
le"></w:pstyle></w:ppr><w:r><w:t>A Word Document on a Website</w:t></
w:r><w:bookmarkstart w:id="0" w:name="_GoBack"></w:bookmarkstart><w:b
ookmarkend w:id="0"></w:bookmarkend></w:p><w:p w:rsidp="00764658" w:r
sidr="00764658" w:rsidrdefault="00764658"></w:p><w:p w:rsidp="0076465
8" w:rsidr="00764658" w:rsidrdefault="00764658" w:rsidrpr="00764658">
<w: r> <w:t>This is a Word document, full of content that you want ve
ry much. Unfortunately, it’s difficult to access because I’m putting
it on my website as a .</w:t></w:r><w:prooferr w:type="spellStart"></
w:prooferr><w:r><w:t>docx</w:t></w:r><w:prooferr w:type="spellEnd"></
w:prooferr> <w:r> <w:t xml:space="preserve"> file, rather than just p
ublishing it as HTML</w:t> </w:r> </w:p> <w:sectpr w:rsidr="00764658"
w:rsidrpr="00764658"> <w:pgszw:h="15840" w:w="12240"></w:pgsz><w:pgm
ar w:bottom="1440" w:footer="720" w:gutter="0" w:header="720" w:left=
"1440" w:right="1440" w:top="1440"></w:pgmar> <w:cols w:space="720"><
/w:cols&g; <w:docgrid w:linepitch="360"></w:docgrid> </w:sectpr> </w:
body> </w:document>
There’s clearly a lot of information here but it’s buried. Fortunately, all of the text
in the document, including the title at the top, is contained in <w:t> tags, which
makes it easy to grab:
zipfile import ZipFile
urllib.request import urlopen
io import BytesIO
bs4 import BeautifulSoup
wordFile = urlopen("").read()
wordFile = BytesIO(wordFile)
document = ZipFile(wordFile)
xml_content ='word/document.xml')
wordObj = BeautifulSoup(xml_content.decode('utf-8'))
textStrings = wordObj.findAll("w:t")
for textElem in textStrings:
The output isn’t perfect but it’s getting there, and printing each <w:t> tag on a new
line makes it easy to see how Word is splitting up the text:
A Word Document on a Website
This is a Word document, full of content that you want very much. Unfortunately,
it’s difficult to access because I’m putting it on my website as a . docx
file, rather than just publishing it as HTML
Notice that the word “docx” is on its own line. In the original XML, it is
surrounded with the tag <w:proofErr w:type="spellStart"/>. This is Word’s
way of highlighting “docx” with the red squiggly underline, indicating that it
believes there’s a spelling error in the name of its own file format.
The title of the document is preceded by the style descriptor tag <w:pstyle
w:val="Title">. Although this doesn’t make it extremely easy for us to identify
titles (or other styled text) as such, using BeautifulSoup’s navigation features can be
textStrings = wordObj.findAll("w:t")
for textElem in textStrings:
closeTag = ""
style = textElem.parent.previousSibling.find("w:pstyle")
if style is not None and style["w:val"] == "Title":
closeTag = "</h1>"
except AttributeError:
#No tags to print
This function can be easily expanded to print tags around a variety of different text
styles or label them in some other way.
This “padding” bit will come back to haunt us with the ISO standards, a little
According to, which
uses web crawlers to gather these sorts of statistics.
Ecma was one of the original contributors to the ISO standard, so it’s no surprise
its website is encoded with a flavor of ISO.
Part II. Advanced Scraping
You’ve laid some web-scraping groundwork; now comes the fun part. Up until this
point our web scrapers have been relatively dumb. They’re unable to retrieve
information unless it’s immediately presented to them in a nice format by the server.
They take all information at face value and simply store it without any analysis.
They get tripped up by forms, website interaction, and even JavaScript. In short,
they’re no good for retrieving information unless that information really wants to be
This part of the book will help you analyze raw data to get the story beneath the
data — the story that websites often hide beneath layers of JavaScript, login forms,
and antiscraping measures.
You’ll learn how to use web scrapers to test your sites, automate processes, and
access the Internet on a large scale. By the end of this section, you should have the
tools to gather and manipulate nearly any type of data, in any form, across any part
of the Internet.
Chapter 7. Cleaning Your Dirty Data
So far in this book we’ve ignored the problem of badly formatted data by using
generally well-formatted data sources, dropping data entirely if it deviated from
what we were expecting. But often, in web scraping, you can’t be too picky about
where you get your data from.
Due to errant punctuation, inconsistent capitalization, line breaks, and misspellings,
dirty data can be a big problem on the Web. In this chapter, I’ll cover a few tools
and techniques to help you prevent the problem at the source by changing the way
you write code, and clean the data once it’s in the database.
Cleaning in Code
Just as you write code to handle overt exceptions, you should practice defensive
coding to handle the unexpected.
In linguistics, an n-gram is a sequence of n words used in text or speech. When
doing natural-language analysis, it can often be handy to break up a piece of text by
looking for commonly used n-grams, or recurring sets of words that are often used
In this section, we will focus on obtaining properly formatted n-grams rather than
using them to do any analysis. Later, in Chapter 8, you can see 2-grams and 3-grams
in action to do text summarization and analysis.
The following will return a list of 2-grams found in the Wikipedia article on the
Python programming language:
from urllib.request import urlopen
from bs4 import BeautifulSoup
def ngrams(input, n):
input = input.split(' ')
output = []
for i in range(len(input)-n+1):
return output
html = urlopen("")
bsObj = BeautifulSoup(html)
content = bsObj.find("div", {"id":"mw-content-text"}).get_text()
ngrams = ngrams(content, 2)
print("2-grams count is: "+str(len(ngrams)))
The ngrams function takes in an input string, splits it into a sequence of words
(assuming all words are separated by spaces), and adds the n-gram (in this case, a
2-gram) that each word starts into an array.
This returns some genuinely interesting and useful 2-grams from the text:
['of', 'free'], ['free', 'and'], ['and', 'open-source'], ['open-source', 'softwa
but also a lot of junk:
['software\nOutline\nSPDX\n\n\n\n\n\n\n\n\nOperating', 'system\nfamilies\n\n\n\n
olaris\nPlan'], ['system\nfamilies\n\n\n\nAROS\nBSD\nDarwin\neCos\nFreeDOS\nGNU\
nHaiku\nInferno\nLinux\nMach\nMINIX\nOpenSolaris\nPlan', '9\nReactOS\nTUD:OS\n\n
\n\n\n\n\n\n\nDevelopment\n\n\n\nBasic'], ['9\nReactOS\nTUD:OS\n\n\n\n\n\n\n\n\n
Development\n\n\n\nBasic', 'For']
In addition, because there is a 2-gram created for each and every word encountered
(except for the last one), there are 7,411 2-grams in the article at the time of this
writing. Not a very manageable dataset!
Using some regular expressions to remove escape characters (such as \n) and
filtering to remove any Unicode characters, we can clean up the output somewhat:
def ngrams(input, n):
content = re.sub('\n+', " ", content)
content = re.sub(' +', " ", content)
content = bytes(content, "UTF-8")
content = content.decode("ascii", "ignore")
input = input.split(' ')
output = []
for i in range(len(input)-n+1):
return output
This first replaces all instances of the newline character (or multiple newline
characters) with a space, then replaces all instances of multiple spaces in a row
with a single space, ensuring that all words have one space between them. Then,
escape characters are eliminated by encoding the content with UTF-8.
These steps greatly improve the output of the function, but there are still some
['Pythoneers.[43][44]', 'Syntax'], ['7', '/'], ['/', '3'], ['3', '=='], ['==', '
At this point, the decisions that need to be made in order to process this data
become more interesting. There are a few more rules we can add to get closer to
ideal data:
Single character “words” should be discarded, unless that character is “i” or
Wikipedia citation marks (numbers enclosed in brackets) should be discarded
Punctuation marks should be discarded (note: this rule is somewhat of a
simplification and will be explored in greater detail in Chapter 9, but is fine for
the purpose of this example)
Now that the list of “cleaning tasks” is getting longer, it’s best to move these out
and put them in a separate function, cleanInput:
from urllib.request import urlopen
from bs4 import BeautifulSoup
import re
import string
def cleanInput(input):
input = re.sub('\n+', " ", input)
input = re.sub('\[[0-9]*\]', "", input)
input = re.sub(' +', " ", input)
input = bytes(input, "UTF-8")
input = input.decode("ascii", "ignore")
cleanInput = []
input = input.split(' ')
for item in input:
item = item.strip(string.punctuation)
if len(item) > 1 or (item.lower() == 'a' or item.lower() == 'i'):
return cleanInput
def ngrams(input, n):
input = cleanInput(input)
output = []
for i in range(len(input)-n+1):
return output
Note the use of import string and string.punctuation to get a list of all
punctuation characters in Python. You can view the output of string.punctuation
from a Python terminal:
>>> import string
>>> print(string.punctuation)
!"#$%&'()*+,-./:;<=>[email protected][\]^_`{|}~
By using item.strip(string.punctuation) inside a loop iterating through all
words in the content, any punctuation characters on either side of the word will be
stripped, although hyphenated words (where the punctuation character is bounded
by letters on either side) will remain untouched.
The output of this effort results in much cleaner 2-grams:
['Linux', 'Foundation'], ['Foundation', 'Mozilla'], ['Mozilla', 'Foundation'], [
'Foundation', 'Open'], ['Open', 'Knowledge'], ['Knowledge', 'Foundation'], ['Fou
ndation', 'Open'], ['Open', 'Source']
Data Normalization
Everyone has encountered a poorly designed web form: “Enter your phone number.
Your phone number must be in the form ‘xxx-xxx-xxxx’.”
As a good programmer, you will likely think to yourself, “Why don’t they just strip
out the non-numeric characters I put in there and do it themselves?” Data
normalization is the process of ensuring that strings that are linguistically or
logically equivalent to each other, such as the phone numbers “(555) 123-4567”
and “555.123.4567,” are displayed, or at least compared, as equivalent.
Using the n-gram code from the previous section, we can add on some data
normalization features.
One obvious problem with this code is that it contains many duplicate 2-grams.
Every 2-gram it encounters gets added to the list, with no record of its frequency.
Not only is it interesting to record the frequency of these 2-grams, rather than just
their existence, but it can be useful in charting the effects of changes to the cleaning
and data normalization algorithms. If data is normalized successfully, the total
number of unique n-grams will be reduced, while the total count of n-grams found
(i.e., the number of unique or non-unique items identified as a n-gram) will not be
reduced. In other words, there will be fewer “buckets” for the same number of ngrams.
Unfortunately for the purposes of this exercise, Python dictionaries are unsorted.
“Sorting a dictionary” doesn’t make any sense, unless you are copying the values in
the dictionary to some other content type and sorting that. An easy solution to this
problem is the OrderedDict, from Python’s collections library:
from collections import OrderedDict
ngrams = ngrams(content, 2)
ngrams = OrderedDict(sorted(ngrams.items(), key=lambda t: t[1], reverse=True))
Here I’m taking advantage of Python’s sorted function in order to put the items into
a new OrderedDict object, sorted by the value. The results:
("['Software', 'Foundation']", 40), ("['Python', 'Software']", 38), ("['of', 'th
e']", 35), ("['Foundation', 'Retrieved']", 34), ("['of', 'Python']", 28), ("['in
', 'the']", 21), ("['van', 'Rossum']", 18)
As of this writing, there are 7,696 total 2-grams and 6,005 unique 2-grams, with the
most popular 2-gram being “Software Foundation,” followed by “Python
Software.” However, analysis of the results shows that “Python Software” actually
appears in the form of “Python software” an additional two times. Similarly, both
“van Rossum” and “Van Rossum” make an appearance in the list separately.
Adding the line:
input = input.upper()
to the cleanInput function keeps the total number of 2-grams found steady at
7,696, while decreasing the number of unique 2-grams to 5,882.
Beyond this, it’s usually good to stop and consider how much computing power you
want to expend normalizing data. There are a number of situations in which
different spellings of words are equivalent, but in order to resolve this equivalency
you need to run a check on every single word to see if it matches any of your
preprogrammed equivalencies.
For example, “Python 1st” and “Python first” both appear in the list of 2-grams.
However, to make a blanket rule that says “All ‘first,’ ’second,’ ‘third,’ etc. will be
resolved to 1st, 2nd, 3rd, etc. (or vice versa)” would result in an additional 10 or
so checks per word.
Similarly, the inconsistent use of hyphens (“co-ordinated” versus “coordinated”),
misspellings, and other natural language incongruities will affect the groupings of
n-grams, and might muddy the results of the output if the incongruities are common
One solution, in the case of hyphenated words, might be to remove hyphens entirely
and treat the word as a single string, which would require only a single operation.
However, this would also mean that hyphenated phrases (an all-too-common
occurrence) will be treated as a single word. Going the other route and treating
hyphens as spaces might be a better option. Just be prepared for the occasional “co
ordinated” and “ordinated attack” to slip in!
Cleaning After the Fact
There is only so much you can, or want to do, in code. In addition, you might be
dealing with a dataset that you didn’t create, or a dataset that would be a challenge
to even know how to clean without seeing it first.
A knee-jerk reaction that many programmers have in this sort of situation is “write
a script,” which can be an excellent solution. However, there are also third-party
tools, such as OpenRefine, that are capable of not only cleaning data quickly and
easily, but allow your data to be easily seen and used by nonprogrammers.
OpenRefine is an open source project started by a company called Metaweb in
2009. Google acquired Metaweb in 2010, changing the name of the project from
Freebase Gridworks to Google Refine. In 2012, Google dropped support for
Refine and changed the name again, to OpenRefine, where anyone is welcome to
contribute to the development of the project.
OpenRefine is unusual in that although its interface is run in a browser, it is
technically a desktop application that must be downloaded and installed. You can
download the application for Linux, Windows, and Mac OS X from its website.
If you’re a Mac user and run into any trouble opening the file, go to System
Preferences → Security & Privacy → General → and check “Anywhere” under
“Allow apps downloaded from.” Unfortunately, during the transition from a
Google project to an open source project, OpenRefine appears to have lost its
legitimacy in the eyes of Apple.
In order to use OpenRefine, you’ll need to save your data as a CSV (file refer back
to “Storing Data to CSV” in Chapter 5 if you need a refresher on how to do this).
Alternatively, if you have your data stored in a database, you might be able to
export it to a CSV file.
Using OpenRefine
In the following examples, we’ll use data scraped from Wikipedia’s “Comparison
of Text Editors” table; see Figure 7-1. Although this table is relatively well
formatted, it contains many edits by people over a long time, so it has a few minor
formatting inconsistencies. In addition, because its data is meant to be read by
humans rather than machines, some of the formatting choices (e.g., using “Free”
rather than “$0.00”) is inappropriate for programming inputs.
Figure 7-1. Data from Wikipedia’s “comparison of text editors” as shown in the main
OpenRefine screen
The first thing to note about OpenRefine is that each column label has an arrow next
to it. This arrow provides a menu of tools that can be used with that column for
filtering, sorting, transforming, or removing data.
Data filtering can be performed using two methods: filters and facets. Filters are
good for using regular expressions to filter the data; for example, “Only show me
data that contains four or more comma-seperated programming languages in the
Programming language column,” seen in Figure 7-2.
Figure 7-2. The regular expression “.+,.+,.+” selects for values that have at least three
comma-separated items
Filters can be combined, edited, and added easily by manipulating the blocks in the
righthand column. They also can be combined with facets.
Facets are great for including or excluding data based on the entire contents of the
column. (e.g., “Show all rows that use the GPL or MIT license, and were first
released after 2005,” seen in Figure 7-3). They have built-in filtering tools. For
instance, filtering on a numeric value provides you with slide bars to select the
value range that you want to include.
Figure 7-3. This displays all text editors using the GPL or MIT license that had their first
public release after 2005
However you filter your data, it can be exported at any point to one of several types
of formats that OpenRefine supports. This includes CSV, HTML (an HTML table),
Excel, and several other formats.
Data filtering can be done successfully only if the data is relatively clean to start
with. For instance, in the facet example in the previous section, a text editor that
had a release date of “01-01-2006” would not have been selected in the “First
public release” facet, which was looking for a value of “2006” and ignoring values
that didn’t look like that.
Data transformation is performed in OpenRefine using the OpenRefine Expression
Language, called GREL (the “G” is left over from OpenRefine’s previous name,
Google Refine). This language is used to create short lambda functions that
transform the values in the cells based on simple rules. For example:
if(value.length() != 4, "invalid", value)
When this function is applied to the “First stable release” column, it preserves the
values of the cells where the date is in a “YYYY” format, and marks all other
columns as “invalid.”
Arbitrary GREL statements can be applied by clicking the down arrow next to any
column’s label and going to edit cells → transform.
Figure 7-4. Inserting a GREL statement into a project (a preview display is shown below
the statement)
However, marking all less than ideal values as invalid, while making them easy to
spot, doesn’t do us much good. We’d rather try to salvage information from the
badly formatted values if possible. This can be done using GREL’s match function:
This attempts to match the string value against the regular expression given. If the
regular expression matches the string, an array is returned. Any substrings that
match the “capture group” in the regular expression (demarcated by the parentheses
in the expression, in this example, “[0-9]{4}”) are returned as array values.
This code, in effect, finds all instances of four decimals in a row and returns the
first one. This is usually sufficient to extract years from text or badly formatted
dates. It also has the benefit of returning “null” for nonexistent dates. (GREL does
not throw a null pointer exception when performing operations on a null variable)
Many other data transformations are possible with cell editing and GREL. A
complete guide to the language can be found on OpenRefine’s GitHub page.
Chapter 8. Reading and Writing Natural Languages
So far the data we have worked with generally has been in the form of numbers or
countable values. In most cases, we’ve simply stored the data without conducting
any analysis after the fact. In this chapter, we’ll attempt to tackle the tricky subject
of the English language.1
How does Google know what you’re looking for when you type “cute kitten” into
its Image Search? Because of the text that surrounds the cute kitten images. How
does YouTube know to bring up a certain Monty Python sketch when you type “dead
parrot” into its search bar? Because of the title and description text that
accompanies each uploaded video.
In fact, even typing in terms such as “deceased bird monty python” immediately
brings up the same “Dead Parrot” sketch, even though the page itself contains no
mention of the words “deceased” or “bird.” Google knows that a “hot dog” is a
food and that a “boiling puppy” is an entirely different thing. How? It’s all
Although you might not think that text analysis has anything to do with your project,
understanding the concepts behind it can be extremely useful for all sorts of
machine learning, as well as the more general ability to model real-world problems
in probabilistic and algorithmic terms.
For instance, the Shazam music service can identify audio as containing a certain
song recording, even if that audio contains ambient noise or distortion. Google is
working on automatically captioning images based on nothing but the image itself.2
By comparing known images of, say, hot dogs to other images of hot dogs, the
search engine can gradually learn what a hot dog looks like and observe these
patterns in additional images it is shown.
Summarizing Data
In Chapter 7, we looked at breaking up text content into n-grams, or sets of phrases
that are n-words in length. At a very basic level, this can be used to determine
which sets of words and phrases tend to be most commonly used in a section of
text. In addition, it can be used to create natural-sounding data summaries by going
back to the original text and extracting sentences around some of these most popular
One piece of sample text we’ll be using to do this is the inauguration speech of the
ninth president of the United States, William Henry Harrison. Harrison’s
presidency sets two records in the history of the office: one for the longest
inauguration speech, and another for the shortest time in office, 32 days.
We’ll use the full text of this speech as the source for many of the code samples in
this chapter.
Slightly modifying the n-gram used to find code in Chapter 7, we can produce code
that looks for sets of 2-grams and sorts them using Python’s sorting function in the
“operator” module:
from urllib.request import urlopen
from bs4 import BeautifulSoup
import re
import string
import operator
def cleanInput(input):
input = re.sub('\n+', " ", input).lower()
input = re.sub('\[[0-9]*\]', "", input)
input = re.sub(' +', " ", input)
input = bytes(input, "UTF-8")
input = input.decode("ascii", "ignore")
cleanInput = []
input = input.split(' ')
for item in input:
item = item.strip(string.punctuation)
if len(item) > 1 or (item.lower() == 'a' or item.lower() == 'i'):
return cleanInput
def ngrams(input, n):
input = cleanInput(input)
output = {}
for i in range(len(input)-n+1):
ngramTemp = " ".join(input[i:i+n])
if ngramTemp not in output:
output[ngramTemp] = 0
output[ngramTemp] += 1
return output
content = str(
ngrams = ngrams(content, 2)
sortedNGrams = sorted(ngrams.items(), key = operator.itemgetter(1),
The output produces, in part:
[('of the', 213),
he constitution',
), ('the people',
3), ('of a', 22),
the', 65), ('to the', 61), ('by the', 41), ('t
('of our', 29), ('to be', 26), ('from the', 24
('and the', 23), ('it is', 23), ('that the', 2
their', 19)
Of these 2-grams, “the constitution” seems like a reasonably popular subject in the
speech, but “of the,” “in the,” and “to the” don’t seem especially noteworthy. How
can you automatically get rid of unwanted words in an accurate way?
Fortunately, there are people out there who carefully study the differences between
“interesting” words and “uninteresting” words, and their work can help us do just
that. Mark Davies, a linguistics professor at Brigham Young University, maintains
the Corpus of Contemporary American English, a collection of over 450 million
words from the last decade or so of popular American publications.
The list of 5,000 most frequently found words is available for free, and fortunately,
this is far more than enough to act as a basic filter to weed out the most common 2grams. Just the first 100 words vastly improves the results, with the addition of an
isCommon function:
def isCommon(ngram):
commonWords = ["the", "be", "and", "of", "a", "in", "to", "have", "it",
"i", "that", "for", "you", "he", "with", "on", "do", "say", "this",
"they", "is", "an", "at", "but","we", "his", "from", "that", "not",
"by", "she", "or", "as", "what", "go", "their","can", "who", "get",
"if", "would", "her", "all", "my", "make", "about", "know", "will",
"as", "up", "one", "time", "has", "been", "there", "year", "so",
"think", "when", "which", "them", "some", "me", "people", "take",
"out", "into", "just", "see", "him", "your", "come", "could", "now",
"than", "like", "other", "how", "then", "its", "our", "two", "more",
"these", "want", "way", "look", "first", "also", "new", "because",
"day", "more", "use", "no", "man", "find", "here", "thing", "give",
"many", "well"]
for word in ngram:
if word in commonWords:
return True
return False
This produces the following 2-grams that were found more than twice in the text
('united states', 10), ('executive department', 4), ('general governm
ent', 4), ('called upon', 3), ('government should', 3), ('whole count
ry', 3), ('mr jefferson', 3), ('chief magistrate', 3), ('same causes'
, 3), ('legislative body', 3)
Appropriately enough, the first two items in the list are “United States” and
“executive department,” which we would expect for a presidential inauguration
It’s important to note that we are using a list of common words from relatively
modern times to filter the results, which might not be appropriate given that the text
was written in 1841. However, because we’re using only the first 100 or so words
on the list — which we can assume are more stable over time than, say, the last 100
words — and we appear to be getting satisfactory results, we can likely save
ourselves the effort of tracking down or creating a list of the most common words
from 1841 (although such an effort might be interesting).
Now that some key topics have been extracted from the text, how does this help us
write text summaries? One way is to search for the first sentence that contains each
“popular” n-gram, the theory being that the first instance will yield a satisfactory
overview of the body of the content. The first five most popular 2-grams yield these
bullet points:
The Constitution of the United States is the instrument containing this grant of
power to the several departments composing the government.
Such a one was afforded by the executive department constituted by the
The general government has seized upon none of the reserved rights of the
Called from a retirement which I had supposed was to continue for the
residue of my life to fill the chief executive office of this great and free
nation, I appear before you, fellow-citizens, to take the oaths which the
constitution prescribes as a necessary qualification for the performance of its
duties; and in obedience to a custom coeval with our government and what I
believe to be your expectations I proceed to present to you a summary of the
principles which will govern me in the discharge of the duties which I shall
be called upon to perform.
The presses in the necessary employment of the government should never be
used to clear the guilty or to varnish crime.
Sure, it might not be published in CliffsNotes any time soon, but considering that
the original document was 217 sentences in length, and the fourth sentence (“Called
from a retirement...”) condenses the main subject down fairly well, it’s not too bad
for a first pass.
Markov Models
You might have heard of Markov text generators. They’ve become popular for
entertainment purposes, as in the Twitov app, as well as their use for generating
real-sounding spam emails to fool detection systems.
All of these text generators are based on the Markov model, which is often used to
analyze large sets of random events, where one discrete event is followed by
another discrete event with a certain probability.
For example, we might build a Markov model of a weather system as illustrated in
Figure 8-1.
Figure 8-1. Markov model describing a theoretical weather system
In this model, each sunny day has a 70% chance of the following day also being
sunny, with a 20% chance of the following day being cloudy with a mere 10%
chance of rain. If the day is rainy, there is a 50% chance of rain the following day, a
25% chance of sun, and a 25% chance of clouds.
All percentages leading away from any one node must add up to exactly 100%.
No matter how complicated the system, there must always be a 100% chance
that it can lead somewhere else in the next step.
Although there are only three possibilities for the weather at any given time, you
can use this model to generate an infinite list of weather states.
Only the state of the current node you are on influences where you will go to
next. If you’re on the “sunny” node, it doesn’t matter if the preceding 100 days
were sunny or rainy — the chances of sun the next day are exactly the same:
It might be more difficult to reach some nodes than others. The math behind this
is reasonably complicated, but it should be fairly easy to see that “rainy” (with
less than “100%” worth of arrows pointing toward it) is a much less likely state
to reach in this system, at any given point in time, than “sunny” or “cloudy.”
Obviously, this is a very simple system, and Markov models can grow arbitrarily
large in size. In fact, Google’s page rank algorithm is based partly on a Markov
model, with websites represented as nodes and inbound/outbound links represented
as connections between nodes. The “likelihood” of landing on a particular node
represents the relative popularity of the site. That is, if our weather system
represented an extremely small Internet, “rainy” would have a low page rank, while
“cloudy” would have a high page rank.
With all of this in mind, let’s bring it back down to a more concrete example:
analyzing and writing text.
Again using the inauguration speech of William Henry Harrison analyzed in the
previous example, we can write the following code that generates arbitrarily long
Markov chains (with the chain length set to 100) based on the structure of its text:
from urllib.request import urlopen
from random import randint
def wordListSum(wordList):
sum = 0
for word, value in wordList.items():
sum += value
return sum
def retrieveRandomWord(wordList):
randIndex = randint(1, wordListSum(wordList))
for word, value in wordList.items():
randIndex -= value
if randIndex <= 0:
return word
def buildWordDict(text):
#Remove newlines and quotes
text = text.replace("\n", " ");
text = text.replace("\"", "");
#Make sure punctuation marks are treated as their own "words,"
#so that they will be included in the Markov chain
punctuation = [',','.',';',':']
for symbol in punctuation:
text = text.replace(symbol, " "+symbol+" ");
words = text.split(" ")
#Filter out empty words
words = [word for word in words if word != ""]
wordDict = {}
for i in range(1, len(words)):
if words[i-1] not in wordDict:
#Create a new dictionary for this word
wordDict[words[i-1]] = {}
if words[i] not in wordDict[words[i-1]]:
wordDict[words[i-1]][words[i]] = 0
wordDict[words[i-1]][words[i]] = wordDict[words[i-1]][words[
i]] + 1
return wordDict
text = str(urlopen("")
.read(), 'utf-8')
wordDict = buildWordDict(text)
#Generate a Markov chain of length 100
length = 100
chain = ""
currentWord = "I"
for i in range(0, length):
chain += currentWord+" "
currentWord = retrieveRandomWord(wordDict[currentWord])
The output of this code changes every time it is run, but here’s an example of the
uncannily nonsensical text it will generate:
I sincerely believe in Chief Magistrate to make all necessary sacrifices and
oppression of the remedies which we may have occurred to me in the arrangement
and disbursement of the democratic claims them , consolatory to have been best
political power in fervently commending every other addition of legislation , by
the interests which violate that the Government would compare our aboriginal
neighbors the people to its accomplishment . The latter also susceptible of the
Constitution not much mischief , disputes have left to betray . The maxim which
may sometimes be an impartial and to prevent the adoption or
So what’s going on in the code?
The function buildWordDict takes in the string of text, which was retrieved from
the Internet. It then does some cleaning and formatting, removing quotes, and putting
spaces around other punctuation so it is effectively treated as a separate word.
After this, it builds a two-dimensional dictionary — a “dictionary of dictionaries
— that has the following form:
{word_a : {word_b : 2, word_c : 1, word_d : 1},
word_e : {word_b : 5, word_d : 2},...}
In this example dictionary, “word_a” was found four times, two instances of which
were followed by “word_b,” one instance followed by “word_c,” and one instance
followed by “word_d.” “Word_e” was followed seven times, five times by
“word_b” and twice by “word_d.”
If we were to draw a node model of this result, the node representing “word_a”
would have a 50% arrow pointing toward “word_b” (which followed it two out of
four times), a 25% arrow pointing toward “word_c,” and a 25% arrow pointing
toward “word_d.”
Once this dictionary is built up, it can be used as a lookup table to see where to go
next, no matter which word in the text you happen to be on.3 Using the sample
dictionary of dictionaries, we might currently be on “word_e,” which means that
we’ll pass the dictionary {word_b : 5, word_d: 2} to the
retrieveRandomWord function. This function in turn retrieves a random word from
the dictionary, weighted by the number of times it occurs.
By starting with a random starting word (in this case, the ubiquitous “I”), we can
traverse through the Markov chain easily, generating as many words as we like.
Six Degrees of Wikipedia: Conclusion
In Chapter 3, we created a scraper that collects links from one Wikipedia article to
the next, starting with the article on Kevin Bacon, and stores them in a database.
Why are we bringing it up again? Because it turns out the problem of choosing a
path of links that starts on one page and ends up on the target page (i.e., finding a
string of pages between and is the
same as finding a Markov chain where both the first word and last word are
defined. These sorts of problems are directed graph problems, where A → B does
not necessarily mean that B → A. The word “football” might often be followed by
the word “player,” but you’ll find that the word “player” is much less often
followed by the word “football.” Although Kevin Bacon’s Wikipedia article links
to the article on his home city, Philadelphia, the article on Philadelphia does not
reciprocate by linking back to him.
In contrast, the original Six Degrees of Kevin Bacon game is an undirected
graph problem. If Kevin Bacon starred in Flatliners with Julia Roberts, then Julia
Roberts necessarily starred in Flatliners with Kevin Bacon, so the relationship
goes both ways (it has no “direction”). Undirected graph problems tend to be less
common in computer science than directed graph problems, and both are
computationally difficult to solve.
Although much work has been done on these sorts of problems and multitudes of
variations on them, one of the best and most common ways to find shortest paths in
a directed graph — and thus find paths between the Wikipedia article on Kevin
Bacon and all other Wikipedia articles — is through a breadth-first search.
A breadth-first search is performed by first searching all links that link directly to
the starting page. If those links do not contain the target page (the page you are
searching for), then a second level of links — pages that are linked by a page that is
linked by the starting page — is searched. This process continues until either the
depth limit (6 in this case) is reached or the target page is found.
A complete solution to the breadth-first search, using a table of links as described
in Chapter 5, is as follows:
from urllib.request import urlopen
from bs4 import BeautifulSoup
import pymysql
conn = pymysql.connect(host='', unix_socket='/tmp/mysql.sock',
user='root', passwd=None, db='mysql', charset='utf8')
cur = conn.cursor()
cur.execute("USE wikipedia")
class SolutionFound(RuntimeError):
def __init__(self, message):
self.message = message
def getLinks(fromPageId):
cur.execute("SELECT toPageId FROM links WHERE fromPageId = %s",
if cur.rowcount == 0:
return None
return [x[0] for x in cur.fetchall()]
def constructDict(currentPageId):
links = getLinks(currentPageId)
if links:
return dict(zip(links, [{}]*len(links)))
return {}
#The link tree may either be empty or contain multiple links
def searchDepth(targetPageId, currentPageId, linkTree, depth):
if depth == 0:
#Stop recursing and return, regardless
return linkTree
if not linkTree:
linkTree = constructDict(currentPageId)
if not linkTree:
#No links found. Cannot continue at this node
return {}
if targetPageId in linkTree.keys():
print("TARGET "+str(targetPageId)+" FOUND!")
raise SolutionFound("PAGE: "+str(currentPageId))
for branchKey, branchValue in linkTree.items():
#Recurse here to continue building the tree
linkTree[branchKey] = searchDepth(targetPageId, branchKey,
branchValue, depth-1)
except SolutionFound as e:
raise SolutionFound("PAGE: "+str(currentPageId))
return linkTree
searchDepth(134951, 1, {}, 4)
print("No solution found")
except SolutionFound as e:
The functions getLinks and constructDict are helper functions that retrieve
links from the database given a page, and format those links into a dictionary. The
main function, searchDepth, works recursively to simultaneously construct and
search a tree of links, working one level at a time. It operates on the following
If the given recursion limit has been reached (i.e., if it has called itself too many
times), return without doing any work.
If the dictionary of links it has been given is empty, populate it with links for the
current page. If the current page has no links, return.
If the current page contains a link to the page we are searching for, throw an
exception that alerts copies of itself on up the stack that the solution has been
found. Each stack then prints the current page it is on, and throws the exception
again, resulting in a perfect list of pages leading to the solution being printed on
the screen.
If the solution is not found, call itself while subtracting one from the depth count
in order to search the next level of links.
The output for searching for a link between the page on Kevin Bacon (page ID 1, in
my database) and the page on Eric Idle (page ID 78520 in my database) is:
PAGE: 156224
PAGE: 155545
This translates into the relationship of links: Kevin Bacon → San Diego Comic
Con International → Brian Froud → Terry Jones → Eric Idle
In addition to solving “6 degree” problems and modeling which words tend to
follow which other words in sentences, directed and undirected graphs can be used
to model a variety of different situations encountered in web scraping. Which
websites link to which other websites? Which research papers cite which other
research papers? Which products tend to be shown with which other products on a
retail site? What is the strength of this link? Is the link reciprocal?
Recognizing these fundamental types of relationships can be extremely helpful for
making models, visualizations, and predictions based on scraped data.
Natural Language Toolkit
So far, this chapter has focused primarily on the statistical analysis of words in
bodies of text. Which words are most popular? Which words are unusual? Which
words are likely to come after which other words? How are they grouped together?
What we are missing is understanding, to the extent that we can, what the words
The Natural Language Toolkit (NLTK) is a suite of Python libraries designed to
identify and tag parts of speech found in natural English text. Its development began
in 2000, and over the past 15 years dozens of developers around the world have
contributed to the project. Although the functionality it provides is tremendous
(entire books are devoted to NLTK), this section will focus on just a few of its
Installation and Setup
The NLTK module can be installed the same as other Python modules, either by
downloading the package through the NLTK website directly or by using any
number of third-party installers with the keyword “nltk.” For complete installation
instructions, you see the NLTK website.
After installing the module it’s a good idea to download its preset text repositories
so you can try out some of the features more easily. Type this on the Python
command line:
>>> import nltk
This opens the NLTK Downloader (Figure 8-2).
Figure 8-2. The NLTK Downloader lets you browse and download optional packages and
text libraries associated with the NLTK module
I recommend installing all of the available packages. Because everything is text
based the downloads are very small; you never know what you’ll end up using, and
you can easily uninstall packages at any time.
Statistical Analysis with NLTK
NLTK is great for generating statistical information about word counts, word
frequency, and word diversity in sections of text. If all you need is a relatively
straightforward calculation (e.g., the number of unique words used in a section of
text), importing NLTK might be overkill — it’s a very large module. However, if
you need to do relatively extensive analysis of a text, you have a number of
functions at your fingertips that will give you just about any metric you want.
Analysis with NLTK always starts with the Text object. Text objects can be
created from simple Python strings in the following way:
from nltk import word_tokenize
from nltk import Text
tokens = word_tokenize("Here is some not very interesting text")
text = Text(tokens)
The input for the word_tokenize function can be any Python text string. If you
don’t have any long strings handy but still want to play around with the
features, NLTK has quite a few books already built into the library, which can
accessed using the import function:
from import *
This loads the nine books:
*** Introductory Examples for the NLTK Book ***
Loading text1, ..., text9 and sent1, ..., sent9
Type the name of the text or sentence to view it.
Type: 'texts()' or 'sents()' to list the materials.
text1: Moby Dick by Herman Melville 1851
text2: Sense and Sensibility by Jane Austen 1811
text3: The Book of Genesis
text4: Inaugural Address Corpus
text5: Chat Corpus
text6: Monty Python and the Holy Grail
text7: Wall Street Journal
text8: Personals Corpus
text9: The Man Who Was Thursday by G . K . Chesterton 1908
We will be working with text6, “Monty Python and the Holy Grail” (the screenplay
for the 1975 movie) in all of the following examples.
Text objects can be manipulated much like normal Python arrays, as if they were an
array containing words of the text. Using this property, you can count the number of
unique words in a text and compare it against the total number of words:
>>> len(text6)/len(words)
The preceding shows that each word in the script was used about eight times on
average. You can also put the text into a frequency distribution object to see what
some of the most common words are and the frequencies for various words:
>>> from nltk import FreqDist
>>> fdist = FreqDist(text6)
>>> fdist.most_common(10)
[(':', 1197), ('.', 816), ('!', 801), (',', 731), ("'", 421), ('[', 3
19), (']', 312), ('the', 299), ('I', 255), ('ARTHUR', 225)]
>>> fdist["Grail"]
Because this is a screenplay, some artifacts of how it is written can pop up. For
instance, “ARTHUR” in all caps crops up frequently because it appears before
each of King Arthur’s lines in the script. In addition, a colon (:) appears before
every single line, acting as a separator between the name of the character and the
character’s line. Using this fact, we can see that there are 1,197 lines in the movie!
What we have called 2-grams in previous chapters NLTK refers to as bigrams
(from time to time you might also hear 3-grams referred to as “trigrams,” but I
personally prefer n-gram rather than bigram or trigram). You can create, search, and
list 2-grams extremely easily:
from nltk import bigrams
bigrams = bigrams(text6)
bigramsDist = FreqDist(bigrams)
bigramDist[("Sir", "Robin")]
To search for the 2-grams “Sir Robin” we need to break it up into an array (“Sir”,
“Robin”), to match the way the 2-grams are represented in the frequency
distribution. There is also a trigrams module that works in the exact same way.
For the general case, you can also import the ngrams module:
from nltk import ngrams
fourgrams = ngrams(text6, 4)
fourgramsDist = FreqDist(fourgrams)
fourgramsDist[("father", "smelt", "of", "elderberries")]
Here, the ngrams function is called to break up a text object into n-grams of any
size, governed by the second parameter. In this case, I’m breaking the text into 4grams. Then, I can demonstrate that the phrase “father smelt of elderberries” occurs
in the screenplay exactly once.
Frequency distributions, text objects, and n-grams also can be iterated through and
operated on in a loop. The following prints out all 4-grams that begin with the word
“coconut,” for instance:
from import *
from nltk import ngrams
fourgrams = ngrams(text6, 4)
for fourgram in fourgrams:
if fourgram[0] == "coconut":
The NLTK library has a vast array of tools and objects designed to organize, count,
sort, and measure large swaths of text. Although we’ve barely scratched the surface
of their uses, most of these tools are very well designed and operate rather
intuitively for someone familiar with Python.
Lexicographical Analysis with NLTK
So far, we’ve compared and categorized all the words we’ve encountered based
only on the value they represent by themselves. There is no differentiation between
homonyms or the context in which the words are used.
Although some people might be tempted to dismiss homonyms as rarely
problematic, you might be surprised at how frequently they crop up. Most native
English speakers probably don’t even register that a word is a homonym, much less
consider that it might possibly be confused for another word in a different context.
“He was objective in achieving his objective of writing an objective philosophy,
primarily using verbs in the objective case” is easy for humans to parse but might
make a web scraper think the same word is being used four times and cause it to
simply discard all the information about the meaning behind each word.
In addition to sussing out parts of speech, being able to distinguish a word being
used in one way versus another might be useful. For example, you might want to
look for company names made up of common English words, or analyze someone’s
opinions about a company “ACME Products is good” and “ACME Products is not
bad” can have the same meaning, even if one sentence uses “good” and the other
uses “bad”.
NLTK uses by default a popular system of tagging parts of speech developed by the
University of Pennsylvania’s Penn Treebank Project. Although some of the tags make
sense (e.g., CC is a coordinating conjunction), others can be confusing (e.g., RP is a
particle). Use the following as a reference for the tags referred to in this section:
Coordinating conjunction
Cardinal number
Existential “there”
Foreign word
Preposition, subordinating conjunction
Adjective, comparative
Adjective, superlative
List item marker
Noun, singular or mass
Noun, plural
Proper noun, singular
NNPS Proper noun, plural
Possessive ending
Personal pronoun
PRP$ Possessive pronoun
Adverb, comparative
Adverb, superlative
SYM Symbol
Verb, base form
Verb, past tense
Verb, gerund or present participle
Verb, past participle
Verb, non-third person singular present
Verb, third person singular present
WDT wh-determiner
Possessive wh-pronoun
WRB Wh-adverb
In addition to measuring language, NLTK can assist in finding meaning in the words
based on context and its own very large dictionaries. At a basic level, NLTK can
identify parts of speech:
>>> from import *
>>> from nltk import word_tokenize
>>> text = word_tokenize("Strange women lying in ponds distributing swords is no
basis for a system of government. Supreme executive power derives from a
from the masses, not from some farcical aquatic ceremony.")
>>> from nltk import pos_tag
>>> pos_tag(text)
[('Strange', 'NNP'), ('women', 'NNS'), ('lying', 'VBG'), ('in', 'IN')
, ('ponds', 'NNS'), ('distributing', 'VBG'), ('swords', 'NNS'), ('is'
, 'VBZ'), ('no', 'DT'), ('basis', 'NN'), ('for', 'IN'), ('a', 'DT'),
('system', 'NN'), ('of', 'IN'), ('government', 'NN'), ('.', '.'), ('S
upreme', 'NNP'), ('executive', 'NN'), ('power', 'NN'), ('derives', 'N
NS'), ('from', 'IN'), ('a', 'DT'), ('mandate', 'NN'), ('from', 'IN'),
('the', 'DT'), ('masses', 'NNS'), (',', ','), ('not', 'RB'), ('from'
, 'IN'), ('some', 'DT'), ('farcical', 'JJ'), ('aquatic', 'JJ'), ('cer
emony', 'NN'), ('.', '.')]
Each word is separated into a tuple containing the word and a tag identifying the
part of speech (see the preceding Penn Treebank Tags sidebar for more information
about these tags). Although this might seem like a very straightforward lookup the
complexity needed to perform the task correctly becomes apparent with the
following example:
>>> text = word_tokenize("The dust was thick so he had to dust")
>>> pos_tag(text)
[('The', 'DT'), ('dust', 'NN'), ('was', 'VBD'), ('thick', 'JJ'), ('so
', 'RB'), ('he', 'PRP'), ('had', 'VBD'), ('to', 'TO'), ('dust', 'VB')
Notice that the word “dust” is used twice in the sentence: once as a noun, and again
as a verb. NLTK identifies both usages correctly, based on their context in the
sentence. NLTK identifies parts of speech using a context-free grammar defined by
the English language. Context-free grammars are, essentially, sets of rules that
define which things are allowed to follow which other things in ordered lists. In
this case, they define which parts of speech are allowed to follow which other
parts of speech. Whenever an ambiguous word such as “dust” is encountered, the
rules of the context-free grammar is consulted and an appropriate part of speech
that follows the rules is selected.
You can have NLTK generate brand-new context-free grammars when training it, for
example, on a foreign language. If you tag large sections of text by hand in the
language using the appropriate Penn Treebank Tags, you can feed them back into
NLTK and train it to properly tag other text it might encounter. This type of training is
a necessary component of any machine-learning activity that we will revisit in
Chapter 11, when training scrapers to recognize CAPTCHA characters.
So, what’s the point of knowing whether a word is a verb or a noun in a given
context? It might be neat in a computer science research lab, but how does it help
with web scraping?
A very common problem in web scraping deals with search. You might be scraping
text off a site and want to be able to search it for instances of the word “google,”
but only when it’s being used as a verb, not a proper noun. Or you might be looking
only for instances of the company Google and don’t want to rely on people’s
correct use of capitalization in order to find those instances. Here, the pos_tag
function can be extremely useful:
from nltk import word_tokenize, sent_tokenize, pos_tag
sentences = sent_tokenize("Google is one of the best companies in the world.
I constantly google myself to see what I'm up to.")
nouns = ['NN', 'NNS', 'NNP', 'NNPS']
for sentence in sentences:
if "google" in sentence.lower():
taggedWords = pos_tag(word_tokenize(sentence))
for word in taggedWords:
if word[0].lower() == "google" and word[1] in nouns:
This prints only sentences that contain the word “google” (or “Google”) as some
sort of a noun, not a verb. Of course, you could be more specific and demand that
only instances of Google tagged with “NNP” (a proper noun) are printed, but even
NLTK makes mistakes at times, and it can be good to leave yourself a little wiggle
room, depending on the application.
Much of the ambiguity of natural language can be resolved using NLTK’s pos_tag
function. By searching text not just for instances of your target word or phrase but
instances of your target word or phrase plus its tag, you can greatly increase the
accuracy and effectiveness of your scraper’s searches.
Additional Resources
Processing, analyzing, and understanding natural language by machine is one of the
most difficult tasks in computer science, and countless volumes and research
papers have been written on the subject. I hope that the coverage here will inspire
you to think beyond conventional web scraping, or at least give some initial
direction about where to begin when undertaking a project that requires natural
language analysis.
There are many excellent resources on introductory language processing and
Python’s Natural Language Toolkit. In particular, Steven Bird, Ewan Klein, and
Edward Loper’s book Natural Language Processing with Python presents both a
comprehensive and introductory approach to the topic.
In addition, James Pustejovsky and Amber Stubbs’ Natural Language Annotations
for Machine Learning provides a slightly more advanced theoretical guide. You’ll
need a knowledge of Python to implement the lessons; the topics covered work
perfectly with Python’s Natural Language Toolkit.
Although many of the techniques described in this chapter can be applied to all or
most languages, it’s okay for now to focus on natural language processing in
English only. Tools such as Python’s Natural Language Toolkit, for example, focus
on English. Fifty-six percent of the Internet is still in English (with German
following at a mere 6%, according to But who
knows? English’s hold on the majority of the Internet will almost certainly change
in the future, and further updates may be necessary in the next few years.
See “A Picture Is Worth a Thousand (Coherent) Words: Building a Natural
Description of Images,” November 17, 2014 (
The exception is the last word in the text because nothing follows the last word. In
our example text, the last word is a period (.), which is convenient because it has
215 other occurrences in the text and so does not represent a dead end. However, in
real-world implementations of the Markov generator, the last word of the text might
be something you need to account for.
Chapter 9. Crawling Through Forms and Logins
One of the first questions that comes up when you start to move beyond the basics
of web scraping is: “How do I access information behind a login screen?” The Web
is increasingly moving toward interaction, social media, and user-generated
content. Forms and logins are an integral part of these types of sites and almost
impossible to avoid. Fortunately, they are also relatively easy to deal with.
Up until this point, most of our interactions with web servers in our example
scrapers has consisted of using HTTP GET to request information. In this chapter,
we’ll focus on the POST method which pushes information to a web server for
storage and analysis.
Forms basically give users a way to submit a POST request that the web server can
understand and use. Just like link tags on a website help users format GET requests,
HTML forms help them format POST requests. Of course, with a little bit of coding,
it is possible to simply create these requests ourselves and submit them with a
Python Requests Library
Although it’s possible to navigate web forms using only the Python core libraries,
sometimes a little syntactic sugar makes life a lot sweeter. When you start to do
more than a basic GET request with urllib it can help to look outside the Python
core libraries.
The Requests library is excellent at handling complicated HTTP requests, cookies,
headers, and much more.
Here’s what Requests creator Kenneth Reitz has to say about Python’s core tools:
Python’s standard urllib2 module provides most of the HTTP capabilities you
need, but the API is thoroughly broken. It was built for a different time — and a
different web. It requires an enormous amount of work (even method overrides)
to perform the simplest of tasks.
Things shouldn’t be this way. Not in Python.
As with any Python library, the Requests library can be installed with any thirdparty Python library manager, such as pip, or by downloading and installing the
source file.
Submitting a Basic Form
Most web forms consist of a few HTML fields, a submit button, and an “action”
page, where the actual form processing is done. The HTML fields usually consist of
text but might also contain a file upload or some other non-text content.
Most popular websites block access to their login forms in their robots.txt file
(Appendix C discusses the legality of scraping such forms), so to play it safe I’ve
constructed a series of different types of forms and logins at
that you can run your web scrapers against. The most basic of these forms is
located at
The entirety of the form is:
<form method="post" action="processing.php">
First name: <input type="text" name="firstname"><br>
Last name: <input type="text" name="lastname"><br>
<input type="submit" value="Submit">
A couple of things to notice here: first, the name of the two input fields are
firstname and lastname. This is important. The names of these fields determine
the names of the variable parameters that will be POSTed to the server when the
form is submitted. If you want to mimic the action that the form will take when
POSTing your own data, you need to make sure that your variable names match up.
The second thing to note is that the action of the form is actually at
processing.php (the absolute path is Any post requests to
the form should be made on this page, not on the page that the form itself resides.
Remember: the purpose of HTML forms is only to help website visitors format
proper requests to send to the page that does the real action. Unless you are doing
research to format the request itself, you don’t need to bother much with the page
that the form can be found on.
Submitting a form with the Requests library can be done in four lines, including the
import and the instruction to print the content (yes, it’s that easy):
import requests
params = {'firstname': 'Ryan', 'lastname': 'Mitchell'}
r ="", data=params)
After the form is submitted, the script should return with the page’s content:
Hello there, Ryan Mitchell!
This script can be applied to many simple forms encountered on the Internet. The
form to sign up for the O’Reilly Media newsletter, for example, looks like this:
<form action="
quicksignup.cgi" id="example_form2" method="POST">
<input name="client_token" type="hidden" value="oreilly" />
<input name="subscribe" type="hidden" value="optin" />
<input name="success_url" type="hidden" value="
newsletter-thankyou.html" />
<input name="error_url" type="hidden" value="
newsletter-signup-error.html" />
<input name="topic_or_dod" type="hidden" value="1" />
<input name="source" type="hidden" value="orm-home-t1-dotd" />
<input class="email_address long" maxlength="200" name=
"email_addr" size="25" type="text" value=
"Enter your email here" />
<button alt="Join" class="skinny" name="submit" onclick=
"return addClickTracking('orm','ebook','rightrail','dod'
);" value="submit">Join</button>
Although it can look daunting at first, remember that in most cases (we’ll cover the
exceptions later), you’re only looking for two things:
The name of the field (or fields) you want to submit with data (in this case, the
name is email_address)
The action attribute of the form itself; that is, the page that the form actually
posts to (in this
Just add in the required information and run it:
import requests
params = {'email_addr': '[email protected]'}
r ="
quicksignup.cgi", data=params)
In this case, the website returned is simply another form to fill out, before you can
actually make it onto O’Reilly’s mailing list, but the same concept could be applied
to that form as well. However, I would request that you use your powers for good,
and not spam the publisher with invalid signups, if you want to try this at home.
Radio Buttons, Checkboxes, and Other Inputs
Obviously, not all web forms are a collection of text fields followed by a submit
button. Standard HTML contains a wide variety of possible form input fields: radio
buttons, checkboxes, and select boxes, to name a few. In HTML5, there’s the
addition of sliders (range input fields), email, dates, and more. With custom
JavaScript fields the possibilities are endless, with colorpickers, calendars, and
whatever else the developers come up with next.
Regardless of the seeming complexity of any sort of form field, there are only two
things you need to worry about: the name of the element and its value. The
element’s name can be easily determined by looking at the source code and finding
the name attribute. The value can sometimes be trickier, as it might be populated by
JavaScript immediately before form submission. Colorpickers, as an example of a
fairly exotic form field, will likely have a value of something like #F03030.
If you’re unsure of the format of an input field’s value, there are a number of tools
you can use to track the GET and POST requests your browser is sending to and from
sites. The best and perhaps most obvious way to track GET requests, as mentioned
before, is to simply look at the URL of a site. If the URL is something like:
You know that this corresponds to a form of this type:
<form method="GET" action="someProcessor.php">
<input type="someCrazyInputType" name="thing1" value="foo" />
<input type="anotherCrazyInputType" name="thing2" value="bar" />
<input type="submit" value="Submit" />
Which corresponds to the Python parameter object:
{'thing1':'foo', 'thing2':'bar'}
You can see this in Figure 9-1.
If you’re stuck with a complicated-looking POST form, and you want to see exactly
which parameters your browser is sending to the server, the easiest way is to use
your browser’s inspector or developer tool to view them.
Figure 9-1. The Form Data section, highlighted in a box, shows the POST parameters
“thing1” and “thing2” with their values “foo” and “bar”
The Chrome developer tool can be accessed via the menu by going to
View → Developer → Developer Tools. It provides a list of all queries that your
browser produces while interacting with the current website and can be a good
way to view the composition of these queries in detail.
Submitting Files and Images
Although file uploads are common on the Internet, file uploads are not something
often used in web scraping. It is possible, however, that you might want to write a
test for your own site that involves a file upload. At any rate, it’s a useful thing to
know how to do.
There is a practice file upload form at http://pythonscraping/files/form2.html. The
form on the page has the following markup:
<form action="processing2.php" method="post" enctype="multipart/form-data">
Submit a jpg, png, or gif: <input type="file" name="image"><br>
<input type="submit" value="Upload File">
Except for the <input> tag having the type attribute file, it looks essentially the
same as the text-based forms used in the previous examples. Fortunately, the way
the forms are used by the Python Requests library is also very similar:
import requests
files = {'uploadFile': open('../files/Python-logo.png', 'rb')}
r ="",
Note that in lieu of a simple string, the value submitted to the form field (with the
name uploadFile) is now a Python File object, as returned by the open function.
In this example, I am submitting an image file, stored on my local machine, at the
path ../files/Python-logo.png, relative to where the Python script is being run from.
Yes, it’s really that easy!
Handling Logins and Cookies
So far, we’ve mostly discussed forms that allow you submit information to a site or
let you to view needed information on the page immediately after the form. How is
this different from a login form, which lets you exist in a permanent “logged in”
state throughout your visit to the site?
Most modern websites use cookies to keep track of who is logged in and who is
not. Once a site authenticates your login credentials a it stores in your browser a
cookie, which usually contains a server-generated token, timeout, and tracking
information. The site then uses this cookie as a sort of proof of authentication,
which is shown to each page you visit during your time on the site. Before the
widespread use of cookies in the mid-90s, keeping users securely authenticated and
tracking them was a huge problem for websites.
Although cookies are a great solution for web developers, they can be problematic
for web scrapers. You can submit a login form all day long, but if you don’t keep
track of the cookie the form sends back to you afterward, the next page you visit
will act as though you’ve never logged in at all.
I’ve created a simple login form at (the username can be
anything, but the password must be “password”).
This form is processed at, and contains a link to the “main
site” page,
If you attempt to access the welcome page or the profile page without logging in
first, you’ll get an error message and instructions to log in first before continuing.
On the profile page, a check is done on your browser’s cookies to see whether its
cookie was set on the login page.
Keeping track of cookies is easy with theRequests library:
import requests
params = {'username': 'Ryan', 'password': 'password'}
r ="", params)
print("Cookie is set to:")
print("Going to profile page...")
r = requests.get("",
Here I am sending the login parameters to the welcome page, which acts as the
processor for the login form. I retrieve the cookies from the results of the last
request, print the result for verification, and then send them to the profile page by
setting the cookies argument.
This works well for simple situations, but what if you’re dealing with a more
complicated site that frequently modifies cookies without warning, or if you’d
rather not even think about the cookies to begin with? The Requests session
function works perfectly in this case:
import requests
session = requests.Session()
params = {'username': 'username', 'password': 'password'}
s ="", params)
print("Cookie is set to:")
print("Going to profile page...")
s = session.get("")
In this case, the session object (retrieved by calling requests.Session()) keeps
track of session information, such as cookies, headers, and even information about
protocols you might be running on top of HTTP, such as HTTPAdapters.
Requests is a fantastic library, second perhaps only to Selenium (which we’ll cover
in Chapter 10) in the completeness of what it handles without programmers having
to think about it or write the code themselves. Although it might be tempting to sit
back and let the library do all the work, it’s extremely important to always be
aware of what the cookies look like and what they are controlling when writing
web scrapers. It could save many hours of painful debugging or figuring out why a
website is behaving strangely!
HTTP Basic Access Authentication
Before the advent of cookies, one popular way to handle logins was with HTTP
basic access authentication. You still see it from time to time, especially on highsecurity or corporate sites, and with some APIs. I’ve created a page
at that has this type of
authentication (Figure 9-2).
Figure 9-2. The user must provide a username and password to get to the page protected
by basic access authentication
As usual with these examples, you can log in with any username, but the password
must be “password.”
The Requests package contains an auth module specifically designed to handle
HTTP authentication:
import requests
from requests.auth import AuthBase
from requests.auth import HTTPBasicAuth
auth = HTTPBasicAuth('ryan', 'password')
r ="", auth=
Although this appears to be a normal POST request, an HTTPBasicAuth object is
passed as the auth argument in the request. The resulting text will be the page
protected by the username and password (or an Access Denied page, if the request
Other Form Problems
Web forms are a hot point of entry for malicious bots. You don’t want bots creating
user accounts, taking up valuable server processing time, or submitting spam
comments on a blog. For this reason, there are often a number of security features
that are incorporated into HTML forms on modern websites that might not be
immediately apparent.
For help with CAPTCHAs, check out Chapter 11, which covers image processing
and text recognition in Python.
If you encounter a mysterious error, or the server is rejecting your form submission
for an unknown reason, check out Chapter 12, which covers honey pots, hidden
fields, and other security measures that websites take to protect their forms.
Chapter 10. Scraping JavaScript
Client-side scripting languages are languages that are run in the browser itself,
rather than on a web server. The success of a client-side language depends on your
browser’s ability to interpret and execute the language correctly. (This is why it’s
so easy to disable JavaScript in your browser.)
Partly due to the difficulty of getting every browser manufacturer to agree on a
standard, there are far fewer client-side languages than there are server-side
languages. This is a good thing when it comes to web scraping: the fewer languages
there are to deal with the better.
For the most part, there are only two languages you’ll frequently encounter online:
ActionScript (which is used by Flash applications) and JavaScript. ActionScript is
used far less frequently today than it was 10 years ago, and is often used to stream
multimedia files, as a platform for online games, or to display “intro” pages for
websites that haven’t gotten the hint that no one wants to watch an intro page. At
any rate, because there isn’t much demand for scraping Flash pages, I will instead
focus on the client-side language that’s ubiquitous in modern web pages:
JavaScript is, by far, the most common and most well-supported client-side
scripting language on the Web today. It can be used to collect information for user
tracking, submit forms without reloading the page, embed multimedia, and even
power entire online games. Even deceptively simple-looking pages can often
contain multiple pieces of JavaScript. You can find it embedded between
<script> tags in the page’s source code:
alert("This creates a pop-up using JavaScript");
A Brief Introduction to JavaScript
Having at least some idea of what is going on in the code you are scraping can be
immensely helpful. With that in mind it’s a good idea to familiarize yourself with
JavaScript is a weakly typed language, with a syntax that is often compared to C++
and Java. Although certain elements of the syntax, such as operators, loops, and
arrays, might be similar, the weak typing and script-like nature of the language can
make it a difficult beast to deal with for some programmers.
For example, the following recursively calculates values in the Fibonacci
sequence, and prints them out to the browser’s developer console:
function fibonacci(a, b){
var nextNum = a + b;
console.log(nextNum+" is in the Fibonacci sequence");
if(nextNum < 100){
fibonacci(b, nextNum);
fibonacci(1, 1);
Notice that all variables are demarcated by a preceding var. This is similar to the
$ sign in PHP, or the type declaration (int, String, List, etc.) in Java or C++.
Python is unusual in that it doesn’t have this sort of explicit variable declaration.
JavaScript is also extremely good at passing around functions just like variables:
var fibonacci = function() {
var a = 1;
var b = 1;
return function () {
var temp = b;
b = a + b;
a = temp;
return b;
var fibInstance = fibonacci();
console.log(fibInstance()+" is in the Fibonacci sequence");
console.log(fibInstance()+" is in the Fibonacci sequence");
console.log(fibInstance()+" is in the Fibonacci sequence");
This might seem daunting at first, but it becomes simple if you think in terms of
lambda expressions (covered in Chapter 2). The variable fibonacci is defined as
a function. The value of its function returns a function that prints increasingly large
values in the Fibonacci sequence. Each time it is called, it returns the Fibonacci209
calculating function, which executes again and increases the values in the function.
Although it might seem convoluted at first glance, some problems, such as
calculating Fibonacci values, tend to lend themselves to patterns like this. Passing
around functions as variables is also extremely useful when it comes to handling
user actions and callbacks, and it is worth getting comfortable with this style of
programming when it comes to reading JavaScript.
Common JavaScript Libraries
Although the core JavaScript language is important to know, you can’t get very far
on the modern Web without using at least one of the language’s many third-party
libraries. You might see one or more of these commonly used libraries when
looking at page source code.
Executing JavaScript using Python can be extremely time consuming and processor
intensive, especially if you’re doing it on a large scale. Knowing your way around
JavaScript and being able to parse it directly (without needing to execute it to
acquire the information) can be extremely useful and save you a lot of headaches.
jQuery is an extremely common library, used by 70% of the most popular Internet
sites and about 30% of the rest of the Internet.1 A site using jQuery is readily
identifiable because it will contain an import to jQuery somewhere in its code,
such as:
<script src=""></
If you find jQuery is found on a site, you must be careful when scraping it. jQuery is
adept at dynamically creating HTML content that appears only after the JavaScript
is executed. If you scrape the page’s content using traditional methods, you will
retrieve only the preloaded page that appears before the JavaScript has created the
content (we’ll cover this scraping problem in more detail in “Ajax and Dynamic
In addition, these pages are more likely to contain animations, interactive content,
and embedded media that might make scraping challenging.
Google Analytics
Google Analytics is used by about 50% of all websites,2 making it perhaps the most
common JavaScript library and the most popular user tracking tool on the Internet.
In fact, both and use Google
It’s easy telling whether a page is using Google Analytics. It will have JavaScript
at the bottom similar to the following (taken from the O’Reilly Media site):
<!-- Google Analytics -->
<script type="text/javascript">
var _gaq = _gaq || [];
_gaq.push(['_setAccount', 'UA-4591498-1']);
_gaq.push(['_setDomainName', '']);
_gaq.push(['_addIgnoredRef', '']);
_gaq.push(['_setSiteSpeedSampleRate', 50]);
(function() { var ga = document.createElement('script'); ga.type =
'text/javascript'; ga.async = true; ga.src = ('https:' ==
document.location.protocol ? 'https://ssl' : 'http://www') +
''; var s =
s.parentNode.insertBefore(ga, s); })();
This script handles Google Analytics, specific cookies used to track your visit from
page to page. This can sometimes be a problem for web scrapers that are designed
to execute JavaScript and handle cookies (such as those that use Selenium,
discussed later in this chapter).
If a site uses Google Analytics or a similar web analytics system and you do not
want the site to know that it’s being crawled or scraped, make sure to discard any
cookies used for analytics or discard cookies altogether.
Google Maps
If you’ve spent any time on the Internet, you’ve almost certainly seen Google Maps
embedded in a website. Its API makes it extremely easy to embed maps with
custom information on any site.
If you’re scraping any sort of location data, understanding how Google Maps works
makes it easy to obtain well-formatted latitude/longitude coordinates and even
addresses. One of the most common ways to denote a location in Google Maps is
through a marker (also known as a pin).
Markers can be inserted into any Google Map using code such as the following:
var marker = new google.maps.Marker({
position: new google.maps.LatLng(-25.363882,131.044922),
map: map,
title: 'Some marker text'
Python makes it easy to extract all instances of coordinates that occur between
google.maps.LatLng( and ) to obtain a list of latitude/longitude coordinates.
Using Google’s “reverse Geocoding” API, you can resolve these coordinate pairs
to addresses that are well formatted for storage and analysis.
Ajax and Dynamic HTML
Until now the only way we’ve had of communicating with a web server is to send it
some sort of HTTP request via the retrieval of a new page. If you’ve ever submitted
a form or retrieved information from a server without reloading the page, you’ve
likely used a website that uses Ajax.
Contrary to what some believe, Ajax is not a language but a group of technologies
used to accomplish a certain task (much like web scraping, come to think of
it). Ajax stands for Asynchronous JavaScript and XML, and is used to send
information to and receive from a web server without making a separate page
request. Note: you should never say, “This website will be written in Ajax.” It
would be correct to say, “This form will use Ajax to communicate with the web
Like Ajax, dynamic HTML or DHTML is a collection of technologies used for a
common purpose. DHTML is HTML code, CSS language, or both that change due
to client-side scripts changing HTML elements on the page. A button might appear
only after the user moves the cursor, a background color might change on a click, or
an Ajax request might trigger a new block of content to load.
Note that although the word “dynamic” is generally associated with words like
“moving,” or “changing,” the presence of interactive HTML components, moving
images, or embedded media does not necessarily make a page DHTML, even
though it might look dynamic. In addition, some of the most boring, static-looking
pages on the Internet can have DHTML processes running behind the scenes that
depend on the use of JavaScript to manipulate the HTML and CSS.
If you scrape a large number of different websites, you will soon run into a
situation in which the content you are viewing in your browser does not match the
content you see in the source code you’re retrieving from the site. You might view
the output of your scraper and scratch your head, trying to figure out where
everything you’re seeing on the exact same page in your browser has disappeared
The web page might also have a loading page that appears to redirect you to
another page of results, but you’ll notice that the page’s URL never changes when
this redirect happens.
Both of these are caused by a failure of your scraper to execute the JavaScript that
is making the magic happen on the page. Without the JavaScript, the HTML just sort
of sits there, and the site might look very different than what it looks like in your
web browser, which executes the JavaScript without problem.
There are several giveaways that a page might be using Ajax or DHTML to
change/load the content, but in situations like this there are only two solutions:
scrape the content directly from the JavaScript, or use Python packages capable of
executing the JavaScript itself, and scrape the website as you view it in your
Executing JavaScript in Python with Selenium
Selenium is a powerful web scraping tool developed originally for website testing.
These days it’s also used when the accurate portrayal of websites — as they appear
in a browser — is required. Selenium works by automating browsers to load the
website, retrieve the required data, and even take screenshots or assert that certain
actions happen on the website.
Selenium does not contain its own web browser; it requires integration with thirdparty browsers in order to run. If you were to run Selenium with Firefox, for
example, you would literally see a Firefox instance open up on your screen,
navigate to the website, and perform the actions you had specified in the code.
Although this might be neat to watch, I prefer my scripts to run quietly in the
background, so I use a tool called PhantomJS in lieu of an actual browser.
PhantomJS is what is known as a “headless” browser. It loads websites into
memory and executes JavaScript on the page, but does it without any graphic
rendering of the website to the user. By combining Selenium with PhantomJS, you
can run an extremely powerful web scraper that handles cookies, JavaScript,
headers, and everything else you need with ease.
You can download the Selenium library from its website or use a third-party
installer such as pip to install it from the command line.
PhantomJS can be downloaded from its website. Because PhantomJS is a full
(albeit headless) browser and not a Python library, it does require a download and
installation to use and cannot be installed with pip.
Although there are plenty of pages that use Ajax to load data (notably Google), I’ve
created a sample page at to run our scrapers against. This
page contains some sample text, hardcoded into the page’s HTML, that is replaced
by Ajax-generated content after a two-second delay. If we were to scrape this
page’s data using traditional methods, we’d only get the loading page, without
actually getting the data that we want.
The Selenium library is an API called on the a WebDriver. The WebDriver is a bit
like a browser in that it can load websites, but it can also be used like a
BeautifulSoup object to find page elements, interact with elements on the page
(send text, click, etc.), and do other actions to drive the web scrapers.
The following code retrieves text behind an Ajax “wall” on the test page:
from selenium import webdriver
import time
driver = webdriver.PhantomJS(executable_path='')
In previous chapters, we’ve selected page elements using BeautifulSoup selectors, such
as find and findAll. Selenium uses an entirely new set of selectors to find an element
in a WebDriver’s DOM, although they have fairly straightforward names.
In the example, we used the selector find_element_by_id, although the following other
selectors would have worked as well:
Of course, if you want to select multiple elements on the page, most of these element
selectors can return a Python list of elements simply by using elements (i.e., make it
Of course, if you still want to use BeautifulSoup to parse this content you can, by
using WebDriver’s page_source function, which returns the page’s source, as viewed
by the DOM at that current time, as a string:
pageSource = driver.page_source
bsObj = BeautifulSoup(pageSource)
This creates a new Selenium WebDriver, using the PhantomJS library, which tells
the WebDriver to load a page and then pauses execution for three seconds before
looking at the page to retrieve the (hopefully loaded) content.
Depending on the location of your PhantomJS installation, you might also need to
explicitly point Selenium in the right direction when creating a new PhantomJS
driver = webdriver.PhantomJS(executable_path='/path/to/download/
If everything is configured correctly the script should take a few seconds to run and
result in the following text:
Here is some important text you want to retrieve!
A button to click!
Note that although the page itself contains an HTML button, Selenium’s
.text function retrieves the text value of the button in the same way that it retrieves
all other content on the page.
If the time.sleep pause is changed to one second instead of three, the text returned
changes to the original:
This is some content that will appear on the page while it's loading.
You don't care about scraping this.
Although this solution works, it is somewhat inefficient and implementing it could
cause problems on a large scale. Page load times are inconsistent, depending on the
server load at any particular millisecond, and natural variations occur in
connection speed. Although this page load should take just over two seconds, we’re
giving it an entire three seconds to make sure that it loads completely. A more
efficient solution would repeatedly check for the existence of some element on a
fully loaded page and return only when that element exists.
This code uses the presence of the button with id loadedButton to declare that the
page has been fully loaded: from selenium import webdriver.
from import By
from import WebDriverWait
from import expected_conditions as EC
driver = webdriver.PhantomJS(executable_path='')
element = WebDriverWait(driver, 10).until(
EC.presence_of_element_located((By.ID, "loadedButton")))
There are several new imports in this script, most notably WebDriverWait and
expected_conditions, both of which are combined here to form what Selenium
calls an implicit wait.
An implicit wait differs from an explicit wait in that it waits for some state in the
DOM to occur before continuing, while an explicit wait defines a hardcoded time
like in the previous example, which has a wait of three seconds. In an implicit wait,
the triggering DOM state is defined by expected_condition (note that the import
is cast to EC here, a common convention used for brevity). Expected conditions can
be many things in the Selenium library, among them:
An alert box pops up
An element (such as a text box) is put into a “selected” state
The page’s title changes, or some text is now displayed on the page or in a
specific element
An element is now visible to the DOM, or an element disappears from the DOM
Of course, most of these expected conditions require that you specify an element to
watch for in the first place. Elements are specified using locators. Note that
locators are not the same as selectors (see previous sidebar for more on selectors).
A locator is an abstract query language, using the By object, which can be used in a
variety of ways, including to make selectors.
In the following example code, a locator is used to find elements with the id
EC.presence_of_element_located((By.ID, "loadedButton"))
Locators can also be used to create selectors, using the find_element WebDriver
print(driver.find_element(By.ID, "content").text)
Which is, of course, functionally equivalent to the line in the example code:
If you do not need to use a locator, don’t; it will save you an import. However, it is
a very handy tool that is used for a variety of applications and has a great degree of
The following locator selection strategies can used with the By object:
Used in the example; finds elements by their HTML id attribute.
Used to find elements by their HTML class attribute. Why is this function
CLASS_NAME and not simply CLASS? Using the form object.CLASS would
create problems for Selenium’s Java library, where .class is a reserved
method. In order to keep the Selenium syntax consistent between different
languages, CLASS_NAME was used instead.
Find elements by their class, id, or tag name, using
the #idName, .className, tagName convention.
Finds HTML <a> tags by the text they contain. For example, a link that says
“Next” can be selected using (By.LINK_TEXT, "Next").
Similar to LINK_TEXT, but matches on a partial string.
Finds HTML tags by their name attribute. This is handy for HTML forms.
Finds HTML tags by their tag name.
Uses an XPath expression (the syntax of which is described in the upcoming
sidebar) to select matching elements.
XPath (short for XML Path) is a query language used for navigating and selecting
portions of an XML document. It was founded by the W3C in 1999 and is
occasionally used in languages such as Python, Java, and C# when dealing with XML
Although BeautifulSoup does not support XPath, many of the other libraries in this
book do. It can often be used in the same way as CSS selectors (such as
mytag#idname), although it is designed to work with more generalized XML documents
rather than HTML documents in particular.
There are four major concepts in the XPath syntax:
Root nodes versus non-root nodes
will select the div node only if it is at the root of the document
selects all divs anywhere in the document
Attribute selection
selects any nodes with the attribute href
selects all links in the document that point to
Selection of nodes by position
selects the third link in the document
selects the last table in the document
//a[position() < 3]
selects the first three links in the document
Asterisks (*) match any set of characters or nodes, and can be used in a variety
of situations
selects all children of tr tags in all tables (this is good for selecting
cells using both th and td tags)
selects all div tags that have any attributes
Of course, there are many advanced features of the XPath syntax. Over the years it
has developed into a relatively complicated query language, with boolean logic,
functions (such as position()), and a variety of operators not discussed here.
If you have an HTML or XML selection problem that cannot be addressed by the
functions shown here, see Microsoft’s XPath syntax page.
Handling Redirects
Client-side redirects are page redirects that are executed in your browser by
JavaScript, rather than a redirect performed on the server, before the page content
is sent. It can sometimes be tricky to tell the difference when visiting a page in your
web browser. The redirect might happen so fast that you don’t notice any delay in
loading time and assume that a client-side redirect is actually a server-side
However, when scraping the Web, the difference is obvious. A server-side redirect,
depending on how it is handled, can be easily traversed by Python’s urllib library
without any help from Selenium (for more information on doing this, see
Chapter 3). Client-side redirects won’t be handled at all unless something is
actually executing the JavaScript.
Selenium is capable of handling these JavaScript redirects in the same way that it
handles other JavaScript execution; however, the primary issue with these redirects
is when to stop page execution — that is, how to tell when a page is done
redirecting. A demo page at gives an example of this type of
redirect, with a two-second pause.
We can detect that redirect in a clever way by “watching” an element in the DOM
when the page initially loads, then repeatedly calling that element until Selenium
throws a StaleElementReferenceException; that is, the element is no longer
attached to the page’s DOM and the site has redirected:
from selenium import webdriver
import time
from selenium.webdriver.remote.webelement import WebElement
from selenium.common.exceptions import StaleElementReferenceException
def waitForLoad(driver):
elem = driver.find_element_by_tag_name("html")
count = 0
while True:
count += 1
if count > 20:
print("Timing out after 10 seconds and returning")
elem == driver.find_element_by_tag_name("html")
except StaleElementReferenceException:
driver = webdriver.PhantomJS(executable_path='<Path to Phantom JS>')
This script checks the page every half second, with a timeout of 10 seconds,
although the times used for the checking time and timeout can be easily adjusted up
or down as needed.
Dave Methvin’s blog post, "The State of jQuery 2014,” January 13, 2014,
contains a more detailed breakdown of the statistics.
W3Techs, “Usage Statistics and Market Share of Google Analytics for Websites”
Chapter 11. Image Processing and Text Recognition
From Google’s self-driving cars to vending machines that recognize counterfeit
currency, machine vision is a huge field with far-reaching goals and implications.
In this chapter, we will focus on one very small aspect of the field: text recognition,
specifically how to recognize and use text-based images found online by using a
variety of Python libraries.
Using an image in lieu of text is a common technique when you don’t want text to be
found and read by bots. This is often seen on contact forms when an email address
is partially or completely rendered as an image. Depending on how skillfully it is
done, it might not even be noticeable to human viewers but bots have a very
difficult time reading these images and the technique is enough to stop most
spammers from acquiring your email address.
CAPTCHAs, of course, take advantage of the fact that users can read security
images but most bots can’t. Some CAPTCHAs are more difficult than others, an
issue we’ll tackle later in this book.
But CAPTCHAs aren’t the only place on the Web where scrapers need image-totext translation assistance. Even in this day and age, many documents are simply
scanned from hard copies and put on the Web, making these documents inaccessible
as far as much of the Internet is concerned, although they are “hiding in plain sight.”
Without image-to-text capabilities, the only way to make these documents
accessible is for a human to type them up by hand — and nobody has time for that.
Translating images into text is called optical character recognition, or OCR. There
are a few major libraries that are able to perform OCR, and many other libraries
that support them or are built on top of them. This system of libraries can get fairly
complicated at times, so I recommend you read the next section before attempting
any of the exercises in this chapter.
Overview of Libraries
Python is a fantastic language for image processing and reading, image-based
machine-learning, and even image creation. Although there are a large number of
libraries that can be used for image processing, we will focus on two: Pillow and
Either can be installed by downloading from their websites and installing from
source ( and or by using a third-party
Python installer such as pip and using the keywords “pillow” and “pytesseract,”
Although Pillow might not be the most fully featured image-processing library, it
has all of the features you are likely to need and then some unless you are doing
research or rewriting Photoshop in Python, or something like that. It is also a very
well-documented piece of code that is extremely easy to use.
Forked off the Python Imaging Library (PIL) for Python 2.x, Pillow adds support for
Python 3.x. Like its predecessor, Pillow allows you to easily import and manipulate
images with a variety of filters, masks, and even pixel-specific transformations:
from PIL import Image, ImageFilter
kitten ="kitten.jpg")
blurryKitten = kitten.filter(ImageFilter.GaussianBlur)"kitten_blurred.jpg")
In the preceding example, the image kitten.jpg will open in your default image
viewer with a blur added to it and will also be saved in its blurrier state as
kitten_blurred.jpg in the same directory.
We will use Pillow to perform preprocessing on images to make them more
machine readable but as mentioned before, there are many other things you can do
with the library aside from these simple manipulations. For more information,
check out the Pillow documentation.
Tesseract is an OCR library. Sponsored by Google (a company obviously well
known for its OCR and machine learning technologies), Tesseract is widely
regarded to be the best, most accurate, open source OCR system available.
In addition to being accurate, it is also extremely flexible. It can be trained to
recognize any number of fonts (as long as those fonts are relatively consistent
within themselves, as we will see soon) and it can be expanded to recognize any
Unicode character.
Unlike libraries we’ve used so far in this book, Tesseract is a command-line tool
written in Python, rather than a library used with an import statement. After
installation, it must be run with the tesseract command, from outside of Python.
Installing Tesseract
For Windows users there is a convenient executable installer. As of this writing,
the current version is 3.02, although newer versions should be fine as well.
Linux users can install Tesseract with apt-get:
$sudo apt-get tesseract-ocr
Installing Tesseract on a Mac is slightly more complicated, although it can be done
easily with many third-party installers such as Homebrew, which was used in
Chapter 5 to install MySQL. For example, you can install Homebrew and use it to
install Tesseract in two lines:
$ruby -e "$(curl -fsSL \
$brew install tesseract
Tesseract also can be installed from the source, on the project’s download page.
To use some features of Tesseract, such as training the software to recognize new
characters later in this section, you will also need to set a new environment
variable, $TESSDATA_PREFIX, to let it know where data files are stored.
You can do this in most Linux system and on Mac OS X using:
$export TESSDATA_PREFIX=/usr/local/share/
Note that /usr/local/share/ is the default data location for Tesseract, although
you should check to make sure that this is the case for your own installation.
Similarly, on Windows, you can use the following to use the environment variable:
#setx TESSDATA_PREFIX C:\Program Files\Tesseract OCR\
Although NumPy is not required for straightforward OCR, you will need it if you
want to train Tesseract to recognize additional character sets or fonts introduced
later in this chapter. NumPy is a very powerful library used for linear algebra and
other large-scale math applications. NumPy works well with Tesseract because of
its ability to mathematically represent and manipulate images as large arrays of
As always, NumPy can be installed using any third-party Python installer such as
$pip install numpy
Processing Well-Formatted Text
With any luck, most of the text that you’ll need to process will be relatively clean
and well formatted. Well-formatted text generally meets several requirements,
although the line between what is “messy” and what is “well formatted” can be
In general, well-formatted text:
Is written in one standard font (excluding handwriting fonts, cursive fonts, or
excessively “decorative” fonts)
If copied or photographed has extremely crisp lines, with no copying artifacts or
dark spots
Is well-aligned, without slanted letters
Does not run off the image, nor is there cut-off text or margins on the edges of
the image
Some of these things can be fixed in preprocessing. For instance, images can be
converted to grayscale, brightness and contrast can be adjusted, and the image can
be cropped and rotated as needed. However, there are some fundamental
limitations that might require more extensive training. See “Reading CAPTCHAs
and Training Tesseract”.
Figure 11-1 an ideal example of well-formatted text.
Figure 11-1. Sample text saved as a .tiff file, to be read by Tesseract
You can run Tesseract from the command line to read this file and write the results
to a text file:
$tesseract text.tif textoutput | cat textoutput.txt
The output is a line of information about the Tesseract library to indicate that it is
running, followed by the contents of the newly created textoutput.txt:
Tesseract Open Source OCR Engine v3.02.02 with Leptonica
This is some text, written in Arial, that will be read by
Tesseract. Here are some symbols: [email protected]#$%"&‘()
You can see that the results are mostly accurate, although the symbols “^” and “*”
were interpreted as a double quote and single quote, respectively. On the whole,
though, this lets you read the text fairly comfortably.
After blurring the image text, creating some JPG compression artifacts, and adding
a slight background gradient, the results get much worse (see Figure 11-2).
Figure 11-2. Unfortunately, many of the documents you will encounter on the Internet will
look more like this than the previous example
Tesseract is not able to deal with this image nearly as well mainly due to the
background gradient and produces the following output:
This is some text, written In Arlal, that"
Tesseract. Here are some symbols: _
Notice that the text is cut off as soon as the background gradient makes the text more
difficult to distinguish, and that the last character from each line is wrong as
Tesseract tries futilely to make sense of it. In addition, the JPG artifacts and
blurring make it difficult for Tesseract to distinguish between a lowercase “i” and
an uppercase “I” and the number “1”.
This is where using a Python script to clean your images first comes in handy. Using
the Pillow library, we can create a threshold filter to get rid of the gray in the
background, bring out the text, and make the image clearer for Tesseract to read:
from PIL import Image
import subprocess
def cleanFile(filePath, newFilePath):
image =
#Set a threshold value for the image, and save
image = image.point(lambda x: 0 if x<143 else 255)
#call tesseract to do OCR on the newly created image["tesseract", newFilePath, "output"])
#Open and read the resulting data file
outputFile = open("output.txt", 'r')
cleanFile("text_2.png", "text_2_clean.png")
The resulting image, automatically created as text_2_clean.png, is shown in
Figure 11-3:
Figure 11-3. This image was created by passing the previous “messy” version of the
image through a threshold filter
Apart from some barely legible or missing punctuation, the text is readable, at least
to us. Tesseract gives it its best shot:
This us some text‘ written In Anal, that will be read by
Tesseract Here are some symbols: [email protected]#$%"&'()
The periods and commas, being extremely small, are the first victims of this image
wrangling and nearly disappear, both from our view and Tesseract’s. There’s also
the unfortunate misinterpretation of “Arial” as “Anal,” the result of Tesseract
interpreting the “r” and the “i” as the single character “n.”
Still, it’s an improvement over the previous version in which nearly half of the text
was cut off.
Tesseract’s greatest weakness seems to be backgrounds with varying brightness.
Tesseract’s algorithms attempt to adjust the contrast of the image automatically
before reading the text, but you can probably get better results doing this yourself
with a tool like the Pillow library.
Images you should definitely fix before submitting to Tesseract are those that are
tilted, have large areas of non-text, or have other problems.
Scraping Text from Images on Websites
Using Tesseract to read text from an image on your hard drive might not seem all
that exciting, but it can be a very powerful tool when used with a web scraper.
Images can inadvertently obfuscate text on websites (as with the JPG copy of a
menu on a local restaurant site), but they can also purposefully hide the text, as I’ll
show in the next example.
Although Amazon’s robots.txt file allows scraping of the site’s product pages, book
previews typically don’t get picked up by passing bots. That’s because the book
previews are loaded via user-triggered Ajax scripts, and the images are carefully
hidden under layers of divs; in fact, to the average site visitor they probably look
more like a flash presentation than image files. Of course, even if we could get to
the images, there’s the not-so-small matter of reading them as text.
The following script accomplishes just this feat: it navigates to the large-print
edition1 of Tolstoy’s War and Peace, opens the reader, collects image URLs, and
then systematically downloads, reads, and prints the text from each one. Because
this is relatively complex code that draws on multiple concepts from previous
chapters, I’ve added comments throughout to make it a little easier to understand
what’s going on:
import time
from urllib.request import urlretrieve
import subprocess
from selenium import webdriver
#Create new Selenium driver
driver = webdriver.PhantomJS(executable_path='<Path to Phantom JS>')
#Click on the book preview button
imageList = set()
#Wait for the page to load
#While the right arrow is available for clicking, turn through pages
while "pointer" in driver.find_element_by_id("sitbReaderRightPageTurner")
#Get any new pages that have loaded (multiple pages can load at once,
#but duplicates will not be added to a set)
pages = driver.find_elements_by_xpath("//div[@class='pageImage']/div/img")
for page in pages:
image = page.get_attribute("src")
#Start processing the images we've collected URLs for with Tesseract
for image in sorted(imageList):
urlretrieve(image, "page.jpg")
p = subprocess.Popen(["tesseract", "page.jpg", "page"],
f = open("page.txt", "r")
As we have experienced with the Tesseract reader before, this prints many long
passages of the book perfectly, as seen in the preview of page 6:
“A word of friendly advice, mon
cher. Be off as soon as you can,
that's all I have to tell you. Happy
he who has ears to hear. Good-by,
my dear fellow. Oh, by the by!” he
shouted through the doorway after
Pierre, “is it true that the countess
has fallen into the clutches of the
holy fathers of the Society of jesus?”
Pierre did not answer and left Rostopchin's room more sullen and angry than he had ever before shown
However, when the text appears on a colored background on the front and back
covers of the book, it becomes incomprehensible:
WEI‘ nrrd Peace
Len Nlkelayevldu Iolfluy
Readmg shmdd be ax
wlnvame asnossxble Wenfler
an mm m our cram: Llhvary
— Leo Tmsloy was a Russian rwovelwst
I and moval phflmopher med lur
A ms Ideas 01 nonviolenx reswslance m 5 We range 0, “and”
Of course, you can use the Pillow library to selectively clean images, but this can
be extremely labor intensive for a process that was designed to be as human-free as
The next section discusses another approach to solving the problem of mangled
text, particularly if you don’t mind investing a little time up front in training
Tesseract. By providing Tesseract with a large collection of text images with
known values, Tesseract can be “taught” to recognize the same font in the future
with far greater precision and accuracy, even despite occasional background and
positioning problems in the text.
Reading CAPTCHAs and Training Tesseract
Although the word “CAPTCHA” is familiar to most, far fewer people know what it
stands for: Computer Automated Public Turing test to tell Computers and Humans
Apart. Its unwieldy acronym hints at its rather unwieldy role in obstructing
otherwise perfectly usable web interfaces, as both humans and nonhuman robots
often struggle to solve CAPTCHA tests.
The Turing test was first described by Alan Turing in his 1950 paper, “Computing
Machinery and Intelligence.” In the paper, he described a setup in which a human
being could communicate with both humans and artificial intelligence programs
through a computer terminal. If the human was unable to distinguish the humans
from the AI programs during a casual conversation, the AI programs would be
considered to have passed the Turing test, and the artificial intelligence, Turing
reasoned, would be genuinely “thinking” for all intents and purposes.
It’s ironic that in the last 60 years we’ve gone from using these tests to test
machines to using them to test ourselves, with mixed results. Google’s notoriously
difficult reCAPTCHA, currently the most popular among security-conscious
websites, blocks as many as 25% of legitimate human users from accessing a site.2
Most other CAPTCHAs are somewhat easier. Drupal, a popular PHP-based content
management system, for example, has a popular CAPTCHA module, that can
generate CAPTCHA images of varying difficulty. The default image looks like
Figure 11-4.
Figure 11-4. An example of the default text CAPTCHA for Drupal’s CAPTCHA project
What makes this CAPTCHA so easy for both humans and machines to read,
compared to other CAPTCHAs?
Characters do not overlap each other, nor do they cross into each other’s space
horizontally. That is, it is possible to draw a neat rectangle around each
character without overlapping any other character.
There are no background images, lines, or other distracting garbage that could
confuse an OCR program.
It is not obvious from this image, but there are few variations on the font that the
CAPTCHA uses. It alternates between a clean sans-serif font (as seen in the
characters “4” and “M”) and a handwriting-style font, (as seen in the characters
“m,” “C,” and “3”).
There is a high contrast between the white background and the dark-colored
This CAPTCHA does throw a few curves, though, that make it challenging for OCR
programs to read:
Both letters and numbers are used, increasing the number of potential characters.
The randomized tilt of the letters might confuse OCR software, but remains very
easy for humans to read.
The relatively strange handwriting font presents particular challenges, with extra
lines in the “C” and “3” and an unusually small lowercase “m” requiring extra
training for computers to get the hang of.
When we run Tesseract over this image using the command:
$tesseract captchaExample.png output
we get this output.txt file:
It got the 4, C, and 3 right, but it’s clearly not going to be able to fill out a
CAPTCHA-protected field any time soon.
Training Tesseract
In order to train Tesseract to recognize writing, whether it’s an obscure and
difficult-to-read font or a CAPTCHA, you need to give it multiple examples of each
This is the part where you might want to queue up a good podcast or movie because
it’s going to be a couple of hours of fairly boring work. The first step is to
download multiple examples of your CAPTCHA into a single directory. The
number of examples you compile will depend on the complexity of the CAPTCHA;
I used 100 sample files (a total of 500 characters, or about 8 examples per symbol,
on average) for my CAPTCHA training, and that seems to work fairly well.
Tip: I recommend naming the image after the CAPTCHA solution it represents (i.e.,
4MmC3.jpg). I’ve found that this helps to do quick error-checking across large
numbers of files at once — you can view all files as thumbnails and compare the
image against its image name easily. This helps greatly in error checking in
subsequent steps, as well.
The second step is to tell Tesseract exactly what each character is and where it is in
the image. This involves creating box files, one for every CAPTCHA image. A box
file looks like this:
15 26 33 55 0
38 13 67 45 0
79 15 101 26 0
111 33 136 60 0
147 17 176 45 0
The first symbol is the character represented, the next four numbers represent
coordinates for a rectangular box outlining the image, and the last number is a
“page number” used for training with multipage documents (0 for us).
Obviously, these box files are not fun to create by hand, but there are a variety of
tools to help you out. I like the online tool Tesseract OCR Chopper because it
requires no installation or additional libraries, runs on any machine that has a
browser, and is relatively easy to use: upload the image, click the “add” button at
the bottom if you need additional boxes, adjust the size of the boxes if
necessary, and copy and paste your new box file text into a new file.
Box files must be saved in plain text, with the .box file extension. As with the
image files, it’s handy to name the box files by the CAPTCHA solutions they
represent (e.g., Again, this makes it easy to double-check the
contents of the .box file text against the name of the file, and then again against the
image file it is paired with if you sort all of the files in your data directory by their
Again, you’ll need to create about 100 of these files to ensure that you have enough
data. Also, Tesseract does occasionally discard files as being unreadable, so you
might want some buffer room on top of that. If you find that your OCR results aren’t
quite as good as you’d like, or Tesseract is stumbling over certain characters, it’s a
good debugging step to create additional training data and try again.
After creating a data folder full of .box files and image files, copy this data into a
backup folder before doing any further manipulation on it. Although running training
scripts over the data is unlikely to delete anything, it’s better safe than sorry when
hours’ worth of work put into .box file creation is involved. Additionally, it’s good
to be able to scrap a messy directory full of compiled data and try again.
There are half a dozen steps to performing all the data analysis and creating the
training files required for Tesseract. There are tools that do this for you given
corresponding source image and .box files, but none at the moment for Tesseract
3.02, unfortunately.
I’ve written a solution in Python that operates over a file containing both image and
box files and creates all necessary training files automatically.
The main settings and steps that this program takes can be seen in its main and
runAll methods:
def main(self):
languageName = "eng"
fontName = "captchaFont"
directory = "<path to images>"
def runAll(self):
The only three variables you’ll need to set here are fairly straightforward:
The three-letter language code that Tesseract uses to understand which
language it’s looking at. In most cases, you’ll probably want to use “eng” for
The name for your chosen font. This can be anything, but must be a single
word without spaces.
The directory containing all of your image and box files
I recommend you make this an absolute path, but if you use a relative path, it
will need to be relative to where you are running the Python code from. If it is
absolute, you can run the code from anywhere on your machine.
Let’s take a look at the individual functions used.
creates a required file, font_properties, that lets Tesseract know
about the new font we are creating:
captchaFont 0 0 0 0 0
This file consists of the name of the font, followed by 1s and 0s indicating whether
italic, bold, or other versions of the font should be considered (training fonts with
these properties is an interesting exercise, but unfortunately outside the scope of
this book).
creates higher-contrast versions of all image files found, converts
them to grayscale, and performs other operations that make the image files easier to
read by OCR programs. If you are dealing with CAPTCHA images with visual
garbage that might be easy to filter out in post-processing, here would be the place
to add that additional processing.
renames all of your .box files and their corresponding image files
with the names required by Tesseract (the file numbers here are sequential digits to
keep multiple files separate):
looks at all of the created .box files and determines the total set
of characters available to be trained. The resulting Unicode file will tell you how
many different characters you’ve found, and could be a good way to quickly see if
you’re missing anything.
The next three functions, runShapeClustering, runMfTraining, and
runCtTraining, create the files shapetable, pfftable, and normproto,
respectively. These all provide information about the geometry and shape of each
character, as well as provide statistical information that Tesseract uses to calculate
the probability that a given character is one type or another.
Finally, Tesseract renames each of the compiled data folders to be prepended by
the required language name (e.g., shapetable is renamed to eng.shapetable) and
compiles all of those files into the final training data file eng.traineddata.
The only step you have to perform manually is move the created eng.traineddata
file to your tessdata root folder by using the following commands on Linux and
$cp /path/to/data/eng.traineddata $TESSDATA_PREFIX/tessdata
Following these steps, you should have no problem solving CAPTCHAs of the type
that Tesseract has now been trained for. Now when I ask Tesseract to read the
example image, I get the correct response:
$ tesseract captchaExample.png output;cat output.txt
Success! A significant improvement over the previous interpretation of the image as
This is just a quick overview of the full power of Tesseract’s font training and
recognition capabilities. If you are interested in extensively training Tesseract,
perhaps starting your own personal library of CAPTCHA training files, or sharing
new font recognition capabilities with the world, I recommend checking out the
Retrieving CAPTCHAs and Submitting Solutions
Many popular content management systems are frequently spammed with
registrations by bots that are preprogrammed with the well-known location of these
user registration pages. On, for instance, even a
CAPTCHA (admittedly, weak) does little to put a damper on the influx of
So how do these bots do it? We’ve successfully solved CAPTCHAs in images
sitting around on our hard drive, but how do we make a fully functioning bot? This
section ties together many techniques covered in previous chapters. If you haven’t
already, you should at least skim the chapters on submitting forms and downloading
Most image-based CAPTCHAs have several properties:
They are dynamically generated images, created by a server-side program. They
might have image sources that do not look like traditional images, such as <img
src="WebForm.aspx?id=8AP85CQKE9TJ">, but can be downloaded and
manipulated like any other image.
The solution to the image is stored in a server-side database.
Many CAPTCHAs time out if you take too long to solve them. This usually isn’t
a problem for bots, but queuing CAPTCHA solutions for later use, or other
practices that may delay the time between when the CAPTCHA was requested,
and when the solution is submitted, may not be successful.
The general approach to this is to download the CAPTCHA image file to your hard
drive, clean it, use Tesseract to parse the image, and return the solution under the
appropriate form parameter.
I’ve created a page at with a CAPTCHAprotected comment form for the purpose of writing a bot to defeat. The bot looks
like the following:
from urllib.request import urlretrieve
from urllib.request import urlopen
from bs4 import BeautifulSoup
import subprocess
import requests
from PIL import Image
from PIL import ImageOps
def cleanImage(imagePath):
image =
image = image.point(lambda x: 0 if x<143 else 255)
borderImage = ImageOps.expand(image,border=20,fill='white')
html = urlopen("")
bsObj = BeautifulSoup(html)
#Gather prepopulated form values
imageLocation = bsObj.find("img", {"title": "Image CAPTCHA"})["src"]
formBuildId = bsObj.find("input", {"name":"form_build_id"})["value"]
captchaSid = bsObj.find("input", {"name":"captcha_sid"})["value"]
captchaToken = bsObj.find("input", {"name":"captcha_token"})["value"]
captchaUrl = ""+imageLocation
urlretrieve(captchaUrl, "captcha.jpg")
p = subprocess.Popen(["tesseract", "captcha.jpg", "captcha"], stdout=
f = open("captcha.txt", "r")
#Clean any whitespace characters
captchaResponse =" ", "").replace("\n", "")
print("Captcha solution attempt: "+captchaResponse)
if len(captchaResponse) == 5:
params = {"captcha_token":captchaToken, "captcha_sid":captchaSid,
"form_id":"comment_node_page_form", "form_build_id": formBuildId,
"captcha_response":captchaResponse, "name":"Ryan Mitchell",
"subject": "I come to seek the Grail",
"...and I am definitely not a bot"}
r ="",
responseObj = BeautifulSoup(r.text)
if responseObj.find("div", {"class":"messages"}) is not None:
print(responseObj.find("div", {"class":"messages"}).get_text())
print("There was a problem reading the CAPTCHA correctly!")
Note there are two conditions on which this script fails: if Tesseract did not extract
exactly five characters from the image (because we know that all valid solutions to
this CAPTCHA must have five characters), or if it submits the form but the
CAPTCHA was solved incorrectly. The first case happens approximately 50% of
the time, at which point it does not bother submitting the form and fails with an
error message. The second case happens approximately 20% of the time, for a total
accuracy rate of about 30% (or about 80% accuracy for each character
encountered, over 5 characters).
Although this may seem low, keep in mind that there is usually no limit placed on
the number of times users are allowed to make CAPTCHA attempts, and that most
of these incorrect attempts can be aborted without needing to actually send the
form. When a form is actually sent, the CAPTCHA is accurate most of the time. If
that doesn’t convince you, also keep in mind that simple guessing would give you
an accuracy rate of .0000001%. Running a program three or four times rather than
guessing 900 million times is quite the time saver!
When it comes to processing text it hasn’t been trained on, Tesseract fares much
better with large-format editions of books, especially if the images are small. In the
next section we’ll discuss how to train Tesseract on different fonts, which can help
it read much smaller font sizes, including previews for non-large print book
Chapter 12. Avoiding Scraping Traps
There are few things more frustrating than scraping a site, viewing the output, and
not seeing the data that’s so clearly visible in your browser. Or submitting a form
that should be perfectly fine but gets denied by the web server. Or getting your IP
address blocked by a site for unknown reasons.
These are some of the most difficult bugs to solve, not only because they can be so
unexpected (a script that works just fine on one site might not work at all on
another, seemingly identical, site), but because they purposefully don’t have any tell
tale error messages or stack traces to use. You’ve been identified as a bot, rejected,
and you don’t know why.
In this book, I’ve written about a lot of ways to do tricky things on websites
(submitting forms, extracting and cleaning difficult data, executing JavaScript, etc.).
This chapter is a bit of a catchall in that the techniques stem from a wide variety of
subjects (HTTP headers, CSS, and HTML forms, to name a few). However, they all
have something in common: they are meant to overcome an obstacle put in place for
the sole purpose of preventing automated web scraping of a site.
Regardless of how immediately useful this information is to you at the moment, I
highly recommend you at least skim this chapter. You never know when it might
help you solve a very difficult bug or prevent a problem altogether.
A Note on Ethics
In the first few chapters of this book, I discussed the legal gray area that web
scraping inhabits, as well as some of the ethical guidelines to scrape by. To be
honest, this chapter is, ethically, perhaps the most difficult one for me to write. My
websites have been plagued by bots, spammers, web scrapers, and all manner of
unwanted virtual guests, as perhaps yours have been. So why teach people how to
build better bots?
There are a few reasons why I believe this chapter is important to include:
There are perfectly ethical and legally sound reasons to scrape some websites
that do not want to be scraped. In a previous job I had as a web scraper, I
performed an automated collection of information from websites that were
publishing clients’ names, addresses, telephone numbers, and other personal
information to the Internet without their consent. I used the scraped information
to make formal requests to the websites to remove this information. In order to
avoid competition, these sites guarded this information from scrapers vigilantly.
However, my work to ensure the anonymity of my company’s clients (some of
whom had stalkers, were the victims of domestic violence, or had other very
good reasons to want to keep a low profile) made a compelling case for web
scraping, and I was grateful that I had the skills necessary to do the job.
Although it is almost impossible to build a “scraper proof” site (or at least one
that can still be easily accessed by legitimate users), I hope that the information
in this chapter will help those wanting to defend their websites against
malicious attacks. Throughout, I will point out some of the weaknesses in each
web scraping technique, which you can use to defend your own site. Keep in
mind that most bots on the Web today are merely doing a broad scan for
information and vulnerabilities, and employing even a couple of simple
techniques described in this chapter will likely thwart 99% of them. However,
they are getting more sophisticated every month, and it’s best to be prepared.
Like most programmers, I simply don’t believe that withholding any sort of
educational information is a net positive thing to do.
While you’re reading this chapter, keep in mind that many of these scripts and
described techniques should not be run against every site you can find. Not only is
it not a nice thing to do, but you could wind up receiving a cease-and-desist letter
or worse. But I’m not going to pound you over the head with this every time we
discuss a new technique. So, for the rest of this book — as the philosopher Gump
once said — “That’s all I have to say about that.”
Looking Like a Human
The fundamental challenge for sites that do not want to be scraped is figuring out
how to tell bots from humans. Although many of the techniques sites use (such as
CAPTCHAs) can be difficult to fool, there are a few fairly easy things you can do
to make your bot look more human.
Adjust Your Headers
In Chapter 9, we used the requests module to handle forms on a website. The
requests module is also excellent for setting headers. HTTP headers are a list of
attributes, or preferences, sent by you every time you make a request to a web
server. HTTP defines dozens of obscure header types, most of which are not
commonly used. The following seven fields, however, are consistently used by
most major browsers when initiating any connection (shown with example data
from my own browser):
Mozilla/5.0 (Macintosh; Intel Mac OS X 10_9_5) AppleWebKit/537.36 (KHTML, like Gecko)
Chrome/39.0.2171.95 Safari/537.36
gzip, deflate, sdch
And here are the headers that a typical Python scraper using the default urllib
library might send:
If you’re a website administrator trying to block scrapers, which one are you more
likely to let through?
We installed the Requests module in Chapter 9, but if you haven’t done so, you can
find download links and instructions on the module’s website or use any third-party
Python module installer.
Fortunately, headers can be completely customized using the requests module.
The website is great for testing browser
properties viewable by servers. We’ll scrape this website to verify our cookie
settings with the following script:
import requests
from bs4 import BeautifulSoup
session = requests.Session()
headers = {"User-Agent":"Mozilla/5.0 (Macintosh; Intel Mac OS X 10_9_5)
AppleWebKit 537.36 (KHTML, like Gecko) Chrome",
url = "
req = session.get(url, headers=headers)
bsObj = BeautifulSoup(req.text)
The output should show that the headers are now the same ones set in the headers
dictionary object in the code.
Although it is possible for websites to check for “humanness” based on any of the
properties in HTTP headers, I’ve found that typically the only setting that really
matters is the User-Agent. It’s a good idea to keep this one set to something more
inconspicuous than Python-urllib/3.4, regardless of what project you are
working on. In addition, if you ever encounter an extremely suspicious website,
populating one of the commonly used but rarely checked headers such as AcceptLanguage might be the key to convincing it you’re a human.
Let’s say you want to write a machine learning language translator for a research
project, but lack large amounts of translated text to test with. Many large sites present
different translations of the same content, based on the indicated language preferences
in your headers. Simply changing Accept-Language:en-US to Accept-Language:fr in
your headers might get you a “Bonjour” from websites with the scale and budget to
handle translation (large international companies are usually a good bet).
Headers also can prompt websites to change the format of the content they are
presenting. For instance, mobile devices browsing the Web often see a very pareddown version of sites, lacking banner ads, Flash, and other distractions. If you try
changing your User-Agent to something like the following, you might find that sites get
a little easier to scrape!
User-Agent:Mozilla/5.0 (iPhone; CPU iPhone OS 7_1_2 like Mac OS X)
AppleWebKit/537.51.2 (KHTML, like Gecko) Version/7.0 Mobile/11D257
Handling Cookies
Handling cookies correctly can alleviate many of scraping problems, although
cookies can also be a double-edged sword. Websites that track your progression
through a site using cookies might attempt to cut off scrapers that display abnormal
behavior, such as completing forms too quickly, or visiting too many pages.
Although these behaviors can be disguised by closing and reopening connections to
the site, or even changing your IP address (see Chapter 14 for more information on
how to do that), if your cookie gives your identity away, your efforts of disguise
might be futile.
Cookies can also be very necessary to scrape a site. As shown in Chapter 9, staying
logged in on a site requires that you be able to hold and present a cookie from page
to page. Some websites don’t even require that you actually log in and get a new
version of a cookie every time — merely holding an old copy of a “logged in”
cookie and visiting the site is enough.
If you are scraping a single targeted website or a small number of targeted sites, I
recommend examining the cookies generated by those sites and considering which
ones you might want your scraper to handle. There are a number of browser plugins that can show you how cookies are being set as you visit and move around a
site. EditThisCookie, a Chrome extension, is one of my favorites.
Check out the code samples in “Handling Logins and Cookies” in Chapter 9 for
more information about handling cookies using the requests module. Of course,
because it is unable to execute JavaScript, the requests module will be unable to
handle many of the cookies produced by modern tracking software, such as Google
Analytics, which are set only after the execution of client-side scripts (or,
sometimes, based on page events, such as button clicks, that happen while browsing
the page). In order to handle these, you need to use the Selenium and PhantomJS
packages (we covered their installation and basic usage in Chapter 10).
You can view cookies by visiting any site (, in this
example) and calling get_cookies() on the webdriver:
from selenium import webdriver
driver = webdriver.PhantomJS(executable_path='<Path to Phantom JS>')
This provides the fairly typical array of Google Analytics cookies:
[{'value': '1', 'httponly': False, 'name': '_gat', 'path': '/', 'expi
ry': 1422806785, 'expires': 'Sun, 01 Feb 2015 16:06:25 GMT', 'secure'
: False, 'domain': ''}, {'value': 'GA1.2.161952506
2.1422806186', 'httponly': False, 'name': '_ga', 'path': '/', 'expiry
': 1485878185, 'expires': 'Tue, 31 Jan 2017 15:56:25 GMT', 'secure':
False, 'domain': ''}, {'value': '1', 'httponly': F
alse, 'name': 'has_js', 'path': '/', 'expiry': 1485878185, 'expires':
'Tue, 31 Jan 2017 15:56:25 GMT', 'secure': False, 'domain': 'pythons'}]
To manipulate cookies, you can call the delete_cookie(), add_cookie(), and
delete_all_cookies() functions. In addition, you can save and store cookies for
use in other web scrapers. Here’s an example to give you an idea how these
functions work together:
from selenium import webdriver
driver = webdriver.PhantomJS(executable_path='<Path to Phantom JS>')
savedCookies = driver.get_cookies()
driver2 = webdriver.PhantomJS(executable_path='<Path to Phantom JS>')
for cookie in savedCookies:
In this example, the first webdriver retrieves a website, prints the cookies, and then
stores them in the variable savedCookies. The second webdriver loads the same
website (technical note: it must load the website first so that Selenium knows
which website the cookies belong to, even if the act of loading the website does
nothing useful for us), deletes all of its cookies and replaces them with the saved
cookies from the first webdriver. When loading the page again, the timestamps,
codes, and other information in both cookies should be identical. According to
Google Analytics, this second webdriver is now identical to the first one.
Timing Is Everything
Some well-protected websites might prevent you from submitting forms or
interacting with the site if you do it too quickly. Even if these security features
aren’t in place, downloading lots of information from a website significantly faster
than a normal human might is a good way to get yourself noticed, and blocked.
Therefore, although multithreaded programming might be a great way to load pages
faster — allowing you to process data in one thread while repeatedly loading pages
in another — it’s a terrible policy for writing good scrapers. You should always try
to keep individual page loads and data requests to a minimum. If possible, try to
space them out by a few seconds, even if you have to add in an extra:
Although web scraping often involves rule breaking and boundary pushing in order
to get data, this is one rule you don’t want to break. Not only does consuming
inordinate server resources put you in a legally vulnerable position, but you might
cripple or take down smaller sites. Bringing down websites is not an ethically
ambiguous situation: it’s just wrong. So watch your speed!
Common Form Security Features
There are many litmus tests that have been used over the years, and continue to be
used, with varying degrees of success, to separate web scrapers from browserusing humans. Although it’s not a big deal if a bot downloads some articles and
blog posts that were available to the public anyway, it is a big problem if a bot
creates thousands of user accounts and starts spamming all of your site’s members.
Web forms, especially forms that deal with account creation and logins, pose a
significant threat to security and computational overhead if they’re vulnerable to
indiscriminate use by bots, so it’s in the best interest of many site owners (or at
least they think it is) to try to limit access to the site.
These anti-bot security measures centered on forms and logins can pose a
significant challenge to web scrapers.
Keep in mind that this is only a partial overview of some of the security measures
you might encounter when creating automated bots for these forms. Review
Chapter 11, on dealing with CAPTCHAs and image processing, as well as
Chapter 14, on dealing with headers and IP addresses, for more information on
dealing with well-protected forms.
Hidden Input Field Values
“Hidden” fields in HTML forms allow the value contained in the field to be
viewable by the browser but invisible to the user (unless they look at the site’s
source code). With the increase in use of cookies to store variables and pass them
around on websites, hidden fields fell out of a favor for a while before another
excellent purpose was discovered for them: preventing scrapers from submitting
Figure 12-1 shows an example of these hidden fields at work on a Facebook login
page. Although there are only three visible fields in the form (username, password,
and a submit button), the form conveys a great deal of information to the server
behind the scenes.
Figure 12-1. The Facebook login form has quite a few hidden fields
There are two main ways hidden fields are used to prevent web scraping: a field
can be populated with a randomly generated variable on the form page that the
server is expecting to be posted to the form processing page. If this value is not
present in the form, then the server can reasonably assume that the submission did
not originate organically from the form page, but was posted by a bot directly to the
processing page. The best way to get around this measure is to scrape the form page
first, collect the randomly generated variable and then post to the processing page
from there.
The second method is a “honey pot” of sorts. If a form contains a hidden field with
an innocuous name, such as “username” or “email address,” a poorly written bot
might fill out the field and attempt to submit it, regardless of whether it is hidden to
the user or not. Any hidden fields with actual values (or values that are different
from their defaults on the form submission page) should be disregarded, and the
user may even be blocked from the site.
In short: It is sometimes necessary to check the page that the form is on to see if you
missed anything that the server might be expecting. If you see several hidden fields,
often with large, randomly generated string variables, it is likely that the web
server will be checking for their existence on form submission. In addition, there
might be other checks to ensure that the form variables have been used only once,
are recently generated (this eliminates the possibility of simply storing them in a
script and using them over and over again over time), or both.
Avoiding Honeypots
Although CSS for the most part makes life extremely easy when it comes to
differentiating useful information from nonuseful information (e.g., by reading the
id and class tags), it can occasionally be problematic for web scrapers. If a field
on a web form is hidden from a user via CSS, it is reasonable to assume that the
average user visiting the site will not be able to fill it out because it doesn’t show
up in the browser. If the form is populated, there is likely a bot at work and the post
will be discarded.
This applies not only to forms but to links, images, files, and any other item on the
site that can be read by a bot but is hidden from the average user visiting the site
through a browser. A page visit to a “hidden” link on a site can easily trigger a
server-side script that will block the user’s IP address, log that user out of the site,
or take some other action to prevent further access. In fact, many business models
have been based on exactly this concept.
Take for example the page located at This page contains
two links, one hidden by CSS and another visible. In addition, it contains a form
with two hidden fields:
<title>A bot-proof form</title>
body {
.customHidden {
<h2>A bot-proof form</h2>
<a href=
"" style="display:none;">Go here!</a>
<a href="">Click me!</a>
<input type="hidden" name="phone" value="valueShouldNotBeModified"/><p/>
<input type="text" name="email" class="customHidden"
<input type="text" name="firstName"/><p/>
<input type="text" name="lastName"/><p/>
<input type="submit" value="Submit"/><p/>
These three elements are hidden from the user in three different ways:
The first link is hidden with a simple CSS display:none attribute
The phone field is a hidden input field
The email field is hidden by moving it 50,000 pixels to the right (presumably off
the screen of everyone’s monitors) and hiding the tell tale scroll bar
Fortunately, because Selenium actually renders the pages it visits, it is able to
distinguish between elements that are visually present on the page and those that
aren’t. Whether the element is present on the page can be determined by the
is_displayed() function.
For example, the following code retrieves the previously described page and looks
for hidden links and form input fields:
from selenium import webdriver
from selenium.webdriver.remote.webelement import WebElement
driver = webdriver.PhantomJS(executable_path='')
links = driver.find_elements_by_tag_name("a")
for link in links:
if not link.is_displayed():
print("The link "+link.get_attribute("href")+" is a trap")
fields = driver.find_elements_by_tag_name("input")
for field in fields:
if not field.is_displayed():
print("Do not change value of "+field.get_attribute("name"))
Selenium catches each hidden field, producing the following output:
The link is a trap
Do not change value of phone
Do not change value of email
Although you probably don’t want to visit any hidden links you find, you will want
to make sure that you submit any pre-populated hidden form values (or have
Selenium submit them for you) with the rest of the form. To sum up, it is dangerous
to simply ignore hidden fields, although you must be very careful when interacting
with them.
The Human Checklist
There’s a lot of information in this chapter, and indeed in this book, about how to
build a scraper that looks less like a scraper and more like a human. If you keep
getting blocked by websites and you don’t know why, here’s a checklist you can use
to remedy the problem:
First, if the page you are receiving from the web server is blank, missing
information, or is otherwise not what you expect (or have seen in your own
browser), it is likely caused by JavaScript being executed on the site to create
the page. Review Chapter 10.
If you are submitting a form or making a POST request to a website, check the
page to make sure that everything the website is expecting you to submit is being
submitted and in the correct format. Use a tool such as Chrome’s Network
inspector to view an actual POST command sent to the site to make sure you’ve
got everything.
If you are trying to log into a site and can’t make the login “stick,” or the website
is experiencing other strange “state” behavior, check your cookies. Make sure
that cookies are being persisted correctly between each page load and that your
cookies are sent to the site for every request.
If you are getting HTTP errors from the client, especially 403 Forbidden errors,
it might indicate that the website has identified your IP address as a bot and is
unwilling to accept any more requests. You will need to either wait until your IP
address is removed from the list, or obtain a new IP address (either move to a
Starbucks or see Chapter 14). To make sure you don’t get blocked again, try the
Make sure that your scrapers aren’t moving through the site too quickly. Fast
scraping is a bad practice that places a heavy burden on the web administrator’s
servers, can land you in legal trouble, and is the number-one cause of scrapers
getting blacklisted. Add delays to your scrapers and let them run overnight.
Remember: Being in a rush to write programs or gather data is a sign of bad
project management; plan ahead to avoid messes like this in the first place.
The obvious one: change your headers! Some sites will block anything that
advertises itself as a scraper. Copy your own browser’s headers if you’re
unsure about what some reasonable header values are.
Make sure you’re not clicking on or accessing anything that a human normally
would not be able to (refer back to “Avoiding Honeypots” for more
If you find yourself jumping through a lot of difficult hoops to gain access,
consider contacting the website administrator to let them know what you’re
doing. Try emailing [email protected]<domain name> or [email protected]<domain name>
for permission to use your scrapers. Admins are people, too!
Chapter 13. Testing Your Website with Scrapers
When working with web projects that have a large development stack, it’s often
only the “back” of the stack that ever gets tested regularly. Most programming
languages today (including Python) have some type of test framework, but the front
end of websites are often left out of these automated tests, although they might be
the only customer-facing part of the project.
Part of the problem is that websites are often a mishmash of many markup
languages and programming languages. You can write unit tests for sections of your
JavaScript, but it’s useless if the HTML it’s interacting with has changed in such
away that the JavaScript doesn’t have the intended action on the page, even if it’s
working correctly.
The problem of front-end website testing has often been left as an afterthought, or
delegated to lower-level programmers armed with, at most, a checklist and a bug
tracker. However, with just a little more up-front effort, we can replace this
checklist with a series of unit tests, and replace human eyes with a web scraper.
Imagine: test-driven development for web development. Daily tests to make sure
all parts of the web interface are functioning as expected. A suite of tests run every
time someone adds a new website feature, or changes the position of an element. In
this chapter, we’ll cover the basics of testing and how to test all sorts of websites,
from simple to complicated, with Python-based web scrapers.
An Introduction to Testing
If you’ve never written tests for your code before, there’s no better time to start than
now. Having a suite of tests that can be run to ensure that your code performs as
expected (at least, as far as you’ve written tests for) saves you time and worry and
makes releasing new updates easy.
What Are Unit Tests?
The words test and unit test are often used interchangeably. Often, when
programmers refer to “writing tests” what they really mean is “writing unit tests.”
On the other hand, when some programmers refer to writing unit tests, they’re really
writing some other kind of test.
Although definitions and practices tend to vary from company to company, a unit
test generally has the following characteristics:
Each unit test tests one aspect of the functionality of a component. For example,
it might ensure that the appropriate error message is thrown if a negative number
of dollars is withdrawn from a bank account.
Often, unit tests are grouped together in the same class, based on the component
they are testing. You might have the test for a negative dollar value being
withdrawn from a bank account, followed by a unit test for the behavior of an
overdrawn bank account.
Each unit test can be run completely independently, and any setup or teardown
required for the unit test must be handled by the unit test itself. Similarly, unit
tests must not interfere with the success or failure of other tests, and they must be
able to run successfully in any order.
Each unit test usually contains at least one assertion. For example, a unit test
might assert that the answer to 2+2 is 4. Occasionally, a unit test might contain
only a failure state. For example, it might fail if an exception is not thrown, but
pass by default if everything goes smoothly.
Unit tests are separated from the bulk of the code. Although they necessarily
need to import and use the code they are testing, they are generally kept in
separate classes and directories.
Although there are many other types of tests that can be written — integration tests
and validation tests, for example — we will primarily focus on unit testing in this
chapter. Not only have unit tests become extremely popular, with recent pushes
toward test-driven development, but their length and flexibility make them easy to
work with as examples, and Python has some built-in unit testing capabilities as
we’ll see in the next section.
Python unittest
Python’s unit testing module, unittest, comes packaged with all standard Python
installations. Just import and extend unittest.TestCase, and it will do the
Provide setUp and tearDown functions that run before and after each unit test
Provide several types of “assert” statements to allow tests to pass or fail
Run all functions that begin with test_ as unit tests, and ignore functions that
are not prepended as tests.
The following provides a very simple unit test for ensuring that 2+2 = 4, according
to Python:
import unittest
class TestAddition(unittest.TestCase):
def setUp(self):
print("Setting up the test")
def tearDown(self):
print("Tearing down the test")
def test_twoPlusTwo(self):
total = 2+2
self.assertEqual(4, total);
if __name__ == '__main__':
Although setUp and tearDown don’t provide any useful functionality here, they are
included for the purposes of illustration. Note that these functions are run before
and after each individual test, not before and after all the tests in the class.
Testing Wikipedia
Testing the front end of your website (excluding JavaScript, which we’ll cover
next) is as simple as combining the Python unittest library with a web scraper:
from urllib.request import urlopen
from bs4 import BeautifulSoup
import unittest
class TestWikipedia(unittest.TestCase):
bsObj = None
def setUpClass():
global bsObj
url = ""
bsObj = BeautifulSoup(urlopen(url))
def test_titleText(self):
global bsObj
pageTitle = bsObj.find("h1").get_text()
self.assertEqual("Monty Python", pageTitle);
def test_contentExists(self):
global bsObj
content = bsObj.find("div",{"id":"mw-content-text"})
if __name__ == '__main__':
There are two tests this time: the first tests whether the title of the page is the
expected “Monty Python,” and the second makes sure that the page has a content
Note that the content of the page is loaded only once, and that the global object
bsObj is shared between tests. This is accomplished by using the unittestspecified function setUpClass, which is run only once at the start of the class
(unlike setUp, which is run before every individual test). Using setUpClass
instead of setUp saves unnecessary page loads; we can grab the content once and
run multiple tests on it.
Although testing a single page at a time might not seem all that powerful or
interesting, if you recall from Chapter 3, it is relatively easy to build web crawlers
that can iteratively move through all pages of a website. What happens when we
combine a web crawler with a unit test that makes some assertion about each page?
There are many ways to run a test repeatedly, but we must be careful only to load
each page once for each set of tests we want to run on the page, and we must also
avoid holding large amounts of information in memory at once. The following setup
does just that:
class TestWikipedia(unittest.TestCase):
bsObj = None
url = None
def test_PageProperties(self):
global bsObj
global url
url = ""
#Test the first 100 pages we encounter
for i in range(1, 100):
bsObj = BeautifulSoup(urlopen(url))
titles = self.titleMatchesURL()
self.assertEquals(titles[0], titles[1])
url = self.getNextLink()
def titleMatchesURL(self):
global bsObj
global url
pageTitle = bsObj.find("h1").get_text()
urlTitle = url[(url.index("/wiki/")+6):]
urlTitle = urlTitle.replace("_", " ")
urlTitle = unquote(urlTitle)
return [pageTitle.lower(), urlTitle.lower()]
def contentExists(self):
global bsObj
content = bsObj.find("div",{"id":"mw-content-text"})
if content is not None:
return True
return False
def getNextLink(self):
#Returns random link on page, using technique from Chapter 5
if __name__ == '__main__':
There are a few things to notice. First, there is only one actual test in this class. The
other functions are technically only helper functions, even though they’re doing the
bulk of the computational work to determine whether a test passes. Because the test
function performs the assertion statements, the results of the test are passed back to
the test function where the assertions happen.
Also, while contentExists returns a boolean, titleMatchesURL returns the
values themselves back for evaluation. To see why we would want to pass values
back rather than just a boolean, compare the results of a boolean assertion:
FAIL: test_PageProperties (__main__.TestWikipedia)
---------------------------------------------------------------------Traceback (most recent call last):
File "", line 22, in test_PageProperties
AssertionError: False is not true
with the results of an assertEquals statement:
FAIL: test_PageProperties (__main__.TestWikipedia)
---------------------------------------------------------------------Traceback (most recent call last):
File "", line 23, in test_PageProperties
self.assertEquals(titles[0], titles[1])
AssertionError: 'lockheed u-2' != 'u-2 spy plane'
Which one is easier to debug?
In this case, the error is occurring because of a redirect, when the article redirects to an article titled “Lockheed U-2”
Testing with Selenium
As with Ajax scraping in Chapter 10, JavaScript presents particular challenges
when doing website testing. Fortunately, Selenium has an excellent framework in
place for handling particularly complicated websites; in fact, the library was
originally designed for website testing!
Although obviously written in the same language, the syntax of Python unittests and
Selenium unit tests have surprisingly little in common. Selenium does not require
that its unit tests be contained as functions within classes; its “assert” statements do
not require parentheses; and tests pass silently, only producing some kind of
message on a failure:
driver = webdriver.PhantomJS()
assert "Monty Python" in driver.title
When run, this test should produce no output.
In this way, Selenium tests can be written more casually than Python unittests, and
assert statements can even be integrated into regular code, where it is desirable for
code execution to terminate if some condition is not met.
Interacting with the Site
Recently, I wanted to contact a local small business through its website’s contact
form but found that the HTML form was broken; nothing happened when I clicked
the submit button. After a little investigation, I saw they were using a simple mailto
form that was designed to send them an email with the form’s contents. Fortunately,
I was able to use this information to send them an email, explain the problem with
their form, and hire them, despite the technical issue.
If I were to write a traditional scraper that used or tested this form, my scraper
would likely just copy the layout of the form and send an email directly —
bypassing the form altogether. How could I test the functionality of the form and
ensure that it was working perfectly through a browser?
Although previous chapters have discussed navigating links, submitting forms, and
other types of interaction-like activity, at its core everything we’ve done is
designed to bypass the browser interface, not use it. Selenium, on the other hand,
can literally enter text, click buttons, and do everything through the browser (in this
case, the headless PhantomJS browser), and detect things like broken forms, badly
coded JavaScript, HTML typos, and other issues that might stymie actual
Key to this sort of testing is the concept of Selenium elements. This object was
briefly encountered in Chapter 10, and is returned by calls like:
usernameField = driver.find_element_by_name('username')
Just as there are a number of actions you can take on various elements of a website
in your browser, there are many actions Selenium can perform on any given
element. Among these are:
myElement.send_keys_to_element("content to enter")
In addition to performing a one-time action on an element, strings of actions can be
combined into action chains, which can be stored and executed once or multiple
times in a program. Action chains are useful in that they can be a convenient way to
string long sets of multiple actions, but they are functionally identical to calling the
action explicitly on the element, as in the preceding examples.
To see this difference, take a look at the form page
at (which was previously used as an example in
Chapter 9). We can fill out the form and submit it in the following way:
selenium import webdriver
selenium.webdriver.remote.webelement import WebElement
selenium.webdriver.common.keys import Keys
selenium.webdriver import ActionChains
driver = webdriver.PhantomJS(executable_path='<Path to Phantom JS>')
firstnameField = driver.find_element_by_name("firstname")
lastnameField = driver.find_element_by_name("lastname")
submitButton = driver.find_element_by_id("submit")
### METHOD 1 ###
### METHOD 2 ###
actions = ActionChains(driver).click(firstnameField).send_keys("Ryan")
Method 1 calls send_keys on the two fields, then clicks the submit button, while
method 2 uses a single action chain to click on and enter text in each field, which
happens in a sequence after the perform method is called. This script will operate
in the same way if either the first method or the second method is used and will
print the line:
Hello there, Ryan Mitchell!
There is another variation in the two methods, in addition to the objects they use to
handle the commands: notice that the first method clicks the “submit” button, while
the second uses the “return” keystroke to submit the form while the text box is
submitted. Because there are many ways to think about the sequence of events that
complete the same action, there are many ways to complete the same action using
Drag and drop
Clicking on buttons and entering text is one thing, but where Selenium really shines
is in its ability to deal with relatively novel forms of web interaction. Selenium
allows for the manipulation of drag-and-drop interfaces with ease. Using its dragand-drop function requires you to specify a “source” element (the element to be
dragged) and either an offset to drag it across, or a target element to drag it to.
The demo page located at presents an example of this type of
from selenium import webdriver
from selenium.webdriver.remote.webelement import WebElement
from selenium.webdriver import ActionChains
driver = webdriver.PhantomJS(executable_path='<Path to Phantom JS>')
element = driver.find_element_by_id("draggable")
target = driver.find_element_by_id("div2")
actions = ActionChains(driver)
actions.drag_and_drop(element, target).perform()
Two messages are printed out from the message div on the demo page. The first
Prove you are not a bot, by dragging the square from the blue area to the red
Then, very quickly, after the task is completed, the content is printed out again,
which now reads:
You are definitely not a bot!
Of course, like the demo page suggests, dragging elements to prove you’re not a bot
is a common theme in many CAPTCHAs. Although bots have been able to drag
objects around for a long time (it’s just a matter of clicking, holding, and moving),
somehow the idea of using “drag this” as a verification of humanity just won’t die.
In addition, these draggable CAPTCHA libraries rarely use any difficult-for-bots
tasks, like “drag the picture of the kitten onto the picture of the cow” (which
requires you to identify the pictures as “a kitten” and “a cow,” while parsing
instructions); instead, they often involve number ordering or some other fairly
trivial task like the one in the preceding example.
Of course, their strength lies in the fact that there are so many different variations,
and they are so infrequently used — no one will likely bother making a bot that can
defeat all of them. At any rate, this example should be enough to illustrate why you
should never use this technique for large-scale websites.
Taking screenshots
In addition to the usual testing capabilities, Selenium has an interesting trick up its
sleeve that might make your testing (or impressing your boss) a little easier:
screenshots. Yes, photographic evidence can be created from unit tests run without
the need for actually pressing the PrtScn key:
driver = webdriver.PhantomJS()
This script navigates to, waits five seconds for any
JavaScript to execute, then stores a screenshot of the home page in the local tmp
folder (the folder must already exist for this to store correctly).
Unittest or Selenium?
The syntactical rigor and verboseness of a Python unittest might be desirable for
most large test suites, while the flexibility and power of a Selenium test might be
your only option for testing some website features. So which to use?
Here’s the secret: you don’t have to choose. Selenium can easily be used to obtain
some information about a website, while unittest can evaluate whether that
information meets the criteria for passing the test. There is no reason you can’t
import Selenium tools into a Python unittest, combining the best of both worlds.
For example, the following script creates a unit test for a website’s draggable
interface, asserting that it correctly says, “You are not a bot!” after one element has
been dragged to another:
from selenium import webdriver
from selenium.webdriver.remote.webelement import WebElement
from selenium.webdriver import ActionChains
import unittest
class TestAddition(unittest.TestCase):
driver = None
def setUp(self):
global driver
driver = webdriver.PhantomJS(executable_path='<Path to Phantom JS>')
url = ''
def tearDown(self):
print("Tearing down the test")
def test_drag(self):
global driver
element = driver.find_element_by_id("draggable")
target = driver.find_element_by_id("div2")
actions = ActionChains(driver)
actions.drag_and_drop(element, target).perform()
self.assertEqual("You are definitely
not a bot!", driver.find_element_by_id(
if __name__ == '__main__':
Virtually anything on a website can be tested with the combination of Python’s
unittest and Selenium. In fact, combined with some of the image-processing
libraries from Chapter 11, you can even take a screenshot of the website and test on
a pixel-by-pixel basis what it should contain!
Chapter 14. Scraping Remotely
That this chapter is the last in the book is somewhat appropriate. Up until now we
have been running all the Python applications from the command line, within the
confines of our home computers. Sure, you might have installed MySQL in an
attempt to replicate the environment of a real-life server. But it’s just not the same.
As the saying goes: “If you love something, set it free.”
In this chapter, I’ll cover several methods for running scripts from different
machines, or even just different IP addresses on your own machine. Although you
might be tempted to put this step off as something you don’t need right now, you
might be surprised at how easy it is to get started with the tools you already have
(such as a personal website on a paid hosting account), and how much easier your
life becomes once you stop trying to run Python scrapers from your laptop.
Why Use Remote Servers?
Although using a remote server might seem like an obvious step when launching a
web app intended for use by a wide audience, more often than not the tools we
build for our own purposes are just left running locally. People who decide to push
onto a remote platform usually base their decision on two primary motivations: the
need for greater power and flexibility, and the need to use an alternative IP address.
Avoiding IP Address Blocking
When building web scrapers, the rule of thumb is: everything can be faked. You can
send emails from addresses you don’t own, automate mouse movement data from a
command line, or even horrify web administrators by sending their website traffic
from Internet Explorer 5.0.
The one thing that cannot be faked is your IP address. Anyone can send you a letter
with the return address: “The President, 1600 Pennsylvania Avenue Northwest,
Washington, DC 20500.” However, if the letter is postmarked from Albuquerque,
N.M., you can be fairly certain you’re not corresponding with the President of the
United States.1
Most efforts to stop scrapers from accessing websites focus on detecting the
difference between humans and bots. Going so far as to block IP addresses is a
little like a farmer giving up spraying pesticides in favor of just torching the field.
It’s a last-ditch but very effective method of simply discarding packets sent from
troublesome IP addresses. However, there are problems with this solution:
IP address access lists are painful to maintain. Although large websites most
often have their own programs automating some of the routine management of
these lists (bots blocking bots!), someone has to at least occasionally check
them, or at least monitor their growth for problems.
Each address adds a tiny amount of processing time to receive packets, as the
server must check received packets against the list to decide whether to approve
them. Many addresses multiplied by many packets can add up quickly. To save
on processing time and complexity, admins often group these IP addresses into
blocks and make rules such as “all 256 addresses in this range are blocked” if
there are a few tightly clustered offenders. Which leads us to the third point.
IP address blocking can lead to unintended consequences. For example, while I
was an undergrad at Olin College of Engineering, one student wrote some
software that attempted to rig votes for popular content on (this
was before Reddit was in vogue). A single IP address blocked led to an entire
dormitory being unable to access the site. The student simply moved his
software to another server; in the meantime, Digg lost page visits from many
regular users in its prime target demographic.
Despite its drawbacks, IP address blocking remains an extremely common method
for server administrators to stop suspected web scrapers from accessing servers.
Portability and Extensibility
Some tasks are simply too large for a home computer and Internet connection.
Although you don’t want to put a large load on any single website, you might be
collecting data across a wide range of sites, and require a lot more bandwidth and
storage than your current setup can provide.
Moreover, by offloading computationally intensive processing, you can free up your
home machine’s cycles for more important tasks (World of Warcraft, anyone?). You
don’t have to worry about maintaining power and an Internet connection (launch
your app at a Starbucks, pack up your laptop, and leave, knowing that everything’s
still running safely), and you can access your collected data anywhere there’s an
Internet connection.
If you have an application that requires so much computing power that a single
Amazon extra-large computing instance won’t satisfy you, you can also look into
distributed computing. This allows multiple machines to work in parallel to
accomplish your goals. As a simple example, you might have one machine crawl
one set of sites and another crawl a second set of sites, and have both of them store
collected data in the same database.
Of course, as noted in previous chapters, many can replicate what Google search
does, but very few can replicate the scale at which Google search does it.
Distributed computing is a very large field of computer science that is outside the
scope of this book. However, learning how to launch your application onto a
remote server is a necessary first step, and you might be surprised at what
computers are capable of these days.
The Onion Router network, better known by the acronym Tor, is a network of
volunteer servers set up to route and reroute traffic through many layers (hence the
onion reference) of different servers in order to obscure its origin. Data is
encrypted before it enters the network so that if any particular server is
eavesdropped on the nature of the communication cannot be revealed. In addition,
although the inbound and outbound communications of any particular server can be
compromised, one would need to know the details of inbound and outbound
communication for all the servers along the path of communication in order to
decipher the true start and endpoints of a communication — a near-impossible feat.
Tor is commonly used by human rights workers and political whistleblowers to
communicate with journalists, and receives much of its funding from the U.S.
government. Of course, it is also commonly used for illegal activities as well, and
so remains a constant target for government surveillance (although to date the
surveillance has had only mixed success).
Limitations of Tor Anonymity
Although the reason we are using Tor in this book is to change our IP address,
not achieve complete anonymity per se, it is worth taking a moment to address
some of the strengths and limitations of Tor’s ability to anonymize traffic.
Although you can assume when using Tor that the IP address you are coming
from, according to a web server, is not an IP address that can be traced back to
you, any information you share with that web server might expose you. For
instance, if you log into your own Gmail account and then make incriminating
Google searches, those searches can now be tied back to your identity.
Beyond the obvious, however, even the act of logging into Tor might be
hazardous to your anonymity. In December 2013, a Harvard undergraduate
student, in an attempt to get out of final exams, emailed a bomb threat to the
school through the Tor network, using an anonymous email account. When the
Harvard IT team looked at their logs, they found traffic going out to the Tor
network from only a single machine, registered to a known student, during the
time that the bomb threat was sent. Although they could not identify the
eventual destination of this traffic (only that it was sent across Tor), the fact that
the times matched up and there was only a single machine logged in at the time
was damning enough to prosecute the student.
Logging into Tor is not an automatic invisibility cloak, nor does it give you free
reign to do as you please on the Internet. Although it is a useful tool, be sure to
use it with caution, intelligence, and, of course, morality.
Having Tor installed and running is a requirement for using Python with Tor, as we
will see in the next section. Fortunately, the Tor service is extremely easy to install
and start running with. Just go to the Tor downloads page (
and download, install, open, and connect! Keep in mind that your Internet speed
might appear to be slower while using Tor. Be patient — it might be going around
the world several times!
PySocks is a remarkably simple Python module that is capable of routing traffic
through proxy servers and that works fantastically in conjunction with Tor. You can
download it from its website or use any number of third-party module managers to
install it.
Although not much in the way of documentation exists for this module, using it is
extremely straightforward. The Tor service must be running on port 9150 (the
default port) while running this code:
import socks
import socket
from urllib.request import urlopen
socks.set_default_proxy(socks.SOCKS5, "localhost", 9150)
socket.socket = socks.socksocket
The website displays only the IP address for the client
connecting to the server and can be useful for testing purposes. When this script is
run, it should display an IP address that is not your own.
If you want to use Selenium and PhantomJS with Tor, you don’t need PySocks at all
— just make sure that Tor is currently running and add the optional service_args
parameters, specifying that Selenium should connect through port 9150:
from selenium import webdriver
service_args = [ '--proxy=localhost:9150', '--proxy-type=socks5', ]
driver = webdriver.PhantomJS(executable_path='<path to PhantomJS>', \
Again, this should print out an IP address that is not your own but the one that your
running Tor client is currently using.
Remote Hosting
Although complete anonymity is lost once you pull out your credit card, hosting
your web scrapers remotely dramatically improves their speed. This is both
because you’re able to purchase time on much larger machines than you likely own,
but also because the connection no longer has to bounce through layers of a Tor
network in order to reach its destination.
Running from a Website Hosting Account
If you have a personal or business website, you might already likely have the means
to run your web scrapers from an external server. Even with relatively lockeddown web servers, where you have no access to the command line, it is possible to
trigger scripts to start and stop through a web interface.
If your website is hosted on a Linux server, it likely already runs Python. If you’re
hosting on a Windows server, you might be out of luck; you’ll need to check
specifically to see if Python is installed, or if the server administrator is willing to
install it.
Most small web hosting providers come with software called cPanel, used to
provide basic administration services and information about your website and
related services. If you have access to cPanel, you can make sure that Python is set
up to run on your server by going to “Apache Handlers” and adding a new handler
(if it is not already present):
Handler: cgi-script
Extension(s): .py
This tells your server that all Python scripts should be executed as a CGI-script.
CGI, which stands for Common Gateway Interface, is simply any program that can
be run on a server and dynamically generate content that is displayed on a website.
By explicitly defining Python scripts as CGI scripts, you’re giving the server
permission to execute them, rather than just display them in a browser or send the
user a download.
Simply write your Python script, upload it to the server, and set the file permissions
to 755 to allow it to be executed. To execute the script, simply navigate to the place
you uploaded it to through your browser (or even better, write a scraper to do it for
you). If you’re worried about the general public accessing and executing the script,
you have two options:
Store the script at an obscure or hidden URL and make sure to never link to the
script from any other accessible URL to avoid search engines indexing it.
Protect the script with a password, or require that a password or secret token be
sent to it before it can execute.
Of course, running a Python script from a service that is specifically designed to
display websites is a bit of a hack. For instance, you’ll probably notice that your
web scraper-cum-website is a little slow to load. In fact, the page doesn’t actually
load (complete with the output of all “print” statements you might have written in)
until the entire scrape is complete. This might take minutes, hours, or never
complete at all, depending on how it is written. Although it certainly gets the job
done, you might want more real-time output. For that you’ll need an actual server.
Running from the Cloud
Back in the olden days of computing, programmers paid for or reserved time on
computers in order to execute their code. With the advent of personal computers,
this became unnecessary — you simply write and execute code on your own
computer. Now, the ambitions of the applications have outpaced the development of
the microprocessor to such a degree that programmers are once again moving to
pay-per-hour computing instances.
This time around, however, users aren’t paying for time on a single, physical
machine but on its equivalent computing power, often spread among many
machines. The nebulous structure of this system allows computing power to be
priced according to times of peak demand. For instance, Amazon allows for
bidding on “spot instances” when low costs are more important than immediacy.
Compute instances are also more specialized, and can be selected based on the
needs of your application, with options like “high memory,” “fast computing,”
and “large storage.” Although web scrapers don’t usually use much in the way of
memory, you may want to consider large storage or fast computing in lieu of a more
general-purpose instance for your scraping application. If you’re doing large
amounts of natural language processing, OCR work, or path finding (such as with
the “Six Degrees of Wikipedia” problem), a fast computing instance might work
well. If you’re scraping large amounts of data, storing files, or large-scale
analytics, you might want to go for an instance with storage optimization.
Although the sky is the limit as far as spending goes, at the time of this writing,
instances start at just 1.3 cents an hour (for an Amazon EC2 micro instance), and
Google’s cheapest instance is 4.5 cents an hour, with a minimum of just 10 minutes.
Thanks to the economies of scale, buying a small compute instance with a large
company is about the same as buying your own physical, dedicated, machine —
except that now, you don’t need to hire an IT guy to keep it running.
Of course, step-by-step instructions for setting up and running cloud computing
instances are somewhat outside of the scope of this book, but you will likely find
that step-by-step instructions are not needed. With both Amazon and Google (not to
mention the countless smaller companies in the industry) vying for cloud computing
dollars, they’ve made setting up new instances as easy as following a simple
prompt, thinking of an app name, and providing a credit card number. As of this
writing, both Amazon and Google also offer hundreds of dollars worth of free
computing hours to further tempt new clients.
Once you have an instance set up, you should be the proud new owner of an IP
address, username, and public/private keys that can be used to connect to your
instance through SSH. From there, everything should be literally the same as
working with a server that you physically own — except, of course, you no longer
have to worry about hardware maintenance or running your own plethora of
advanced monitoring tools.
Additional Resources
Many years ago, running “in the cloud” was mostly the domain of those who felt
like slogging through the documentation and already had some server administration
experience. However, today, the tools have improved dramatically, due to
increased popularity and competition among cloud computing providers.
Still, for building large-scale or more-complex scrapers and crawlers, you might
want a little more guidance on creating a platform for collecting and storing data.
Google Compute Engine by Marc Cohen, Kathryn Hurley, and Paul Newson is a
straightforward resource on using Google Cloud Computing with both Python and
JavaScript. Not only does it cover Google’s user interface, but also the commandline and scripting tools that you can use to give your application greater flexibility.
If you prefer to work with Amazon, Mitch Garnaat’s Python and AWS Cookbook is
a brief but extremely useful guide that will get you started with Amazon Web
Services and show you how to get a scalable application up and running.
Moving Forward
The Web is constantly changing. The technologies that bring us images, video, text,
and other data files are constantly being updated and reinvented. To keep pace, the
collection of technologies used to scrape data from the Internet must also change.
Future versions of this text omit JavaScript entirely as an obsolete and rarely used
technology and instead focus on HTML8 hologram parsing. However, what won’t
change is the mindset and general approach needed to successfully scrape any
website. When encountering any web scraping project, now and into the distant
future, you should always ask yourself:
What is the question I want answered, or the problem I want solved?
What data will help me achieve this, and where is it?
How is the website displaying this data? Can I identify exactly which part of the
website’s code contains this information?
How can I isolate the data and retrieve it?
What processing or analysis needs to be done to make it more useful?
How can I make this process better, faster, and more robust?
In addition, you will need to understand not just how to use the tools presented in
this book in isolation, but how they can work together to solve a larger problem.
Sometimes the data is easily available and well formatted, allowing a simple
scraper to do the trick. Other times you have to put some thought into it.
In Chapter 10, for example, I combined the Selenium library to identify Ajaxloaded images on Amazon, and Tesseract to use optical character recognition to
read them. In the “Six Degrees of Wikipedia” problem, I used regular expressions
to write a crawler that stored link information in a database, and then used a graphsolving algorithm in order to answer the question, “What is the shortest path of
links between Kevin Bacon and Eric Idle”?
There is rarely an unsolvable problem when it comes to automated data collection
on the Internet. Just remember: the Internet is one giant API with a somewhat poor
user interface.
Technically, IP addresses can be spoofed in outgoing packets, which is a
technique used in distributed denial-of-service attacks, where the attackers don’t
care about receiving return packets (which, if sent, will be sent to the wrong
address). But web scraping is by definition an activity in which a response from the
web server is required, so we think of IP addresses are one thing that can’t be
Appendix A. Python at a Glance
According to the Association for Computing Machinery, Python is the top teaching
language in the United States, with it being more commonly taught in introductory
programming classes than BASIC, Java, and even C. This section will cover the
basics of installing, using, and running Python 3.x and Python programs.
Installation and “Hello, World!”
If you have OS X or almost any flavor of Linux, you likely already have Python
installed. If in doubt, type the following into your command line:
$python --version
We will be using Python 3.x throughout this book. If you see that you have Python
2.x installed, you upgrade on Linux simply by calling apt-get:
$sudo apt-get install python3
Note this might involve needing to execute your Python code by typing $python3 into the command line, rather than $python
The OS X operating system, as of this writing, has Python 2.7 installed by default.
To install Python 3.x, download an installer from the Python website. Again, if you
are running both Python 2.x and Python 3.x simultaneously, you might need to
explicitly call Python 3.x using $python3.
For Windows users without Python installed, there are precompiled installers
available for download. Simply download, open, and install. You might also need
to set the Windows system path to the directory in which you install Python in order
to tell the operating system where it’s at, but the website and installers have fairly
straightforward instructions for doing so.
For more complete information about installing and upgrading Python on all
platforms, visit the Python downloads page.
Unlike Java, Python can be used as a scripting language without creating any new
classes or functions. Opening up a new text file, writing:
print("Hello, Internet!")
and saving it as is perfectly legitimate. If you want to get a little more
elaborate, you can create a new function, and use it to do the same thing:
def hello():
print("Hello, Internet!")
There are a few things to notice here:
Python does not use semicolons to indicate the end of a line, nor does it use braces
to indicate the beginning and ending of a loop or function. Instead, it relies on line
breaks and tabs to control execution. Opening a line with def, followed by a
function name, arguments (in this case, we have a (), indicating that there are no
arguments), and a colon indicates that the body of a function will follow. Each line
of a function following a function declaration must be indented, and outdenting
indicates the end of the function body.
If the idea of using whitespace in this way seems a little pedantic (or insane) at
first, keep at it. In my experience, writing Python actually improves the readability
of my code writing in other languages. Just remember to add semicolons when you
switch back over to something like Java or C!
Python is a weakly typed language, meaning that variable types (string, integer,
object, etc.) do not need to be explicitly declared when the variables are
initialized. As with other weakly typed languages, this can occasionally cause
debugging issues, but it does make variable declaration a breeze:
greeting = "Hello, Internet!"
And that’s it! You’re a Pythoneer!
Obviously, this isn’t quite the truth, but one of the great things about Python is that it
is such a simple language, programmers of other languages can read and interpret it
without much prior exposure. This aspect of the language can probably be best
summed up in Python’s most famous “Easter Egg”:
import this
The output of which is:
The Zen of Python, by Tim Peters
Beautiful is better than ugly.
Explicit is better than implicit.
Simple is better than complex.
Complex is better than complicated.
Flat is better than nested.
Sparse is better than dense.
Readability counts.
Special cases aren’t special enough to break the rules.
Although practicality beats purity.
Errors should never pass silently.
Unless explicitly silenced.
In the face of ambiguity, refuse the temptation to guess.
There should be one-- and preferably only one --obvious way to do it.
Although that way may not be obvious at first unless you’re Dutch.1
Now is better than never.
Although never is often better than *right* now.
If the implementation is hard to explain, it’s a bad idea.
If the implementation is easy to explain, it may be a good idea.
Namespaces are one honking great idea -- let’s do more of those!
This line might be a reference to the Dutch computer scientist Edsger Dijkstra,
who said in a 1978 talk: “I thought that it was a firm principle of language
design...that in all respects equivalent programs should have few possibilities for
different representations... Otherwise completely different styles of programming
arise unnecessarily, thereby hampering maintainability, readability and what have
you” (
Or it may simply be due to the fact that the original creator of Python, Guido van
Rossum, is Dutch. No one seems to be entirely sure on this subject, however.
Appendix B. The Internet at a Glance
As the types of transactions the Internet is required to handle become increasingly
complex, the terms and technologies used to describe these transactions also
increases in complexity. Far removed from its roots as a way to exchange research
messages, the Internet must now handle large file uploads, streaming video, secure
banking transactions, credit card purchases, and the transmission of sensitive
corporate documents.
Despite these extra layers of complexity, however, the Internet remains, at its core,
a series of messages. Some messages contain requests for information, some
contain messages intended for a distant recipient, some contain file information or
instructions for a particular application on the machine it’s being sent to. These
requests are sent from a client (desktop or mobile device) machine to a server, and
vice versa. They can also be sent between servers themselves, perhaps in order to
gather more information requested by a client.
Figure B-1 depicts a few common types of Internet transactions: a request for the
location of a server at a particular domain name, the request of a web page and its
associated image file across two servers, and uploading an image file.
Figure B-1. A few types of common Internet transactions between clients and servers
There are many different protocols, or languages, that govern these communications
between clients and servers. You might receive your mail via the SMTP protocol
while making a phone call over VOIP and uploading files via FTP. All of these
protocols define different fields for headers, data encodings, sending and receiving
addresses or names, and other types of things. The protocol used for requesting,
sending, and receiving websites is the Hypertext Transfer Protocol (HTTP).
For the vast majority of the scrapers in this book (and, likely, the majority of the
scrapers you will write), HTTP is used to communicate with remote web servers.
For this reason, it’s important to examine the protocol in a little more detail.
An HTTP message contains two main parts: header fields and a data field. Each
header field contains a title and a value. The possible titles for these fields are predefined by the HTTP standard. For instance, you might have a header field like:
Content-Type: application/json
indicating that the data in the HTTP packet will use the JSON format. There are
60+ possible header fields that might be found in an HTTP packet, but we will only
use a small set of them in this book. The following table shows a few of the HTTP
header fields you should be familiar with:
Name Description Example
User- A string
Mozilla/5.0 (X11; Ubuntu; Linux x86_64; rv:28.0)
Agent indicating
Gecko/20100101 Firefox/28.0
system you
are making
the request
Cookie Variables
used by the 20549163.147923691.1398729710.1398729710.1398858679.2”
to hold
session data
and other
Status A code
“200” (Okay), “404” (Not Found)
the success
or failure of
the page
Once a packet, carried by HTTP, reaches your browser, the contents of the packet
must be interpreted as a website. The structure of websites are governed by HTML,
or HyperText Markup Language. Although HTML is often referred to as a
programming language, HTML is a markup language. It defines the structure of a
document using tags in order to demarcate items such as title, content, sidebar,
footer, and more.
All HTML pages (at least all well-formatted ones) are surrounded by opening and
closing <html></html> tags, with <head> and <body> tags in between. Other tags
populate these page headers and page bodies to form the structure and content of the
<title>An Example Page</title>
<h1>An Example Page</h1>
<div class="body">
Some example content is here
In this example, the page title (this is the text that is seen in a tab in your browser)
is “An Example Page”, with that same title echoed in the header <h1> tag.
Immediately below that is a div (“divider”) tag of the class “body,” containing
what might be a main article or longer piece of text.
Noting when page content is copied exactly in different areas of the page can be useful
in web scraping — part of the page might be more convenient or more reliable to
scrape than another part of the page.
In “An Example Page,” of course, the title is copied fairly obviously in two separate
areas of the page, but more subtle cases are common on the Web. You might be
scraping the names of employees off a corporate website, for example, and notice that
the names are inconsistently formatted:
<span id="mary_smith">Dr. Mary Smith, CEO</span>
<span id="john_jones">President of Finance Mr. John E. Jones</span>
<span id="stacy_roberts">Stacy Roberts III, Marketing</span>
Separating the content of these tags into “firstname, lastname” word pairs would be
difficult. How do you deal with middle names, titles, positions, and other garbage in
this mess of formatting? Fortunately, it appears that the id tags are well-formatted, and
separating them out in Python becomes a trivial matter of splitting the string on the
underscore character.
CSS, or Cascading Style Sheets, go hand in hand with HTML in order to define the
look of a website. CSS defines things like the color, position, size, and background
of various objects on a website.
Using the HTML in the previous example, this might look something like:
border:2px solid;
This creates a medium-size red title, with a border around the content text on the
Unfortunately, going into greater detail about HTTP, HTML, or CSS is outside the
scope of this book. However, if you’re not familiar with the subjects, I recommend
you take a look at W3Schools to look up unfamiliar terms or lines of HTML/CSS as
you read through the book. Using your browser’s “view source” feature is a great
help to start getting familiar with the syntax as well.
Appendix C. The Legalities and Ethics of Web
In 2010, software engineer Pete Warden built a web crawler to gather data from
Facebook. He collected data from approximately 200 million Facebook users —
names, location information, friends, and interests. Of course, Facebook noticed
and sent him cease-and-desist letters, which he obeyed. When asked why he
complied with the cease and desist, he said: “Big data? Cheap. Lawyers? Not so
In this chapter, we’ll look at U.S. laws (and some international ones) that are
relevant to web scraping, and learn how to analyze the legality and ethics of a given
web scraping situation.
Before you read this section, consider the obvious: I am a software engineer, not a
lawyer. Do not interpret anything you read here or in any other chapter of the book
as professional legal advice or act on it accordingly. Although I believe I’m able to
discuss the legalities and ethics of web scraping knowledgeably, you should consult
a lawyer (not a software engineer) before undertaking any legally ambiguous web
scraping projects.
Trademarks, Copyrights, Patents, Oh My!
Time for some Intellectual Property 101! There are three basic types of IP:
trademarks (indicated by a ™ or ® symbol), copyrights (the ubiquitous ©), and
patents (sometimes indicated by text describing that the invention is patent
protected, but often by nothing at all).
Patents are used to declare ownership over inventions only. You cannot patent
images, text, or any information itself. Although some patents, such as software
patents, are less tangible than what we think of as “inventions,” keep in mind that it
is the thing (or technique) that is patented — not the information contained in the
patent. Unless you are either building things from scraped blueprints, or someone
patents a method of web scraping, you are unlikely to inadvertently infringe on a
patent by scraping the Web.
Trademarks also are unlikely to be an issue, but still something that must be
considered. According to the U.S Patent and Trademark Office:
A trademark is a word, phrase, symbol, and/or design that identifies and
distinguishes the source of the goods of one party from those of others. A
service mark is a word, phrase, symbol, and/or design that identifies and
distinguishes the source of a service rather than goods. The term “trademark” is
often used to refer to both trademarks and service marks.
In addition to the traditional words/symbols branding that we think of when we
think of trademarks, other descriptive attributes can be trademarked. This includes,
for example, the shape of a container (think Coca-Cola bottles) or even a color
(most notably, the pink color of Owens Corning’s Pink Panther fiberglass
Unlike with patents, the ownership of a trademark depends heavily on the context in
which it is used. For example, if I wish to publish a blog post with an
accompanying picture of the Coca-Cola logo, I could do that (as long as I wasn’t
implying that my blog post was sponsored by, or published by, Coca-Cola). If I
wanted to manufacture a new soft drink with the same Coca-Cola logo displayed on
the packaging, that would clearly be a trademark infringement. Similarly, although I
could package my new soft drink in Pink Panther pink, I could not use that same
color to create a home insulation product.
Copyright Law
Both trademarks and patents have something in common in that they have to be
formally registered in order to be recognized. Contrary to popular belief, this is not
true with copyrighted material. What makes images, text, music, etc. copyrighted?
It’s not the “All Rights Reserved” warning at the bottom of the page, nor anything
special about “published” versus “unpublished” material. Every piece of material
you create is automatically subject to copyright law as soon as you bring it into
The Berne Convention for the Protection of Literary and Artistic Works, named
after Berne, Switzerland, where it was first adopted in 1886, is the international
standard for copyright. This convention says, in essence, that all member countries
must recognize the copyright protection of the works of citizens of other member
countries as if they were citizens of their own country. In practice, this means that,
as a U.S. citizen, you can be held accountable in the United States for violating the
copyright of material written by someone in, say, France (and vice versa).
Obviously, copyright is a concern for web scrapers. If I scrape content from
someone’s blog and publish it on my own blog, I could very well be opening myself
up to a lawsuit. Fortunately, there are several layers of protection that I have that
might make my blog-scraping project defensible, depending on how it functions.
First, copyright protection extends to creative works only. It does not cover
statistics or facts. Fortunately, much of what web scrapers are after are facts and
statistics. Although a web scraper that gathers poetry from around the Web and
displays that poetry on your own website might be violating copyright law, a web
scraper that gathers information on the frequency of poetry postings over time is
not. The poetry, in its raw form, is a creative work. The average word count of
poems published on a website, by month, is not a creative work.
Even content that is posted verbatim (as opposed to aggregated/calculated content
from raw scraped data) might not be violating copyright law if that data is prices,
names of company executives, or some other factual piece of information.
Even copyrighted content can be used directly, within reason, under the Digital
Millennium Copyright Act. The DMCA outlines some rules for the automated
handling of copyrighted material. The DMCA is long, with many specific rules
governing everything from ebooks to telephones. However, there are three main
sections that are of particular interest:
Under the “safe harbor” protection, if you scrape material from a source that you
are led to believe contains only copyright-free material, but some user has
submitted copyright material to, you are protected as long as you removed the
copyrighted material when notified.
You cannot circumvent security measures (such as password protection) in order
to gather content.
You can use content under the “fair use” rule, which takes into account the
percentage of the copyrighted work used and the purpose it is being used for.
In short, you should never directly publish copyrighted material without permission
from the original author or copyright holder. If you are storing copyrighted material
that you have free access to in your own nonpublic database for the purposes of
analysis, that is fine. If you are publishing that database to your website for viewing
or download, that is not fine. If you are analyzing that database and publishing
statistics about word counts, a list of authors by prolificacy, or some other metaanalysis of the data, that is fine. If you are accompanying that meta-analysis with a
few select quotes, or brief samples of data to make your point, that is likely also
fine, but you might want to examine the “fair use” clause of the DMCA to make
Trespass to Chattels
Trespass to chattels is fundamentally different from what we think of as
“trespassing laws” in that it applies not to real estate or land but to moveable
property (such as a server). It applies when your access to property is interfered
with in some way that does not allow you to access or use it.
In this era of cloud computing, it’s tempting not to think of web servers as real,
tangible, resources. However, not only do servers consist of expensive
components, but they need to be stored, monitored, cooled, and supplied with vast
amounts of electricity. By some estimates, 10% of global electricity usage is
consumed by computers (if a survey of your own electronics doesn’t convince you,
consider Google’s vast server farms, all of which need to be connected to large
power stations).
Although servers are expensive resources, they’re interesting from a legal
perspective in that webmasters generally want people to consume their resources
(i.e., access their websites); they just don’t want them to consume their resources
too much. Checking out a website via your browser is fine, launching a full-scale
DDOS against it obviously is not.
There are three criteria that need to be met for a web scraper to violate trespass to
Lack of consent
Because web servers are open to everyone, they are generally “giving
consent” to web scrapers as well. However, many websites’ Terms of Service
agreements specifically prohibit the use of scrapers. In addition, any ceaseand-desist notifications obviously revoke this consent.
Actual harm
Servers are costly. In addition to server costs, if your scrapers take a website
down, or limit its ability to serve other users, this can add to the “harm” you
If you’re writing the code, you know what it does!
You must meet all three of these criteria for trespass to chattels to apply. However,
if you are violating a Terms of Service agreement, but not causing actual harm,
don’t think that you’re immune from legal action. You might very well be violating
copyright law, the DMCA, the Computer Fraud and Abuse Act (more on that later),
or one of the other myriad of laws that apply to web scrapers.
Back in the olden days, web servers were far more powerful than personal computers.
In fact, part of the definition of “server” was “big computer.” Now, the tables have
turned somewhat. My personal computer, for instance, has a 3.5 GHz processor and 8
GB of RAM. An Amazon medium instance, in contrast (as of the writing of this book),
has 4 GB of RAM and about 3 GHz of processing capacity.
With a decent Internet connection and a dedicated machine, even a single personal
computer can place a heavy load on many websites, even crippling them or taking
them down completely. Unless there’s a medical emergency and the only cure is
aggregating all the data from Joe Schmo’s website in two seconds flat, there’s really no
reason to hammer on a site.
A watched bot never completes. Sometimes it’s better to leave crawlers running
overnight than in the middle of the afternoon or evening for a few reasons:
If you have about eight hours, even at the glacial pace of two seconds per page,
you can crawl over 14,000 pages. When time is less of an issue, you’re not
tempted to push the speed of your crawlers.
Assuming the target audience of the website is in your general location (adjust
accordingly for remote target audiences), the website’s traffic load is probably far
lower during the night, meaning that your crawling will not be compounding peak
traffic hour congestion.
You save time by sleeping, instead of constantly checking your logs for new
information. Think of how excited you’ll be to wake up in the morning to brandnew data!
Consider the following scenarios:
You have a web crawler that traverses Joe Schmo’s website, aggregating some or
all of its data
You have a web crawler that traverses hundreds of small websites, aggregating
some or all of their data.
You have a web crawler that traverses a very large site, such as Wikipedia.
In the first scenario, it’s best to leave the bot running slowly, and during the night.
In the second scenario, it’s best to crawl each website in a round-robin fashion, rather
than crawling them slowly, one at a time. Depending on how many websites you’re
crawling, this means that you can collect data as fast as your Internet connection and
machine can manage, yet the load is reasonable for each individual remote server. You
can accomplish this programatically, either using multiple threads (where each
individual thread crawls a single site and pauses its own execution), or using Python
lists to keep track of sites.
In the second scenario, the load your Internet connection and home machine can place
on a site like Wikipedia is unlikely to be noticed, or cared much about. However, if
you’re using a distributed network of machines, this is obviously a different matter.
Use caution, and ask a company representative whenever possible.
The Computer Fraud and Abuse Act
In the early 1980s, computers started moving out of academia and into the business
world. For the first time, viruses and worms were seen as more than an
inconvenience (or even a fun hobby) and as a serious criminal matter that could
cause actual monetary damages. In response, the Computer Fraud and Abuse Act
was created in 1986.
Although you might think that the act only applies to some stereotypical version of a
malicious hacker unleashing viruses, the act has strong implications for web
scrapers as well. Imagine a scraper that scans the web looking for login forms with
easy-to-guess passwords, or collects government secrets accidentally left in a
hidden but public location. All of these activities are illegal (and rightly so) under
the CFAA.
The act defines seven main criminal offenses, which can be summarized as
The knowing unauthorized access of computers owned by the U.S. government
and obtaining information from those computers.
The knowing unauthorized access of a computer, obtaining financial information.
The knowing unauthorized access of a computer owned by the U.S. government,
affecting the use of that computer by the government.
Knowingly accessing any protected computer with the attempt to defraud.
Knowingly accessing a computer without authorization and causing damage to
that computer.
Shares or traffics passwords or authorization information for computers used by
the U.S. government or computers that affect interstate or foreign commerce.
Attempts to extort money or “anything of value” by causing damage, or
threatening to cause damage, to any protected computer.
In short: stay away from protected computers, do not access computers (including
web servers) that you are not given access to, and especially, stay away from
government or financial computers.
robots.txt and Terms of Service
A website’s terms of service and robots.txt files are in interesting territory, legally
speaking. If a website is publicly accessible, the webmaster’s right to declare what
software can and cannot access it is debatable. Saying “it is fine if you use your
browser to view this site, but not if you use a program you wrote to view it” is
Most sites have a link to their Terms of Service in the footer on every page. The
TOS contains more than just the rules for web crawlers and automated access; it
often has information about what kind of information the website collects, what it
does with it, and usually a legal disclaimer that the services provided by the
website come without any express or implied warranty.
If you are interested in search engine optimization (SEO) or search engine
technology, you’ve probably heard of the robots.txt file. If you go to just about any
large website and look for its robots.txt file, you will find it in the root web folder:
The syntax for robots.txt files was developed in 1994 during the initial boom of
web search engine technology. It was about this time that search engines scouring
the entire Internet, such as AltaVista and DogPile, started competing in earnest with
simple lists of sites organized by subject, such as the one curated by Yahoo! This
growth of search across the Internet meant an explosion in not only the number of
web crawlers, but in the availability of information collected by those web
crawlers to the average citizen.
While we might take this sort of availability for granted today, some webmasters
were shocked when information they published deep in the file structure of their
website became available on the front page of search results in major search
engines. In response, the syntax for robots.txt files, called the Robots Exclusion
Standard, was developed.
Unlike the TOS, which often talks about web crawlers in broad terms and in very
human language, robots.txt can be parsed and used by automated programs
extremely easily. Although it might seem like the perfect system to solve the
problem of unwanted bots once and for all, keep in mind the following:
There is no official governing body for the syntax of robots.txt. It is a commonly
used and generally well-followed convention, but there is nothing to stop anyone
from creating their own version of a robots.txt file (apart from the fact that no
bot will recognize or obey it until it gets popular). That being said, it is a very
widely accepted convention, mostly because it is relatively straightforward, and
there is no incentive for companies to invent their own standard, or try to
improve on it.
There is no way to enforce a robots.txt file. It is merely a sign that says “Please
don’t go to these parts of the site.” There are many web scraping libraries
available that obey robots.txt (although this is often merely a default setting that
can be overridden). Beyond that, there are often more barriers to following a
robots.txt file (after all, you need to scrape, parse, and apply the contents of the
page to your code logic) than there are to just going ahead and scraping
whatever page you want to.
The Robot Exclusion Standard syntax is fairly straightforward. Like in Python (and
many other languages), comments begin with a # symbol, end with a newline, and
can be used anywhere in the file.
The first line of the file, apart from any comments, is started with User-agent:,
which specifies the user that the following rules apply to. This is followed by a set
of rules, either Allow: or Disallow:, depending on whether the bot is allowed on
that section of the site or not. An asterisk (*), indicates a wildcard, and can be used
to describe either a User-agent or a URL.
If a rule follows a rule that it seems to contradict, the last rule takes
precedence. For example:
#Welcome to my robots.txt file!
User-agent: *
Disallow: *
User-agent: Googlebot
Allow: *
Disallow: /private
In this case, all bots are disallowed from anywhere on the site, except for the
Googlebot, which is allowed anywhere except for the /private directory.
In Twitter’s robots.txt file, it has explicit instructions for the bots of Google,
Yahoo!, Yandex (a popular Russian search engine), Microsoft, and other bots or
search engines not covered by any of the preceding categories. The Google section
(which looks identical to the permissions allowed to all other categories of bots)
looks like:
#Google Search Engine Robot
User-agent: Googlebot
Allow: /?_escaped_fragment_
Allow: /?lang=
Allow: /hashtag/*?src=
Allow: /search?q=%23
Disallow: /search/realtime
Disallow: /search/users
Disallow: /search/*/grid
Disallow: /*?
Disallow: /*/followers
Disallow: /*/following
Notice that Twitter restricts access to the portions of its site that it has an API in
place for. Because Twitter has a well-regulated API (and one that it can make
money off of by licensing), it is in the company’s best interest to disallow any
“home-brewed APIs” that gather information by independently crawling its site.
Although a file telling your crawler where it can’t go might seem restrictive at first,
it can actually be a blessing in disguise for web crawler development. If you find a
robots.txt file that disallows crawling in a particular section of the site, the
webmaster is saying, essentially, that they are fine with crawlers in all other
sections of the site (after all, if they weren’t fine with it, they would have restricted
access when they were writing robots.txt in the first place).
For example, the section of Wikipedia’s robots.txt file that applies to general web
scrapers (as opposed to search engines) is wonderfully permissive. It even goes as
far as containing human-readable text to welcome bots (that’s us!) and blocks
access to only a few pages, such as the login page, search page, and “random
article” page:
# Friendly, low-speed bots are welcome viewing article pages, but not
# dynamically generated pages please.
# Inktomi's "Slurp" can read a minimum delay between hits; if your bot supports
# such a thing using the 'Crawl-delay' or another instruction, please let us
# know.
# There is a special exception for API mobileview to allow dynamic mobile web &
# app views to load section content.
# These views aren't HTTP-cached but use parser cache aggressively and don't
# expose special: pages etc.
User-agent: *
Allow: /w/api.php?action=mobileview&
Disallow: /w/
Disallow: /trap/
Disallow: /wiki/Especial:Search
Disallow: /wiki/Especial%3ASearch
Disallow: /wiki/Special:Collection
Disallow: /wiki/Spezial:Sammlung
Disallow: /wiki/Special:Random
Disallow: /wiki/Special%3ARandom
Disallow: /wiki/Special:Search
Disallow: /wiki/Special%3ASearch
Disallow: /wiki/Spesial:Search
Disallow: /wiki/Spesial%3ASearch
Disallow: /wiki/Spezial:Search
Disallow: /wiki/Spezial%3ASearch
Disallow: /wiki/Specjalna:Search
Disallow: /wiki/Specjalna%3ASearch
Disallow: /wiki/Speciaal:Search
Whether you choose to write web crawlers that obey robots.txt is up to you, but I
highly recommend it, particularly if you have crawlers that indiscriminately crawl
the Web.
Three Web Scrapers
Because web scraping is such a limitless field, there are a staggering number of
ways to land yourself in legal hot water. In this section, we’ll take a look at three
cases that touched on some form of law that generally applies to web scrapers, and
how it was used in that case.
eBay versus Bidder’s Edge and Trespass to Chattels
In 1997, the Beanie Baby market was booming, the tech sector was bubbling, and
online auction houses were the hot new thing on the Internet. A company called
Bidder’s Edge formed and created a new kind of meta-auction site. Rather than
force you to go from auction site to auction site, comparing prices, it would
aggregate data from all current auctions for a specific product (say, a hot new Furby
doll or a copy of Spice World) and point you to the site that had the lowest price.
Bidder’s Edge accomplished this with an army of web scrapers, constantly making
requests to the web servers of the various auction sites in order to get price and
product information. Of all the auction sites, eBay was the largest, and Bidder’s
Edge hit eBay’s servers about 100,000 times a day. Even by today’s standards, this
is a lot of traffic. According to eBay, this was 1.53% of its total Internet traffic at
the time, and it certainly wasn’t happy about it.
eBay sent Bidder’s Edge a cease-and-desist letter, coupled with an offer to license
its data. However, negotiations for this licensing failed and Bidder’s Edge
continued to crawl eBay’s site.
eBay tried blocking IP addresses used by Bidder’s Edge, blocking 169 different IP
addresses, although Bidder’s Edge was able to get around this by using proxy
servers (servers that forward requests on behalf of another machine, but using the
proxy server’s own IP address). As I’m sure you can imagine, this was a frustrating
and unsustainable solution for both parties — Bidder’s Edge was constantly trying
to find new proxy servers and buy new IP addresses while old ones were blocked,
while eBay was forced to maintain large firewall lists (and adding computationally
expensive IP address-comparing overhead to each packet check).
Finally, in December of 1999, eBay sued Bidder’s Edge under trespass to chattels.
Because eBay’s servers were real, tangible resources that it owned, and it didn’t
appreciate Bidder’s Edge’s abnormal use of them, trespass to chattels seemed like
the ideal law to use. In fact, in modern times, trespass to chattels goes hand in hand
with web-scraping lawsuits, and is most often thought of as an IT law.
The courts ruled that in order for eBay to win its case using trespass to chattels,
eBay had to show two things:
Bidder’s Edge did not have permission to use eBay’s resources
eBay suffered financial loss as a result of Bidder’s Edge’s actions
Given the record of eBay’s cease-and-desist letters, coupled with IT records
showing server usage and actual costs associated with the servers, this was
relatively easy for eBay to do. Of course, no large court battles end easily: counter309
suits were filed, many lawyers were paid, and the matter was eventually settled out
of court for an undisclosed sum in March 2001.
So does this mean that any unauthorized use of another person’s server is
automatically a violation of trespass to chattels? Not necessarily. Bidder’s Edge
was an extreme case; it was using so many of eBay’s resources that the company
had to buy additional servers, pay more for electricity, and perhaps hire additional
personnel (although 1.53% might not seem like a lot, in large companies, it can add
up to a significant amount)
In 2003, the California Supreme Court ruled on another case, Intel Corp versus
Hamidi, in which a former Intel employee (Hamidi) sent emails Intel didn’t like,
across Intel’s servers, to Intel employees. The court said:
Intel’s claim fails not because e-mail transmitted through the Internet enjoys
unique immunity, but because the trespass to chattels tort — unlike the causes of
action just mentioned — may not, in California, be proved without evidence of
an injury to the plaintiff’s personal property or legal interest therein.
Essentially, Intel had failed to prove that the costs of transmitting the six emails sent
by Hamidi to all employees (each one, interestingly enough, with an option to be
removed from Hamidi’s mailing list — at least he was polite!) contributed to any
financial injury for them. It did not deprive Intel of any property or use of their
United States v. Auernheimer and The Computer Fraud and
Abuse Act
If there is information readily accessible on the Internet to a human using a web
browser, it’s unlikely that accessing the same exact information in an automated
fashion would land you in hot water with the Feds. However, as easy as it can be
for a sufficiently curious person to find a small security leak, that small security
leak can very quickly become a much larger and much more dangerous one when
automated scrapers enter the picture.
In 2010, Andrew Auernheimer and Daniel Spitler noticed a nice feature of iPads:
when you visited AT&T’s website on them, AT&T would redirect you to a URL
containing your iPad’s unique ID number:<idNumber>&IMEI=
This page would contain a login form, with the email address of the user whose ID
number was in the URL. This allowed users to gain access to their accounts simply
by entering their password.
Although there were a very large number of potential iPad ID numbers, it was
possible, given enough web scrapers, to iterate through the possible numbers,
gathering email addresses along the way. By providing users with this convenient
login feature, AT&T, in essence, made its customer email addresses public to the
Auernheimer and Spitler created a scraper that collected 114,000 of these email
addresses, among them the private email addresses of celebrities, CEOs, and
government officials. Auernheimer (but not Spitler) then sent the list, and
information about how it was obtained, to Gawker Media, who published the story
(but not the list) under the headline: “Apple’s Worst Security Breach: 114,000 iPad
Owners Exposed.”
In June 2011, Auernheimer’s home was raided by the FBI in connection with the
email address collection, although they ended up arresting him on drug charges. In
November 2012, he was found guilty of identity fraud and conspiracy to access a
computer without authorization, and some time later sentenced to 41 months in
federal prison and ordered to pay $73,000 in restitution.
His case caught the attention of civil rights lawyer Orin Kerr, who joined his legal
team and appealed the case to the Third Circuit Court of Appeals. On April 11,
2014 (these legal processes can take quite a while), The Third Circuit agreed with
the appeal, saying:
Auernheimer’s conviction on Count 1 must be overturned because visiting a
publicly available website is not unauthorized access under the Computer Fraud
and Abuse Act, 18 U.S.C. § 1030(a)(2)(C). AT&T chose not to employ
passwords or any other protective measures to control access to the e-mail
addresses of its customers. It is irrelevant that AT&T subjectively wished that
outsiders would not stumble across the data or that Auernheimer hyperbolically
characterized the access as a “theft.” The company configured its servers to
make the information available to everyone and thereby authorized the general
public to view the information. Accessing the e-mail addresses through AT&T’s
public website was authorized under the CFAA and therefore was not a crime.
Thus, sanity prevailed in the legal system, Auernheimer was released from prison
that same day, and everyone lived happily ever after.
Although it was ultimately decided that Auernheimer did not violate the Computer
Fraud and Abuse Act, he had his house raided by the FBI, spent many thousands of
dollars in legal fees, and spent three years in and out of courtrooms and prisons. As
web scrapers, what lessons can we take away from this to avoid similar situations?
Scraping any sort of sensitive information, whether it’s personal data (in this case,
email addresses), trade secrets, or government secrets, is probably not something
you want to do without having a lawyer on speed dial. Even if it’s publicly
available, think: “Would the average computer user be able to easily access this
information if they wanted to see it?” “Is this something the company wants users to
I have on many occasions called companies to report security vulnerabilities in
their websites and web applications. This line works wonders: “Hi, I’m a security
professional who discovered a potential security vulnerability on your website.
Could you direct me to someone so that I can report it, and get the issue resolved?”
In addition to the immediate satisfaction of recognition for your (white hat) hacking
genius, you might be able to get free subscriptions, cash rewards, and other goodies
out of it!
In addition, Auernheimer’s release of the information to Gawker Media (before
notifying AT&T) and his showboating around the exploit of the vulnerability also
made him an especially attractive target for AT&T’s lawyers.
If you find security vulnerabilities in a site, the best thing to do is to alert the
owners of the site, not the media. You might be tempted to write up a blog post and
announce it to the world, especially if a fix to the problem is not put in place
immediately. However, you need to remember that it is the company’s
responsibility, not yours. The best thing you can do is take your web scrapers (and,
if applicable, your business) away from the site!
Field v. Google: Copyright and robots.txt
Blake Field, an attorney, filed a lawsuit against Google on the basis that its sitecaching feature violated copyright law by displaying a copy of his book after he
had removed it from his website. Copyright law allows the creator of an original
creative work to have control over the distribution of that work. Field’s argument
was that Google’s caching (after he had removed it from his website) removed his
ability to control its distribution.
The Google Web Cache
When Google web scrapers (also known as “Google bots”) crawl websites, they
make a copy of the site and host it on the Internet. Anyone can access this
cache, using the URL format:
If a website you are searching for, or scraping, is unavailable, you might want to
check there to see if a usable copy exists!
Knowing about Google’s caching feature and not taking action did not help Field’s
case. After all, he could have prevented Google’s bots from caching his website
simply by adding the robots.txt file, with simple directives about which pages
should and should not be scraped.
More importantly, the court found that the DMCS Safe Harbor provision allowed
Google to legally cache and display sites such as Field’s: “[a] service provider
shall not be liable for monetary relief... for infringement of copyright by reason of
the intermediate and temporary storage of material on a system or network
controlled or operated by or for the service provider.”
" (quotation marks), find() and findAll() with BeautifulSoup
$ (dollar sign), Regular Expressions and BeautifulSoup
() (parentheses), Regular Expressions
* (asterisk), Regular Expressions, Executing JavaScript in Python with
+ (plus sign), Regular Expressions
- (hyphen), Data Normalization
. (period), Regular Expressions
403 Forbidden error, The Human Checklist
404 Page Not Found error, Connecting Reliably
500 Internal Server Error, Connecting Reliably
; (semicolon), Installation and “Hello, World!”
?! (does not contain), Regular Expressions and BeautifulSoup
[] (square brackets), Regular Expressions
\ (forward slash), Regular Expressions and BeautifulSoup
^ (caret), Regular Expressions and BeautifulSoup
_ (underscore), find() and findAll() with BeautifulSoup
| (pipe), Regular Expressions
a tag, Accessing Attributes, Executing JavaScript in Python with Selenium
Accept header, Adjust Your Headers
Accept-Encoding header, Adjust Your Headers
Accept-Language header, Adjust Your Headers
action chains, Interacting with the Site
ActionScript , Scraping JavaScript
add_cookie function, Handling Cookies
Ajax (Asynchronous JavaScript and XML), Ajax and Dynamic HTML-Ajax
and Dynamic HTML
API keys
Echo Nest example, Authentication, A Few Examples
Google example, Getting Started-A Few Examples
Twitter example, Getting Started
about, Using APIs-How APIs Work, More About APIs
authentication and, Authentication
common conventions, Common Conventions-Authentication
Echo Nest example, Authentication, Echo Nest-A Few Examples
Google examples, How APIs Work, Google APIs-A Few Examples
HTTP methods and, Methods
parsing JSON, Parsing JSON
responses , Responses
Twitter example, Twitter-A Few Examples
Wikipedia example, Bringing It All Back Home-Bringing It All Back Home
ASCII character encoding, A brief overview of encoding types-Encodings in
assertions (unit tests), What Are Unit Tests?
asterisk (*), Regular Expressions, Executing JavaScript in Python with
Asynchronous JavaScript and XML (Ajax), Ajax and Dynamic HTML-Ajax
and Dynamic HTML
AttributeError exception, Connecting Reliably
accessing, Another Serving of BeautifulSoup, Accessing Attributes
finding tags based on, Another Serving of BeautifulSoup-Other
BeautifulSoup Objects
Auernheimer, Andrew, United States v. Auernheimer and The Computer Fraud
and Abuse Act
auth module, HTTP Basic Access Authentication
about, Authentication
handling logins, Handling Logins and Cookies-HTTP Basic Access
HTTP basic access, HTTP Basic Access Authentication
Twitter example, A Few Examples
backing up data, Training Tesseract
BeautifulSoup library
about, An Introduction to BeautifulSoup
children() function, Dealing with children and other descendants
descendants() function, Dealing with children and other descendants
find() function, find() and findAll() with BeautifulSoup-Other BeautifulSoup
findAll() function, Another Serving of BeautifulSoup-Other BeautifulSoup
Objects, Lambda Expressions
get_text() function, Another Serving of BeautifulSoup
installing, Installing BeautifulSoup
next_siblings() function, Dealing with siblings
previous_siblings() function, Dealing with siblings
regular expressions and, Regular Expressions and BeautifulSoup
running, Running BeautifulSoup
searching for tags, Another Serving of BeautifulSoup-Regular Expressions
XPath and, Executing JavaScript in Python with Selenium
BeautifulSoup object, Another Serving of BeautifulSoup, Other BeautifulSoup
Objects, Testing Wikipedia
Berne Convention for the Protection of Literary and Artistic Works,
Copyright Law
body tag, The Internet at a Glance
.box file extension, Training Tesseract
building web scrapers
advanced HTML parsing, Advanced HTML Parsing-Beyond BeautifulSoup
crawling across the Web, Starting to Crawl-Crawling with Scrapy
first web scraper, Your First Web Scraper-Connecting Reliably
reading documents, Reading Documents
storing data, Storing Data-Email
using APIs, Using APIs-More About APIs
By object, Executing JavaScript in Python with Selenium
BytesIO object, Microsoft Word and .docx
CAPTCHA characters
about, Image Processing and Text Recognition, Reading CAPTCHAs and
Training Tesseract-Reading CAPTCHAs and Training Tesseract
dragging, Drag and drop
machine training and, Lexicographical Analysis with NLTK, Training
Tesseract-Training Tesseract
retrieving, Retrieving CAPTCHAs and Submitting Solutions-Retrieving
CAPTCHAs and Submitting Solutions
caret (^), Regular Expressions and BeautifulSoup
Carroll , Lewis, An Introduction to BeautifulSoup
Cascading Style Sheets (CSS)
about, Another Serving of BeautifulSoup, The Internet at a Glance
dynamic HTML and, Ajax and Dynamic HTML
hidden fields and, Avoiding Honeypots
CGI (Common Gateway Interface), Running from a Website Hosting Account
checklist, human, The Human Checklist
children (tags), Dealing with children and other descendants
children() function, Dealing with children and other descendants
Chrome developer tool , Radio Buttons, Checkboxes, and Other Inputs
class attribute, Another Serving of BeautifulSoup, find() and findAll() with
BeautifulSoup, Executing JavaScript in Python with Selenium
cleaning dirty data
cleaning after the fact, Cleaning After the Fact-Cleaning
cleaning in code, Cleaning in Code-Data Normalization
client-side processing
handling redirects, Crawling Across the Internet, Handling Redirects
scripting languages and, Scraping JavaScript
cloud computing, Running from the Cloud
colorpickers, Radio Buttons, Checkboxes, and Other Inputs
comma-separated values (CSV) files
reading, CSV-Reading CSV Files
storing data to, Storing Data to CSV-Storing Data to CSV
Comment object, Other BeautifulSoup Objects
Common Gateway Interface (CGI), Running from a Website Hosting Account
Computer Fraud and Abuse Act , The Computer Fraud and Abuse Act, United
States v. Auernheimer and The Computer Fraud and Abuse Act-United States
v. Auernheimer and The Computer Fraud and Abuse Act
Connection header, Adjust Your Headers
Connection object, Integrating with Python
connection/cursor model, Integrating with Python
context-free grammars, Lexicographical Analysis with NLTK
handling, Handling Logins and Cookies-Handling Logins and Cookies,
Handling Cookies-Handling Cookies
verifying settings, Adjust Your Headers
copyright law, Copyright Law-Copyright Law, Field v. Google: Copyright and
cPanel software, Running from a Website Hosting Account
crawling across the Web (see web crawlers)
CREATE DATABASE statement, Some Basic Commands
CREATE INDEX statement, Database Techniques and Good Practice
CREATE TABLE statement, Some Basic Commands
Google accounts, Getting Started
Twitter accounts, A Few Examples
CSS (Cascading Style Sheets)
about, Another Serving of BeautifulSoup, The Internet at a Glance
dynamic HTML and, Ajax and Dynamic HTML
hidden fields and, Avoiding Honeypots
CSV (comma-separated values) files
reading , CSV-Reading CSV Files
storing data to, Storing Data to CSV-Storing Data to CSV
csv library, Reading CSV Files-Reading CSV Files
Cursor object, Integrating with Python
Dark Web, Crawling an Entire Site
data gathering, Crawling an Entire Site, Collecting Data Across an Entire SiteCollecting Data Across an Entire Site
data management
about, Storing Data
email and, Email-Email
MySQL and, MySQL-“Six Degrees” in MySQL
storing data to CSV, Storing Data to CSV-Storing Data to CSV
storing media files, Media Files-Media Files
data normalization, Data Normalization-Data Normalization
data warehouses, Crawling Across the Internet
database size versus query time, Database Techniques and Good Practice
Davies, Mark, Summarizing Data
deactivate command, Installing BeautifulSoup
Deep Web, Crawling an Entire Site
DELETE method (HTTP), Methods
DELETE statement, Some Basic Commands
delete_all_cookies function, Handling Cookies
delete_cookie function, Handling Cookies
delimiters, Storing Data to CSV
descendants (tags), Dealing with children and other descendants
descendants() function, Dealing with children and other descendants
DESCRIBE statement, Some Basic Commands
DHTML (dynamic HTML), Ajax and Dynamic HTML-Ajax and Dynamic
dictionaries, Database Techniques and Good Practice
DictReader object, Reading CSV Files
Digital Millennium Copyright Act (DMCA), Copyright Law
directed graph problems, Six Degrees of Wikipedia: Conclusion
display:none attribute, Avoiding Honeypots
distributed computing, Portability and Extensibility
DMCA (Digital Millennium Copyright Act), Copyright Law
.doc format, Microsoft Word and .docx
documents, reading (see reading documents)
.docx format, Microsoft Word and .docx-Microsoft Word and .docx
dollar sign ($), Regular Expressions and BeautifulSoup
downloading files from Internet, Media Files
drag-and-drop interfaces, Drag and drop
dynamic HTML (DHTML), Ajax and Dynamic HTML-Ajax and Dynamic
Easter Egg, Installation and “Hello, World!”
eBay v. Bidder's Edge, eBay versus Bidder’s Edge and Trespass to Chattels
Echo Nest API, Authentication, Echo Nest-A Few Examples
EditThisCookie Chrome extension, Handling Cookies
elements (Selenium), Executing JavaScript in Python with Selenium,
Interacting with the Site
identifying addresses, Regular Expressions
sending and receiving, Email-Email
email package, Email
encoding (document)
about, Document Encoding
text files and, Text Encoding and the Global Internet-Encodings in action
environment variables, Installing Tesseract
escape characters, Regular Expressions and BeautifulSoup, Cleaning in Code
ethical guidelines, A Note on Ethics-A Note on Ethics, The Legalities and
Ethics of Web Scraping-Field v. Google: Copyright and robots.txt
exception handling
external links, Crawling Across the Internet
handling redirects, Handling Redirects
network connections, Connecting Reliably
suggestions for, Traversing a Single Domain, Collecting Data Across an
Entire Site
explicit wait, Executing JavaScript in Python with Selenium
eXtensible Markup Language (XML), Responses
external links
cautions using, Crawling Across the Internet
crawling across the Internet, Crawling Across the Internet-Crawling with
crawling with Scrapy, Crawling with Scrapy-Crawling with Scrapy
finding, Crawling Across the Internet
Facebook social media site, The Legalities and Ethics of Web Scraping
fair use clause, Copyright Law
fast scraping, Timing Is Everything, The Human Checklist
Fibonacci sequence, A Brief Introduction to JavaScript
Field v. Google, Field v. Google: Copyright and robots.txt
file attribute, Submitting Files and Images
File object, Submitting Files and Images
file uploads, Submitting Files and Images
filtering data, Filtering-Filtering, Processing Well-Formatted Text
finally statement, Integrating with Python
find() function , find() and findAll() with BeautifulSoup-Other BeautifulSoup
findAll() function, Another Serving of BeautifulSoup-Other BeautifulSoup
Objects, Lambda Expressions
for loops, Collecting Data Across an Entire Site
about, Crawling Through Forms and Logins
file uploads and, Submitting Files and Images
handling logins and cookies, Handling Logins and Cookies-HTTP Basic
Access Authentication
hidden fields in, Hidden Input Field Values-Avoiding Honeypots
images in, Submitting Files and Images, Image Processing and Text
input fields supported, Radio Buttons, Checkboxes, and Other Inputs
malicious bots, Other Form Problems
security considerations, Common Form Security Features-Avoiding
submitting basic, Submitting a Basic Form-Submitting a Basic Form
forward slash (\), Regular Expressions and BeautifulSoup
frequency distributions, Statistical Analysis with NLTK-Statistical Analysis
with NLTK
handling in JavaScript, A Brief Introduction to JavaScript
lambda expressions and, Lambda Expressions
gathering data, Crawling an Entire Site, Collecting Data Across an Entire SiteCollecting Data Across an Entire Site
GET method (HTTP)
about, Methods
Google example, A Few Examples
retrieving data , API Calls
tracking requests, Radio Buttons, Checkboxes, and Other Inputs
get_cookies function, Handling Cookies
get_text() function, Another Serving of BeautifulSoup
API examples, How APIs Work, Google APIs-A Few Examples
building, Crawling Across the Internet
Markov model example, Markov Models
Google Analytics, Google Analytics, Handling Cookies
Google Maps , Google Maps
GREL (OpenRefine Expression Language), Cleaning
h1 tag, Running BeautifulSoup, Collecting Data Across an Entire Site
head tag, Encodings in action, The Internet at a Glance
headers (HTTP), Adjust Your Headers-Adjust Your Headers, The Human
hidden fields in forms, Hidden Input Field Values-Avoiding Honeypots
Homebrew package manager, Installing MySQL
homonyms, Lexicographical Analysis with NLTK
honeypots, Hidden Input Field Values-Avoiding Honeypots
Host header, Adjust Your Headers
hotlinking, Media Files
href attribute, Accessing Attributes
HTML (HyperText Markup Language), The Internet at a Glance
HTML Parser library, Beyond BeautifulSoup
HTML parsing
accessing attributes, Accessing Attributes
avoiding the need for, You Don’t Always Need a Hammer
BeautifulSoup example, Another Serving of BeautifulSoup-Dealing with
your parents
lambda expressions, Lambda Expressions
regular expressions, Regular Expressions-Regular Expressions and
html tag, The Internet at a Glance
HTTP (Hypertext Transfer Protocol)
API functionality and, How APIs Work
basic access authentication, HTTP Basic Access Authentication
error handling, Connecting Reliably, The Human Checklist
headers supported, Adjust Your Headers-Adjust Your Headers
methods supported, Methods
HTTPBasicAuth object, HTTP Basic Access Authentication
human checklist, The Human Checklist
HyperText Markup Language (HTML), The Internet at a Glance
Hypertext Transfer Protocol (see HTTP)
hyphen (-), Data Normalization
id attribute, Executing JavaScript in Python with Selenium
image processing
scraping text from images, Scraping Text from Images on Websites-Scraping
Text from Images on Websites
submitting image files, Submitting Files and Images
text recognition and, Image Processing and Text Recognition-Retrieving
CAPTCHAs and Submitting Solutions
img tag, Accessing Attributes
implicit wait, Executing JavaScript in Python with Selenium
indexing, Database Techniques and Good Practice
Innes, Nick, PDF
input tag, Submitting Files and Images
INSERT INTO statement, Some Basic Commands, Integrating with Python
Intel Corp v. Hamidi, eBay versus Bidder’s Edge and Trespass to Chattels
intellectual property, Trademarks, Copyrights, Patents, Oh My!-Copyright
internal links
crawling an entire site, Crawling an Entire Site-Collecting Data Across an
Entire Site
crawling with Scrapy, Crawling with Scrapy-Crawling with Scrapy
traversing a single domain, Traversing a Single Domain-Traversing a Single
about, The Internet at a Glance-The Internet at a Glance
cautions downloading files from, Media Files
crawling across, Crawling Across the Internet-Crawling with Scrapy
moving forward, Moving Forward
IP address blocking, avoiding, Avoiding IP Address Blocking-Avoiding IP
Address Blocking
ISO character sets, A brief overview of encoding types-Encodings in action
is_displayed function, Avoiding Honeypots
Item object, Crawling with Scrapy, Crawling with Scrapy file, Crawling with Scrapy
about, Scraping JavaScript-A Brief Introduction to JavaScript
common libraries, Common JavaScript Libraries-Google Maps
executing with Selenium, Executing JavaScript in Python with SeleniumExecuting JavaScript in Python with Selenium
handling redirects, Handling Redirects
importing files, You Don’t Always Need a Hammer
JavaScript Object Notation (JSON)
about, Responses
parsing, Parsing JSON
jQuery library, jQuery
JSON (JavaScript Object Notation)
about, Responses
parsing, Parsing JSON
Kerr, Orin, United States v. Auernheimer and The Computer Fraud and Abuse
keywords, find() and findAll() with BeautifulSoup
lambda expressions, Lambda Expressions, Media Files
legalities of web scraping, The Legalities and Ethics of Web Scraping-Field v.
Google: Copyright and robots.txt
lexicographical analysis with NLTK, Lexicographical Analysis with NLTKLexicographical Analysis with NLTK
bundling with projects, Installing BeautifulSoup
OCR support, Image Processing and Text Recognition-NumPy
logging with Scrapy, Crawling with Scrapy
about, Crawling Through Forms and Logins
handling, Handling Logins and Cookies-Handling Logins and Cookies
troubleshooting, The Human Checklist
lxml library, Beyond BeautifulSoup
machine learning, Lexicographical Analysis with NLTK, Adjust Your Headers
machine training, Lexicographical Analysis with NLTK, Training TesseractTraining Tesseract
Markov text generators, Markov Models-Six Degrees of Wikipedia:
media files, storing, Media Files-Media Files
Mersenne Twister algorithm, Traversing a Single Domain
methods (HTTP), Methods
Microsoft SQL Server, MySQL
Microsoft Word, Microsoft Word and .docx-Microsoft Word and .docx
MIME (Multipurpose Internet Mail Extensions) protocol, Email
MIMEText object, Email
about, MySQL
basic commands, Some Basic Commands-Some Basic Commands
database techniques, Database Techniques and Good Practice-Database
Techniques and Good Practice
installing, Installing MySQL-Installing MySQL
integrating with Python, Integrating with Python-Integrating with Python
Wikipedia example, “Six Degrees” in MySQL-“Six Degrees” in MySQL
name attribute, Radio Buttons, Checkboxes, and Other Inputs
natural language processing
about, Reading and Writing Natural Languages
additional resources, Additional Resources
Markov models, Markov Models-Six Degrees of Wikipedia: Conclusion
Natural Language Toolkit, Natural Language Toolkit-Lexicographical
Analysis with NLTK
summarizing data, Summarizing Data-Summarizing Data
Natural Language Toolkit (NLTK)
about, Natural Language Toolkit
installation and setup, Installation and Setup
lexicographical analysis, Lexicographical Analysis with NLTKLexicographical Analysis with NLTK
statistical analysis, Statistical Analysis with NLTK-Statistical Analysis with
NavigableString object, Other BeautifulSoup Objects
navigating trees, Navigating Trees-Regular Expressions
network connections
about, Connecting-Connecting
connecting reliably, Connecting Reliably-Connecting Reliably
security considerations, Handling Cookies
next_siblings() function, Dealing with siblings
ngrams module, Statistical Analysis with NLTK
n-grams, Cleaning in Code-Cleaning in Code, Summarizing Data
NLTK (Natural Language Toolkit)
about, Natural Language Toolkit
installation and setup, Installation and Setup
lexicographical analysis, Lexicographical Analysis with NLTKLexicographical Analysis with NLTK
statistical analysis, Statistical Analysis with NLTK-Statistical Analysis with
NLTK Downloader interface, Installation and Setup
NLTK module, Installation and Setup
None object, Connecting Reliably
normalizing data, Data Normalization-Data Normalization
NumPy library, NumPy
OAuth authentication, A Few Examples
OCR (optical character recognition)
about, Image Processing and Text Recognition
library support, Overview of Libraries-NumPy
OpenRefine Expression Language (GREL), Cleaning
OpenRefine tool
about, Cleaning After the Fact
cleaning data, Cleaning-Cleaning
filtering data, Filtering-Filtering
installing, Installation
usage considerations, Using OpenRefine
optical character recognition (OCR)
about, Image Processing and Text Recognition
library support, Overview of Libraries-NumPy
Oracle DBMS, MySQL
OrderedDict object, Data Normalization
os module, Media Files
page load times, Executing JavaScript in Python with Selenium, Timing Is
parentheses (), Regular Expressions
parents (tags), Dealing with children and other descendants, Dealing with your
parsing HTML pages (see HTML parsing)
parsing JSON, Parsing JSON
patents, Trademarks, Copyrights, Patents, Oh My!
pay-per-hour computing instances, Running from the Cloud
PDF files, PDF-PDF
PDFMiner3K library, PDF
Penn Treebank Project, Lexicographical Analysis with NLTK
period (.), Regular Expressions
Peters, Tim, Installation and “Hello, World!”
PhantomJS tool, Executing JavaScript in Python with Selenium-Executing
JavaScript in Python with Selenium, PySocks
PIL (Python Imaging Library), Pillow
Pillow library
about, Pillow
processing well-formatted text, Processing Well-Formatted Text-Scraping
Text from Images on Websites
pipe (|), Regular Expressions
plus sign (+), Regular Expressions
POST method (HTTP)
about, Methods
tracking requests, Radio Buttons, Checkboxes, and Other Inputs
troubleshooting, The Human Checklist
variable names and, Submitting a Basic Form
viewing form parameters, Radio Buttons, Checkboxes, and Other Inputs
previous_siblings() function, Dealing with siblings
primary keys in tables, Database Techniques and Good Practice
programming languages, regular expressions and, Regular Expressions and
projects, bundling with libraries, Installing BeautifulSoup
pseudorandom number generators, Traversing a Single Domain
PUT method (HTTP), Methods
PyMySQL library, Integrating with Python-Integrating with Python
PySocks module, PySocks
Python Imaging Library (PIL), Pillow
Python language, installing, Python at a Glance-Installation and “Hello,
query time versus database size, Database Techniques and Good Practice
quotation marks ("), find() and findAll() with BeautifulSoup
random number generators, Traversing a Single Domain
random seeds, Traversing a Single Domain
rate limits
about, Authentication
Google APIs, Google APIs
Twitter API, Twitter
reading documents
document encoding, Document Encoding
Microsoft Word, Microsoft Word and .docx-Microsoft Word and .docx
PDF files, PDF
text files, Text-Encodings in action
recursion limit, Crawling an Entire Site, “Six Degrees” in MySQL
redirects, Crawling Across the Internet, Handling Redirects
Referrer header, Adjust Your Headers
RegexPal website, Regular Expressions
regular expressions
about, Regular Expressions-Regular Expressions and BeautifulSoup
BeautifulSoup example, Regular Expressions and BeautifulSoup
commonly used symbols, Regular Expressions
programming languages and, Regular Expressions and BeautifulSoup
relational data, MySQL
remote hosting
running from a website hosting account, Running from a Website Hosting
running from the cloud, Running from the Cloud
remote servers
avoiding IP address blocking, Avoiding IP Address Blocking-Avoiding IP
Address Blocking
extensibility and, Portability and Extensibility
portability and, Portability and Extensibility
PySocks and, PySocks
Tor and, Tor-Tor
Requests library
about, Python Requests Library
auth module, HTTP Basic Access Authentication
installing, Python Requests Library, Adjust Your Headers
submitting forms, Submitting a Basic Form
tracking cookies, Handling Logins and Cookies-Handling Logins and
requests module, Adjust Your Headers-Handling Cookies
responses, API calls and, Responses
Robots Exclusion Standard, robots.txt and Terms of Service
robots.txt file, Submitting a Basic Form, Scraping Text from Images on
Websites, robots.txt and Terms of Service-robots.txt and Terms of Service,
Field v. Google: Copyright and robots.txt
safe harbor protection, Copyright Law, Field v. Google: Copyright and
Scrapy library, Crawling with Scrapy-Crawling with Scrapy
screenshots, Taking screenshots
script tag, Scraping JavaScript
search engine optimization (SEO), robots.txt and Terms of Service
searching text data, Lexicographical Analysis with NLTK
security considerations
copyright law and, Copyright Law
forms and, Common Form Security Features-Avoiding Honeypots
handling cookies, Handling Cookies
SELECT statement, Some Basic Commands, Some Basic Commands
Selenium library
about, Handling Logins and Cookies
elements and, Executing JavaScript in Python with Selenium, Interacting
with the Site
executing JavaScript, Executing JavaScript in Python with SeleniumExecuting JavaScript in Python with Selenium
handling redirects, Handling Redirects
security considerations, Avoiding Honeypots
testing example, Testing with Selenium-Unittest or Selenium?
Tor support, PySocks
semicolon (;), Installation and “Hello, World!”
SEO (search engine optimization), robots.txt and Terms of Service
server-side processing
handling redirects, Crawling Across the Internet, Handling Redirects
scripting languages and, Scraping JavaScript
sets, Bringing It All Back Home
siblings (tags), Dealing with siblings
Simple Mail Transfer Protocol (SMTP), Email
site maps, Crawling an Entire Site
Six Degrees of Wikipedia, Traversing a Single Domain-Traversing a Single
SMTP (Simple Mail Transfer Protocol), Email
smtplib package, Email
sorted function, Data Normalization
span tag, Another Serving of BeautifulSoup
Spitler, Daniel, United States v. Auernheimer and The Computer Fraud and
Abuse Act
SQL Server (Microsoft), MySQL
square brackets [], Regular Expressions
src attribute, Accessing Attributes, Media Files, Media Files
StaleElementReferenceException, Handling Redirects
statistical analysis with NLTK, Statistical Analysis with NLTK-Statistical
Analysis with NLTK
storing data (see data management)
StringIO object, Reading CSV Files
strings, regular expressions and, Regular Expressions-Regular Expressions and
about, Another Serving of BeautifulSoup, The Internet at a Glance
dynamic HTML and, Ajax and Dynamic HTML
hidden fields and, Avoiding Honeypots
Surface Web, Crawling an Entire Site
creating in databases, Some Basic Commands
inserting data into, Some Basic Commands
primary keys and, Database Techniques and Good Practice
Wikipedia example, “Six Degrees” in MySQL
Tag object, Other BeautifulSoup Objects
accessing attributes, Another Serving of BeautifulSoup, Accessing
finding based on location in document, Navigating Trees-Dealing with your
finding based on name and attribute, Another Serving of BeautifulSoupOther BeautifulSoup Objects
preserving, Another Serving of BeautifulSoup
Terms of Service, robots.txt and Terms of Service-robots.txt and Terms of
Tesseract library
about, Tesseract
installing, Installing Tesseract
processing well-formatted text, Processing Well-Formatted Text-Scraping
Text from Images on Websites
training, Training Tesseract-Training Tesseract
Tesseract OCR Chopper tool, Training Tesseract
about, An Introduction to Testing
Selenium example, Testing with Selenium-Unittest or Selenium?
unit tests, What Are Unit Tests?, Unittest or Selenium?
unittest module, Python unittest, Unittest or Selenium?
Wikipedia example, Testing Wikipedia-Testing Wikipedia
Text object, Statistical Analysis with NLTK
text processing
image-to-text translation, Image Processing and Text RecognitionRetrieving CAPTCHAs and Submitting Solutions
reading text files, Text-Encodings in action
scraping text from images, Scraping Text from Images on Websites-Scraping
Text from Images on Websites
searching text data, Lexicographical Analysis with NLTK
strings and regular expressions, Regular Expressions-Regular Expressions
and BeautifulSoup
well-formatted text, Processing Well-Formatted Text-Scraping Text from
Images on Websites
The Onion Network (Tor), Tor-Tor
threshold filters, Processing Well-Formatted Text
timestamps, Database Techniques and Good Practice
tokens, Authentication, A Few Examples
Tor (The Onion Network), Tor-Tor
trademarks, Trademarks, Copyrights, Patents, Oh My!
traversing the Web (see web crawlers)
tree navigation, Navigating Trees-Regular Expressions
trespass to chattels, Trespass to Chattels-Trespass to Chattels, eBay versus
Bidder’s Edge and Trespass to Chattels
trigrams module, Statistical Analysis with NLTK
try...finally statement, Integrating with Python
Twitov app, Markov Models
Twitter API, Twitter-A Few Examples
underscore (_), find() and findAll() with BeautifulSoup
undirected graph problems, Six Degrees of Wikipedia: Conclusion
Unicode standard, Integrating with Python, A brief overview of encoding
types-Encodings in action, Cleaning in Code
unit tests, What Are Unit Tests?, Unittest or Selenium?
United States v. Auernheimer, United States v. Auernheimer and The
Computer Fraud and Abuse Act-United States v. Auernheimer and The
Computer Fraud and Abuse Act
unittest module, Python unittest, Unittest or Selenium?
UPDATE statement, Some Basic Commands
urllib library, Connecting, Crawling Across the Internet
urllib.error module, Connecting
urllib.parse module, Connecting
urllib.request module, Connecting, Media Files
urllib2 library , Connecting
urlopen function, Connecting, Encodings in action
urlretrieve function, Media Files
USE statement, Some Basic Commands
User-Agent header, Adjust Your Headers
UTF standards, A brief overview of encoding types, Cleaning in Code
environment, Installing Tesseract
handling in JavaScript, A Brief Introduction to JavaScript
lambda expressions and, Lambda Expressions
versions, multiple, Crawling with Scrapy
virtual environments, Installing BeautifulSoup
Warden, Pete, The Legalities and Ethics of Web Scraping
web crawlers
about, Starting to Crawl
cautions using, Crawling Across the Internet
crawling across the Internet, Crawling Across the Internet-Crawling with
crawling an entire site, Crawling an Entire Site-Collecting Data Across an
Entire Site
crawling with Scrapy, Crawling with Scrapy-Crawling with Scrapy
traversing a single domain, Traversing a Single Domain-Traversing a Single
usage considerations, Trespass to Chattels
web scraping, What Is Web Scraping?-Why Web Scraping?
WebDriver, Executing JavaScript in Python with Selenium-Executing
JavaScript in Python with Selenium, Handling Cookies
analyzing for scraping, The Internet at a Glance
crawling entire, Crawling an Entire Site-Crawling Across the Internet
running from hosting accounts, Running from a Website Hosting Account
scraping text from images on, Scraping Text from Images on WebsitesScraping Text from Images on Websites
testing with scrapers, Testing Your Website with Scrapers-Unittest or
well-formatted text, Processing Well-Formatted Text-Scraping Text from
Images on Websites
whitespace, Storing Data to CSV, Installation and “Hello, World!”
cleaning dirty data, Cleaning in Code-Cleaning in Code
Markov model example, Six Degrees of Wikipedia: Conclusion-Six Degrees
of Wikipedia: Conclusion
MySQL example, “Six Degrees” in MySQL-“Six Degrees” in MySQL
revision history example, Bringing It All Back Home-Bringing It All Back
robots.txt file, robots.txt and Terms of Service
testing example, Testing Wikipedia-Testing Wikipedia
traversing a single domain example, Traversing a Single Domain-Traversing
a Single Domain
Word (Microsoft), Microsoft Word and .docx-Microsoft Word and .docx
w:t tag, Microsoft Word and .docx
XML (eXtensible Markup Language), Responses
XPath (XML Path), Executing JavaScript in Python with Selenium
About the Author
Ryan Mitchell is a software engineer at LinkeDrive in Boston, where she develops
the company’s API and data analysis tools. She is a graduate of Olin College of
Engineering, and is a master’s degree student at Harvard University School of
Extension Studies. Prior to joining LinkeDrive, she built web scrapers and bots at
Abine Inc. and regularly consults on web scraping projects, primarily in the
financial and retail industries.
The animal on the cover of Web Scraping with Python is a ground pangolin
(Smutsia temminckii). The pangolin is a solitary, nocturnal mammal and closely
related to armadillos, sloths, and anteaters. They can be found in southern and
eastern Africa. There are three other species of pangolins in Africa and all are
considered to be critically endangered.
Full-grown ground pangolins can average in size between 12-39 inches in length
and weigh between a mere 3.5-73 pounds. They resemble armadillos, covered in
protective scales that are either dark, light brown, or olive in color. Immature
pangolin scales are more pink in color. When threatened, the scales on the tails can
act more like an offensive weapon, as they are able to cut and wound attackers. The
pangolin also has a defense strategy similar to skunks, in which they secrete a foulsmelling acid from glands located close to the anus. This serves as a warning to
potential attackers, but also helps the pangolin mark territory. The underside of the
pangolin is not covered in scales, but, instead, with little bits of fur.
Like their anteater relatives, pangolin diets consist of ants and termites. Their
incredibly long tongues allow them to scavenge logs and anthills for their meals.
The tongue is longer than their body and retracts into their chest cavity while at
Though they are solitary animals, once matured, the ground pangolin lives in large
burrows that run deep underground. In many cases, the burrows once belonged to
aardvarks and warthogs and the pangolin has simply taken over the abandoned
residence. With the three, long, curved claws found on their forefeet, however,
pangolins don’t have a problem digging their own burrows if necessary.
Many of the animals on O’Reilly covers are endangered; all of them are important
to the world. To learn more about how you can help, go to
The cover image is from Lydekker’s Royal Natural History. The cover fonts are
URW Typewriter and Guardian Sans. The text font is Adobe Minion Pro; the
heading font is Adobe Myriad Condensed; and the code font is Dalton Maag’s
Ubuntu Mono.
What Is Web Scraping?
Why Web Scraping?
About This Book
Conventions Used in This Book
Using Code Examples
Safari® Books Online
How to Contact Us
I. Building Scrapers
1. Your First Web Scraper
An Introduction to BeautifulSoup
Installing BeautifulSoup
Running BeautifulSoup
Connecting Reliably
2. Advanced HTML Parsing
You Don’t Always Need a Hammer
Another Serving of BeautifulSoup
find() and findAll() with BeautifulSoup
Other BeautifulSoup Objects
Navigating Trees
Regular Expressions
Regular Expressions and BeautifulSoup
Accessing Attributes
Lambda Expressions
Beyond BeautifulSoup
3. Starting to Crawl
Traversing a Single Domain
Crawling an Entire Site
Collecting Data Across an Entire Site
Crawling Across the Internet
Crawling with Scrapy
4. Using APIs
How APIs Work
Common Conventions
API Calls
Echo Nest
A Few Examples
Getting Started
A Few Examples
Google APIs
Getting Started
A Few Examples
Parsing JSON
Bringing It All Back Home
More About APIs
5. Storing Data
Media Files
Storing Data to CSV
Installing MySQL
Some Basic Commands
Integrating with Python
Database Techniques and Good Practice
“Six Degrees” in MySQL
6. Reading Documents
Document Encoding
Text Encoding and the Global Internet
Reading CSV Files
Microsoft Word and .docx
II. Advanced Scraping
7. Cleaning Your Dirty Data
Cleaning in Code
Data Normalization
Cleaning After the Fact
8. Reading and Writing Natural Languages
Summarizing Data
Markov Models
Six Degrees of Wikipedia: Conclusion
Natural Language Toolkit
Installation and Setup
Statistical Analysis with NLTK
Lexicographical Analysis with NLTK
Additional Resources
9. Crawling Through Forms and Logins
Python Requests Library
Submitting a Basic Form
Radio Buttons, Checkboxes, and Other Inputs
Submitting Files and Images
Handling Logins and Cookies
HTTP Basic Access Authentication
Other Form Problems
10. Scraping JavaScript
A Brief Introduction to JavaScript
Common JavaScript Libraries
Ajax and Dynamic HTML
Executing JavaScript in Python with Selenium
Handling Redirects
11. Image Processing and Text Recognition
Overview of Libraries
Processing Well-Formatted Text
Scraping Text from Images on Websites
Reading CAPTCHAs and Training Tesseract
Training Tesseract
Retrieving CAPTCHAs and Submitting Solutions
12. Avoiding Scraping Traps
A Note on Ethics
Looking Like a Human
Adjust Your Headers
Handling Cookies
Timing Is Everything
Common Form Security Features
Hidden Input Field Values
Avoiding Honeypots
The Human Checklist
13. Testing Your Website with Scrapers
An Introduction to Testing
What Are Unit Tests?
Python unittest
Testing Wikipedia
Testing with Selenium
Interacting with the Site
Unittest or Selenium?
14. Scraping Remotely
Why Use Remote Servers?
Avoiding IP Address Blocking
Portability and Extensibility
Remote Hosting
Running from a Website Hosting Account
Running from the Cloud
Additional Resources
Moving Forward
A. Python at a Glance
Installation and “Hello, World!”
B. The Internet at a Glance
C. The Legalities and Ethics of Web Scraping
Trademarks, Copyrights, Patents, Oh My!
Copyright Law
Trespass to Chattels
The Computer Fraud and Abuse Act
robots.txt and Terms of Service
Three Web Scrapers
eBay versus Bidder’s Edge and Trespass to Chattels
United States v. Auernheimer and The Computer Fraud and Abuse
Field v. Google: Copyright and robots.txt
Was this manual useful for you? yes no
Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Download PDF