Exploiting Web Search Engines to Search Structured

Exploiting Web Search Engines to Search Structured
WWW 2009 MADRID!
Track: Search / Session: Ads and Query Expansion
Exploiting Web Search Engines to Search Structured
Databases
Sanjay Agrawal, Kaushik Chakrabarti, Surajit Chaudhuri, Venkatesh Ganti
Arnd Christian König, Dong Xin
Microsoft Research
One Microsoft Way
Redmond, WA 98052
{sagrawal, kaushik, surajitc, vganti, chrisko, dongxin}@microsoft.com
ABSTRACT
structured database is searched individually and the relevant
structured data items are returned to the web search engine.
The search engine gathers the structured search results and
displays them along side the web search results. The need
for structured data search is illustrated by the proliferation
of vertical search engines for products [3, 1], celebrities [2],
etc. Typically, these structured databases contain information about named entities like products, people, movies and
locations.
Currently, the search in each structured database is “siloed” in that it exclusively uses the information in the specific
structured database to find matching entities. That is, it
matches the query terms only against the information in its
own database. The results from the structured database
search are therefore independent of the results from web
search. We refer to this type of structured data search as
silo-ed search. This approach works well for some queries.
Consider a product database containing the attributes name,
description, and technical specifications for each product. A
search for a specific product such as [canon XTI] on this
database might find several good matches since the name
and description attributes contain the requisite information
to answer this query. However, there is a broad class of
queries where this approach would return incomplete or even
empty results. For example, consider the query [light-weight
gaming laptop]. Dell XPS M1330 may be a light-weight
laptop that is suitable for gaming but the query keywords
{light-weight, gaming} may not occur in its name, description or technical specifications. Silo-ed search over the above
product database would fail to return this relevant product.
One or more reviews of the product may describe it using
those terms and can help deduce that the product is relevant to the query. However, the structured database may
not contain the comprehensive set of reviews of each product
necessary to identify the relevant products. Hence, a silo-ed
search against the structured database may miss relevant
results.
In contrast, several web documents from product review
sites, blogs and discussion forums may mention the relevant
products in the context of the query keywords {light-weight,
gaming, laptop}. Therefore, the documents returned by a
web search engine are likely to mention products that are
relevant to the user query. However, each of the returned
web documents may contain many relevant products and
even some irrelevant ones. To assemble the relevant products mentioned in documents returned by a web search engine, a user has to read through a potentially large number
Web search engines often federate many user queries to relevant structured databases. For example, a product related
query might be federated to a product database containing
their descriptions and specifications. The relevant structured data items are then returned to the user along with
web search results. However, each structured database is
searched in isolation. Hence, the search often produces
empty or incomplete results as the database may not contain
the required information to answer the query.
In this paper, we propose a novel integrated search architecture. We establish and exploit the relationships between
web search results and the items in structured databases to
identify the relevant structured data items for a much wider
range of queries. Our architecture leverages existing search
engine components to implement this functionality at very
low overhead. We demonstrate the quality and efficiency of
our techniques through an extensive experimental study.
Categories and Subject Descriptors
H.3.3 [Information Systems]: Information Search and Retrieval; H.3.1 [Information Systems]: Content Analysis
and Indexing
General Terms
Design, Performance, Experimentation
Keywords
Structured Database Search, Entity Search, Entity Extraction, Entity Ranking
1.
INTRODUCTION
Many user queries issued against web search engines do
not look for web pages per se, but instead are seeking information from structured databases. For example, the intent of a query such as [canon XTI] is to get information
about the specific product. This query can be better served
by information from a product database. In order to provide more informative results for these scenarios, the web
query is federated to one or more structured databases. Each
Copyright is held by the International World Wide Web Conference Committee (IW3C2). Distribution of these papers is limited to classroom use,
and personal use by others.
WWW 2009, April 20–24, 2009, Madrid, Spain.
ACM 978-1-60558-487-4/09/04.
501
WWW 2009 MADRID!
Track: Search / Session: Ads and Query Expansion
at very low (almost negligible) space and time overheads.
We now briefly discuss the intuition behind our techniques
for achieving these design criteria.
Q: light-weight gaming laptop
Dell Xps M1330, …
Dell Inspiron 1525, …
Web Search
Top k
documents
from web
search
Reviews of
gaming laptops
…
light-weight
Light laptops
At CES…
… good for
gaming
Variety of Structured Data Domains: We desire our technology to be effective for a wide variety of structured data
domains, e.g., products, people, books, movies or locations.
One of the main tasks is to establish the relationships between the web documents and entities in these structured
databases. In this paper, we focus on the “mentions relationship”, that is, a web document is related to an entity if
the entity occurs the document. Our goal is to be able to
establish this relationship, i.e., identify mentions of entities
in web documents for a wide class of entities. This task is
an instance of the entity extraction problem [15, 18]. However, current entity extraction techniques cannot be adopted
in our scenario. Current entity extraction techniques use
machine learning and natural language techniques to parse
documents and break it into sentences, and assign parts of
speech tags for extracting entities. These techniques can be
quite resource-intensive. Even if entity extraction is performed at document indexing time in a web search engine,
the additional overhead is typically unacceptable as it adversely affects the document crawl rates.
Our main insight for overcoming this limitation is that
the structured database defines the universe of entities we
need to extract from web documents. For any entity not in
the structured database, the search engine cannot provide
the rich experience to a user as the entity is not associated
with any additional information. Such entities would not
be returned as results. We therefore constrain the set of
entities that need to be identified in a web document to
be from a given structured database. By doing so, we can
avoid the additional effort and time spent by current entity
extraction techniques to extract entities not in the target
structured database. We leverage this constraint to develop
techniques (i) which can handle a wide variety of structured
data domains, and (ii) which are also significantly more
efficient than traditional entity extraction techniques.
Aggregate and
Rank entities
Document-Entity
Relationships
Dell XPS, Laptop ...
Asus EEE, Laptop ...
Dell Inspiron 1525,
Laptop ...
Structured Database
for Products
Figure 1: Overview over Structured Data Search
of documents, usually beyond the top 2 or 3 documents. Assuming that we can identify the mentions of the products in
web documents, we can automatically identify the most relevant products amongst those mentioned in the top N web
search results based on the following insights. First, how and
where a product is mentioned in the individual returned documents provides evidence about its relevance to the query.
For example, a product mentioned closer to the query keywords in the returned document is likely to be more relevant
than that mentioned farther away from the query keywords.
Second, how often a product is mentioned across the top
web search results also provides important hint about its
relevance to the query. A product mentioned often across
many returned web documents is likely to be more relevant
than another product which is mentioned only a few times.
Hence, we can identify the most relevant products by first
assessing the evidence each individual returned document
provides about the products mentioned in it and then “aggregating” those evidences across the returned documents.
Our Design: In this paper, we study the task of efficiently
and accurately identifying relevant information in a structured database for a web search query. Recall that we consider structured databases that contain information about
named entities. Therefore, we also refer to search for relevant information in a structured database as entity search.
Our approach, illustrated in Figure 1, is to first establish
the relationships between web documents and the entities in
structured databases. Subsequently, we leverage the top web
search results and the relationships between those returned
documents and the entities to identify the most relevant entities. Our approach can return entity results for a much
wider range of queries compared to silo-ed search.
While implementing the above high level approach, our
main design goals are: (i) be effective for a wide variety of
structured data domains, (ii) be integrated with a search
engine and to exploit it effectively. Our techniques use a
combination of pre-processing and web search engine adaptations in order to implement the entity search functionality
Integration with Search Engine Architecture: The challenge
here is to keep the space and time overheads, both at
document indexing and query processing times, very low.
In order to address these challenges, we leverage and adapt
existing search engine components. We discuss the details
in Section 2.
In contrast to our approach, prior entity search approaches [11, 6] do not effectively exploit the web search
engines. These systems rely on sophisticated entity extraction and natural language parsing techniques and custom
indexing structures in order to return entities relevant to
a user’s query. These approaches have significant overhead
which makes it difficult to adopt them at web scale. Furthermore, we show in Section 5 that our techniques produce
higher quality results compared with those prior approaches.
We will discuss these approaches in detail in Section 6.
The remainder of the paper is organized as follows. In Section 2, we describe our approach and the integration with
the search engine architecture. In Section 3, we discuss our
techniques for entity extraction. In Section 4, we discuss
our techniques for aggregating relevance of entities across
the returned documents. In Section 5, we describe an ex-
502
WWW 2009 MADRID!
Track: Search / Session: Ads and Query Expansion
perimental study illustrating quality and efficiency of our
system. In Section 6, we discuss related work and conclude
in Section 7.
2.
Search Engine Components
Indexing
components
ARCHITECTURAL OVERVIEW
Query-time
components
Search
queries
Front (FEX)
End
Front End Experience
In this section, we first describe the key insights of our
search engine-integrated approach for entity search. Then
we describe the adaptations to the search engine to efficiently implement it.
Searchers
Aggregators
Entity Ranker
2.1
Key Insights
Our search engine-integrated approach is based on the following observation. Even though the query keywords may
not match with any entity or their descriptions in the entity
database, the web documents relevant to the query would
typically mention the relevant entities.
Consider the example depicted in Figure 1. Here, we issue
the query [light-weight gaming laptop] to the search engine
and obtain the set of highly relevant documents. A large
fraction of these documents will likely contain mentions of
relevant entities: laptops that are light-weight and suitable
for gaming. Furthermore, those mentions will typically occur in close proximity to the query keywords. By aggregating
the occurrences of the entities in the top N (typically, 10)
web search results, we can obtain the most relevant entities
for the above query.1
In order to enable this functionality, our architecture
needs to (a) establish the ‘mention relationships’ between
entities in the structured database and the web search results and (b) aggregate the entities related to the top web
search results and rank them. Given the response time requirements of search engines, it is not feasible to identify
the required relationships at query time. Hence, we must be
able to (i) identify the required relationships off-line (at document indexing time) and (ii) materialize and index them so
that they can retrieved, aggregated and ranked at query time
at a very low overhead. Our key insight here is to leverage
suitable data structures and processing components already
part of industrial search engines to accomplish the above
goals. We will now describe this approach in detail.
2.2
Web
Crawlers
Document
Crawl
Document
Indexers
Document
Search
Generate snippets
Entity Retrieval
Inverted
Index
Entity Extraction
DocIndex
Entities
Figure 2: Architecture
(query dependent) snippets that are shown to the user. (iv)
The front end implements the experience a search user is
presented with.
To implement the entity search functionality within this
framework, we require three modifications to the components outlined above.
Entity Extraction: The entity extraction component
takes a text document (or a document tokenized into a sequence of tokens) as input, analyzes the document and outputs all the mentions of entities from the given structured
database mentioned in the document. For each mention, it
outputs the entity name, id, and the position in document it
occurs at. As shown in Figure 2, we embed this component
in the search engine’s indexer which already processes and
tokenizes documents for indexing.
When a crawled document is being tokenized and indexed
by the indexer, we invoke the entity extraction component
and obtain the list of entities mentioned in the document.
We add the list of entity mentions to the document’s metadata. The indexer stores this metadata along with the
other metadata of the document (like URL, title) in the
DocIndex. We describe this component further in Section 3.
Architecture
Figure 2 illustrates a high level overview of current search
engine architecture as described in [8]. It has four main
components: (i) The crawlers crawl documents on the web
and store them in a repository. (ii) The indexers process
each document from the crawled repository, and build an
inverted index and a document index. For each keyword,
the inverted index maintains the list of all its occurrences
in all documents. The document index, referred to as the
DocIndex in short, is indexed by document identifier and
contains, for each document, the metadata associated with
the document like URL, title, etc. as well as the raw content
of the document. (iii) The searchers use the inverted index
to return the best K document identifiers and then access
the DocIndex to obtain the URLs, titles, and generate the
Entity Retrieval: The task of entity retrieval is to lookup
the DocIndex for a given document identifier and retrieve
the entities extracted from the document (at document indexing time) along with their positions. A direct implementation of this approach would add an additional access to the
DocIndex per document in the web search results. Because
this cost is incurred at query processing time, the additional
overhead could be excessive.
Our observation here is that current search engines access
the DocIndex anyway to generate snippets for the top
ranking documents. Given a document identifier, the query
keywords and their positions in the document (which is all
information present in the inverted index), they retrieve the
document’s metadata and context of the query keywords
1
Note that a product mention close to the query keywords
does not guarantee that the product is relevant to the query.
For example, the document might say “laptop X is a great
gaming laptop, but look somewhere else if you are looking
for light-weight”. However, it is unlikely that multiple top
results will contain such references and hence, after aggregation, such entities are unlikely to be ranked on top.
503
WWW 2009 MADRID!
Track: Search / Session: Ads and Query Expansion
within each document. This information is used to generate
the snippet for the document. In our architecture, the extracted entities for each document are stored by the indexer
in the DocIndex as metadata of the document. Therefore,
we are able to piggyback on accesses to the DocIndex and
modify the snippet generator to also return the entities
extracted from the document along with their positions.
Since the snippet generator is always invoked for the top 10
web search results, the entity retrieval component always
retrieves the extracted entities for the top 10 documents.
If the user requests for subsequent results, the snippet
generator is invoked for subsequent documents as well.
In that case, the entity retrieval component retrieves the
extracted entities for those documents as well. Because
we leverage existing structures and access patterns, our
approach of retrieving entities adds very little overhead to
a search engine’s query processing time.
Document-Entity List
Entity Categorization
(identify true mentions)
Candidate Entity
mentions + context
Lookup Driven Extraction
(use Entity Reference Tables)
…The Xbox 360
features a standard
DVD drive… …PlayStation 3 has a
Documents
Entity Ranker: The entity ranker ranks the set of all
entities returned by the entity retrieval component. As
shown in Figure 2, we embed the entity ranker in the search
engine’s aggregators. The set of highest-ranked entities are
then returned along with the web document results.
11
Attr.
Xbox 360
…
Playstation 3
…
…
…
Entity Reference Table
Figure 3: Components of Scalable Entity Extraction
3.1
Observe that one key aspect of our architecture is the tight
integration with and the re-use of existing search engine
components. This approach allows us to leverage the high
relevance of web search results and the scalability of the
search engine architecture. But it also poses a challenge: our
components must not impose a significant time and space
overhead on the search engine. In the following two sections
we will focus on the two critical parts of the proposed infrastructure – the entity extraction and the entity ranking –
and show how we address this challenge in detail.
3.
much fancier Blu-Ray
drive which drives up
its cost...
Entity
Lookup-Driven Extraction
We now introduce some preliminary notation and define
the lookup driven entity extraction problem. Let E be the
reference set (from the structured database) of entities. We
associate each entity e with a class c = class(e) and denote
the set of all entity classes as C. Each entity e ∈ E is a
sequence of words. An input document d is also a sequence
of words. For any sequence s = [w1 , . . . , wk ] of words, we
use s[i, j] where (1 ≤ i ≤ j ≤ k) to denote the subsequence
[wi , . . . , wj ]. The triplet (d, e, start) is a candidate mention
of an entity e ∈ E by a document d if the contiguous subsequence d[start, start + |e|] of words in d is equal to the
sequence of words in e.
ENTITY EXTRACTION
The goal of the entity extraction task is to extract entities
from the given structured database that are mentioned in
each document. We need this component to be very efficient
because it is invoked for every document inside the search
engine’s indexer. Traditionally, entity extraction techniques
analyze the entire document: breaking it up into sentences,
identifying noun phrases and then filtering them to identify certain classes of entities. Hence, these techniques are
usually not very efficient.
Our main insight to address the above limitations exploits
the constraint that we have to extract only the entities in
the structured database. This constraint enables us to split
the extraction into two steps, as shown in Figure 3. In the
first lookup driven extraction step we efficiently recognize all
phrases in a document which match with any entity in the
database, using purely syntactic matching of entity strings.
However, some of these phrases may not actually be referring to the entity in the structured database. For example,
the phrase “Pi” in a document may not always be referring
to the entity “Pi” in the movie database. In the second entity mention classification step, we analyze (using machine
learning classifiers) the contexts in which the entity strings
are mentioned in order to identify the mentions that actually refer to the entities in our database. We now describe
these two steps.
Definition 1. (Extraction from an input document) An
extraction from a document d with respect to the reference
set E is the set {m1 , . . . , mn } of all candidate mentions in d
of the entities in E.
Example 1. Consider the reference entity table and the
documents in Figure 3. The extraction from d1 and d2 with
respect to the reference table is {(d1, “Xbox 360”, 2), (d3,
“PlayStation 3”, 1)}.
The lookup-driven entity extraction problem reduces to
the well studied multi-pattern matching problem in the
string matching literature [25]. The goal of multi-pattern
matching is to find within a text string d all occurrences of
patterns from a given set. In our scenario, if each entity is
modeled as a pattern, the lookup-driven entity extraction
problem reduces to the multi-pattern matching problem.
Popular techniques for solving the multi-pattern matching
problem (e.g., Aho-Corasick algorithm) is to build a trie
over all patterns. The trie is used to significantly reduce
the number of comparisons between subsequences of words
in an input document and patterns [19]. We adapt this
technique and build a trie over the sequence of tokens in
each entity of the reference table. We encode each distinct
token by a unique integer thus representing each entity as
a sequence of integers. The encoding allows us to reduce
the overall memory required for storing the trie. During
504
WWW 2009 MADRID!
Track: Search / Session: Ads and Query Expansion
the initialization step, we build a trie over all entities in
the reference set. While processing a document, if a token
sub-sequence starting at position p in a document does not
match the prefix of any entity in the reference set, then the
prefix match fails in the trie. At that point, we can start
matching document sub-sequences starting at p + 1 (or even
later if we maintain additional information in the form of
backlinks per internal node in the trie). Since most subsequences in documents do not match even short prefixes
of any entity, we can skip significant parts of the document
and hence can process the document efficiently.
Note that in typical search engines, the crawled documents are partitioned and each partition is potentially
indexed by a different machine. In such cases, we need to
initialize the trie once per partition or at least once per
distinct machine. Since the initialization time (as shown
in Section 5) is usually insignificant compared to the time
required to process documents in a partition, the additional
overhead due to initialization is acceptable.
filters, and ensure that the entity mention classification filters out false positives. We leave the investigation into these
extreme scenarios as future work.
3.2
Entity Mention Classification
After all candidate mentions within a document have been
identified through lookup-driven extraction, we now have to
classify these into candidate mentions that correspond to entities within our reference set E and those that do not. We
refer to the former as true mentions or simply entity mentions and the latter as false mentions. It is important to note
that this classification cannot be done based on the string of
the candidate mention alone, given that many entity strings
are highly ambiguous. For example, the string “Pi” can refer
to either the movie or the mathematical constant. Similarly,
an occurrence of the string “Pretty Woman” may or may not
refer to the film of the same name.
To address this issue, we leverage the fact that E is a set of
entities from known entity classes such as products, actors
and movies. Therefore, determining whether a candidate
mention refers an entity e in the reference set can be cast as
a classification problem: if the candidate mention belongs
to the corresponding class class(e), it is a true mention of
e, otherwise it is a false mention.
We use linear Support Vector Machines (SVM) [27] to
classify candidate mentions as they are known to be accurate for this type of classification tasks [23, 22]. We now
discuss two important issues for training and applying these
classifiers. The first is that of obtaining training data and
the second is that of which feature sets we use as part of
the classifier.
Approximate Match: In many realistic scenarios, e.g. extraction of product names, the requirement that a sub-string
in a document has to match exactly with an entity string in
the reference table is very limiting. For example, consider
the set of consumers and electronics devices. In many documents, users may just refer to the entity Canon EOS Digital
Rebel XTi SLR Camera by writing “Canon XTI”, or “EOS
XTI SLR” and to the entity Sony Vaio F150 Laptop by writing “Vaio F150” or “Sony F150”. Insisting that sub-strings
in these documents match exactly with entity names in the
reference table would cause these product mentions to be
missed. Therefore, it is very important to consider approximate matches between sub-strings in a document and an
entity name in a reference table [16, 10].
We overcome this limitation by first identifying a set of
“synonyms” for each entity in the reference table. Each synonym for an entity e is an identifying set of tokens, which
when mentioned contiguously in a document refer to e [14].
Once such synonyms are identified and the reference set enhanced with them, the task of identifying approximately
matching entity mentions reduces to that of identifying exactly matching entity mentions. This allows us to leverage
the efficient lookup driven extraction techniques discussed
above. On typical reference sets such as consumer and electronics products, the synonym expansion adds on an average
up to two synonyms per reference entity.
Alternatively, we can also adapt other approaches for
efficiently enabling approximate match [16, 10]. These rely
on known similarity functions which measure similarity
between a sub-string in a document and the target entity
string that it could match with [16, 10].
Generating training data: We require a corpus of labeled
training data for true as well as false mentions. Since obtaining these via human annotation is costly, we propose to
use Wikipedia [4] to automatically generate such examples
for many classes of entities. We now describe the process we
adopt.
For many classes of entities (such as movies, actors, painters, writers), Wikipedia maintains manually curated pages listing the most important instances
of each class (e.g., http://en.wikipedia.org/wiki/List_of_painters_by_name). The pages contain both the
entity string as well as a link to the Wikipedia page of that
entity. We can use the above information to automatically
label candidate mentions of entities in one of the above
lists in Wikipedia documents. If the candidate mention
is linked to the Wikipedia page of the entity, then it is
a true mention. If it is linked to a different page within
Wikipedia, it is a false mention. Because of the size of the
overall Wikipedia corpus, this gives us a significant number
of examples of candidate mentions as well as the contexts
they occur in to train our classifier on.
Memory Requirement: We assume that the trie over all
entities in the reference set fits in main memory. In most
scenarios, this is a fine assumption. Even if the number of
entities (product names) runs up to 10 million, we observed
that the memory consumption of the trie is still less than
100 MB. Given the main memory sizes of current commodity hardware, the trie easily fits in main memory. In those
rare cases where the size of the trie may be larger than that
available, we can consider sophisticated data structures such
as compressed full text indices [24]. Alternatively, we can
also consider approximate lookup structures such as bloom
Feature sets: We use the following 4 types of features:
Document Level Features: The information about the
document where the candidate mention occurs is important
for mention classification. For example, movies tend to
occur in documents about entertainment, and writers in
documents about literature. To capture this information,
we identify important phrases (n-grams, for n = 1, 2, 3)
in a document which encapsulate this information. In our
implementation, we trained the classifier using the 30K
505
WWW 2009 MADRID!
Track: Search / Session: Ads and Query Expansion
n-gram features occurring most frequently in our training
set.
Entity Context Features: The context a candidate
mention occurs in is important for classification where the
context is the sequence of words immediately preceding or
following a candidate mention. For example, a mention of a
movie is often preceded by the words ‘starred in’. We select
the 10K most frequent n-grams within preceding contexts
of entity mentions from the training data. Presence of these
n-grams in candidate mention contexts are indicative of
true entity mentions, so we can increase the accuracy of
our classifiers by using these features. We use a second set
of 10K n-gram features for contexts directly following the
candidate mentions.
Entity Token Frequency Features: The frequency
with which we observe a candidate mention in a random
text document is an important feature for entity mention
classification. For example, the movie title ‘The Others’ is
a common English phrase, likely to appear in documents
without referring to the movie of the same name. Hence,
we are likely to see many false mentions for such phrases.
As a consequence, we include the frequency of each entity
string in a background corpus of English text as a feature.
Entity Token Length Features: Short strings and
strings with few words tend to be more frequent as generic
phrases that do not refer to entities than longer ones.
Therefore, we include the number of words in an entity
string and its length in characters as features.
cause we are exploiting highly relevant documents returned
by a search engine, we observe that even our unsupervised
scoring function produces high quality results (as shown in
Section 5). Our unsupervised scoring function is based on 3
main observations.
• Aggregation: The number of times an entity is mentioned in the top N documents is crucial in determining how
relevant the entity is. The higher the number of mentions,
the more relevant the entity.
• Proximity: The proximity of the mentions of an entity to
the query keywords is also very important for determining
the relevance of the entity. An entity mentioned near the
query keywords is likely to be more relevant to the query
than an entity mentioned farther away from the query
keywords.
• Document importance: Among the top N documents,
some documents are more important than others in determining the relevance of an entity. An entity mentioned in a
document with higher relevance to the query and/or higher
static rank is likely to be more relevant than one mentioned
in a less relevant and/or lower static rank document.
Let D be the set of top N documents returned by the
search engine. Let De be the set of documents mentioning
an entity e. The score of an entity e is computed as follows.
score(e) =
X
imp(d) ∗ score(d, e)
d∈De
Feature Extraction using Lookup-Driven Extraction:
One crucial factor for the features we have selected is that
the value of each of these features can be determined using
(a variant of) lookup-driven extraction during classification
time. In each case, we need to tokenize the input document a
candidate mention occurs in and then lookup the tokens and
n-grams of tokens in lookup tables containing the individual features, their weights and additional information (such
as their frequency in the background corpus). Therefore,
the classifier can benefit from the speed of lookup driven
extraction. In Section 5, we will demonstrate that this, in
combination with the fact that linear SVM-classifiers require
us to only sum up the weights associated with each feature
for classification, leads to a very low overhead for this component.
4.
where imp(d) is the importance of a document d, and
score(d, e) is e’s score from the document d. Note that if
an entity is not mentioned in any document, its score is 0
by this definition. We now define the two components of the
score:
Importance of a Document imp(d): Since the search engine
takes both relevance to the query and static rank into account in order to compute the search results, we rely on the
search engine to obtain the importance of a document. We
use the document’s rank among the web search results to
compute the importance of the document. 2
We partition the list of top N documents into ranges of
equal size B, where 1 ≤ B ≤ N . For each document d, we
compute the range range(d) the document d is in. range(d)
e where rank(d) denotes the document’s rank
is d rank(d)
B
among the web search results. Following the normalized
discount cumulative gain (NDCG) function popularly used
for assigning importance to a ranked list of web documents,
log(2)
we define the importance imp(d) of d to be log(1+range(d))
[21].
ENTITY RANKING
The task of the entity ranker is to rank the set of entities
occurring in the top N (typically 10) pages returned by
search engine. It takes as input the ranked list of top N
URLs, the set of entity mentions and the positions they
occur in for each URL, the positions of the query keywords
for each URL and the number K of entities desired. It
computes the “score” of each entity in the input set using
a scoring function. Subsequently, it returns the K (or less)
entities with scores above a specified threshold, ranked in
decreasing order of their scores. Prior approaches have
proposed to learn the scoring function using a supervised
machine learning model [13]. However, these approaches
require large amounts of labeled training data. Since we do
not have learning datasets for many entity domains, we initially adopt an unsupervised approach. We note that when
sufficient training data is available, existing techniques for
learning ranking functions can be leveraged. However, be-
Score from a Document score(d, e): The score score(d, e) of
an entity e gets from a single document d depends on (i)
whether e is mentioned in the document d, and (ii) on the
proximity between the query keywords and the mentions of
e in the document d.
Both the query keywords and the mentions of an entity
may occur in several positions in a document. Aggregating
2
We can alternatively use the search engine’s relevance score
to compute the importance of the document. However, since
the search engine relevance scores are designed for comparison among themselves and not meant for aggregation, they
need to be suitably transformed for our use. Finding the
appropriate transformation function is hard.
506
Track: Search / Session: Ads and Query Expansion
30
0.7
25
0.6
Mean Reciporcal Rank
Number of results
WWW 2009 MADRID!
20
15
10
5
0
0
100
200
Query Number
Integrated Search
0.5
0.4
0.3
0.2
0.1
0
300
Trec QA 1999
SE(Results Only)
Silo-ed Search
Trec QA 2000
Trec QA 2001
SE(Results+Snippets)
RankExp
Figure 4: Comparison of proposed approach with
Figure 5: Quality of entity ranking on Trec Question
silo-ed search
Answering benchmark
• Effectiveness of entity ranking: Our entity ranking
based on aggregation and query term proximity produces
high quality results.
• Low overhead: Our architecture imposes much lower
overhead on the query time compared to other architectures
proposed for entity search. Furthermore, our extraction
techniques imposes low overhead on the search engine indexer component.
• Accuracy of entity extraction: Our entity classifier
is accurate in identifying the true entity mentions in web
pages.
the proximities among the individual keyword and entity
occurrences cleanly is hard. However, search engines already
employ sophisticated snippet generation techniques. These
techniques linguistically analyze the contexts of the query
keyword occurrences in the document and piece together
the contexts to generate the snippet. Therefore, we observe
an entity appearing in the snippet is a good indicator that
it is “close” to the query keywords in the document.
Note that while detecting the mention of an entity in the
snippet, we may not require the entire string (or its synonym) to occur in the snippet. We can detect partial mentions as long as it unambiguously refers to a single entity
mentioned in the document. That is, the partial match is
contained in only one entity mentioned in the document. For
example, consider the entity ‘Michael Jordan’ mentioned in
a document. If the document’s snippet contains only ‘Jordan’, we detect it as a mention of ‘Michael Jordan’ if he is
the only entity mentioned in the document that contains the
token ‘Jordan’. Note that we do not require ‘Jordan’ to unambiguously refer to ‘Michael Jordan’ across all documents.
We define the score(d, e) as follows.
score(d, e)
All experiments reported in this section were conducted on
an Intel 6600 PC (2.4 GHz) with a single core and 4GB
RAM, running Windows 2003 Server.
5.1
= 0, if e is not mentioned in d
= wa , if e occurs in snippet
=
wb , otherwise
where wa and wb are two constants. The value wa is the
score an entity gets from a document when it occurs in the
snippet, and wb is the score when it does not. In our implementation, we set B = 5, wa = 1.0, and wb = 0.3.
5.
Comparison with Silo-ed Search
We compared our integrated entity search approach with
silo-ed search on a product database containing names and
descriptions of 187,606 Consumer and Electronics products.
We implemented silo-ed search on the above database by
storing the product information in a table in Microsoft SQL
Server 2008 (one row per product), concatenating the text
attributes, viz. product name and description, into a single attribute and building a full text index on that concatenated column using SQL Server Integrated Full Text
Search engine (iFTS). Subsequently, given a keyword query,
we issue the corresponding AND query on that column using
iFTS and return the results. For example, for the keyword
query [lightweight gaming laptop], we issue the iFTS query
(’lightweight’ AND ’gaming’ AND ’laptop’) on the concatenated column. We issue AND queries (instead of OR queries
or freetext queries) since users typically expect conjunctive
semantics in web search, i.e., a relevant entity is expected
to contain all (or most) of the keywords in the name and
description fields.
We implemented the integrated entity search approach for
the above database (consisting of the product names and
their synonyms) using the Live Search infrastructure. In
all our experiments, we use the top 10 documents returned
EXPERIMENTAL EVALUATION
We now present the results of an extensive empirical
study to evaluate the approach proposed in this paper. The
major findings of our study can be summarized as follows.
• Superiority over silo-ed search: Our integrated entity
search approach overcomes the limitation of silo-ed search
over entity databases. It returns relevant results when
silo-ed search fails to return any relevant results.
507
WWW 2009 MADRID!
Track: Search / Session: Ads and Query Expansion
0.7
0.9
0.6
0.5
Precision
Precision
0.8
0.7
0.6
0.3
0.5
0.2
0.4
0.1
0.3
0
0
0.2
SE(Results+Snippets)
0.4
Recall
0.6
0.3
0.8
0.5
0.7
0.9
Recall
SE(Results+Snippets)
SE(Results Only)
Figure 6: Quality of entity ranking on IMDB dataset
SE(Results Only)
Figure 7: Quality of entity ranking on Wikipedia
dataset
answer to query q is at rank rq , award a score of
by Live Search engine to aggregate and rank the entities.
This is because the snippet generator is always invoked for
these documents. It might be possible to aggregate over
more documents if the user requests for subsequent results.
Aggregating over more documents (i.e., 20 or 30) typically
improves the quality of entity results. We restrict our evaluation to top 10 documents in this paper.
We took a random sample of 316 Consumer and Electronics queries 3 from the Live search query log. Figure 4 shows
the number of results returned by the two approaches for
the 316 queries. Integrated entity search returned answers
for most queries (for 98.4% of the queries) whereas silo-ed
search often returned empty results (for about 36.4% of the
queries). This is because the name and description of a
product often does not have enough information to answer
a query but web pages mentioning the product name do.
5.2
0.4
1
1
Σ
Q q∈Q rq
1
rq
and
average over the query set Q to get MRR
. Figure
5 shows our entity ranking has high quality (M RR = 0.6),
i.e., the actual answer appeared among the top 1-2 results.
Recall that our entity ranking depends on two main insights,
viz., aggregation and proximity, we evaluate the individual impact of the two insights on quality by measuring the
quality of entity ranking without using snippets (i.e., taking aggregation but not proximity into account) and with
snippets (i.e, taking both aggregation and proximity into
account). Figure 5 shows that using snippets improved the
MRR significantly (e.g., from 0.53 to 0.64 for the TREC 2001
queries). We also compare our technique with the RankExp
entity ranking technique that does not leverage the search
engine results [13]. Since our technique is based on the web
search engine and RankExp is difficult to implement on the
web corpus, we compare with the results of RankExp for
the same queries (on the ACQUAINT corpus) as reported
in [13]. Our entity ranking significantly outperforms the
MRR of RankExp as shown in Figure 5.
IMDB Data: We obtained the movie names and the people
(actors, director, producer, etc.) associated with them from
the MSN Movie database. For example, for the movie “The
Godfather”, the associated people include Francis Ford Coppola, Mario Puzo, Marlon Brando, etc. We treat the movie
name as the query and the associated people as ground truth
for the query. We used 250 such queries along with the
ground truth. The ground truth contained, on average, 16.3
answers per query. Figure 6 shows the precision-recall tradeoff for our integrated entity search technique. We vary the
precision-recall tradeoff by varying the number k of topranked answers returned by the integrated entity search approach (K = 1, 3, 5, 10, 20 and 30). The integrated entity
search technique produces high quality answers: the precision is 86%, 81% and 79% for the top 1, 3 and 5 answers
respectively. Using the snippet (i.e., taking proximity into
account) improves the quality significantly: the precision improved from 67% to 86%, from 70% to 81% and from 71%
to 79% for K = 1, 3 and 5 respectively. Furthermore, the
Entity Ranking
We evaluate the quality of our entity ranking on
3 datasets:
TREC Question Answering benchmark,
Wikipedia data and IMDB data. We focus on queries
with person entities as answers. We used an entity reference
table containing 2.04 million person names (and their synonyms 4 ) obtained from IMDB, Wikipedia, DBLP, ACM as
well as several internal databases. Note that our approach
is designed to return only the entities in the structured
database. We therefore ensure that, for all the 3 datasets,
the relevant entities (along with their synonyms) for all the
queries are present in the entity reference table.
TREC Question Answering benchmark: We used the
factoid questions from the TREC Question Answering track
for years 1999, 2000 and 2001. We focussed on the 160
queries with person entities as answers. We evaluated the
quality of our entity ranking using the “mean reciprocal
rank” (MRR) used in the TREC community: if the first
3
These are the queries classified as Consumer and Electronics queries by the Live Search query classifier.
4
We generated the synonyms of person names by using common first name synonyms (’Michael’ ≡ ’Mike’) and dropping/abbreviating middle names and initials.
508
WWW 2009 MADRID!
Track: Search / Session: Ads and Query Expansion
100%
Structured Database
database size
95%
Precision
90%
85%
80%
Movies
Products
People
75%
70%
Context + Document + Frequency Features
800K
200K
2 million
Average overhead
per
query
for
head
query
workload
0.022 ms
0.019 ms
0.029 ms
Average
overhead per
query for commerce query
workload
0.017 ms
0.070 ms
0.026 ms
65%
60%
Table 1: Query time overhead of our approach
Context + Document + Frequency + Token Features
55%
50%
0%
20%
40%
60%
80%
100%
Recall
Figure 8: Precision/Recall tradeoff for Entity Classification
MRR improves from 0.8 to 0.91.
Wikipedia Data: We extracted 115 queries and their
ground truth from Wikipedia. We focussed on queries
whose ground truth can be extracted easily from Wikipedia
pages using automated scripts. For example, we extracted
ceo queries (e.g., [general electric ceo]) whose ground truth
(the corresponding ceo names) can be extracted from the
list page http://en.wikipedia.org/wiki/List_of_chief_
executive_officers. We extracted nobel prize winner
queries (e.g., [nobel prize 2006]) whose ground truth (the
names of the winners) can be easily extracted from the year
pages (e.g., http://en.wikipedia.org/wiki/2006). The
ground truth contained, on average, 3.6 answers per query.
Figure 7 shows the precision-recall tradeoff for our integrated
entity search technique for various values of K (K = 1, 3,
5, 10, 20 and 30). The integrated entity search approach
produces high quality results for this collection as well: the
precision is 61% and 42% for the top 1 and 3 answers respectively. Again, using the snippets significantly improves
quality: the precision improved from 54% to 61% and from
35% to 42% for k = 1 and 3 respectively. Furthermore, the
MRR improves from 0.67 to 0.74.
5.3
tecture imposes negligible query time overhead on existing
search engines.
Extraction Overhead: Since the extraction takes place
in the indexer component of the search engine, it is important for the extraction overhead to be low so that the search
engine crawl rate is not affected. The extraction time consists of 3 components: the time to scan and decompress the
crawled documents (performed in chunks of documents), the
time to identify all candidate mentions using lookup-driven
extraction and the time to perform entity mention classification using the classifier described in Section 3.2. We
measured the above times for a web sample of 30310 documents (250MB in document text). The times for each of the
individual components are 18306 ms, 51495 ms and 20723
ms respectively. The total extraction time is the sum of the
above three times: 90524 ms. Hence, the average extraction overhead at less than 3ms per web document, which is
acceptable in our architecture.
5.4
Accuracy of Entity Classification
We now evaluate the accuracy of the classifier described
in Section 3.2 to identify true mentions from the candidate
mentions identified by the lookup driven extractor. For this
purpose, we use separate training and test corpora of a total
100K instances of movie mentions found within Wikipedia
documents and generated automatically via the link-analysis
method described in Section 3.2. We use a linear Support
Vector Machine (SVM) [27] trained using Sequential Minimal Optimization [26] as the classifier. This means that
once all features for a given candidate mention have been
obtained, the computation of the corresponding classifier
output corresponds to a summation of the corresponding
feature weights, which can be performed efficiently.
We measured the accuracy of the classifier with regards
to differentiating true mentions from the false ones among
the set of all candidates. 54.2% of all entity candidates in
the test data are true mentions. Our classifier achieves an
accuracy of 97.2%. Furthermore, by varying the threshold
(the distance from the separating hyperplane in case of this
classifier), we can tradeoff additional precision against lower
recall for the extracted true mentions. Figure 8 shows the
resulting precision/recall tradeoff.
Overhead of our Approach
Query Time Overhead: Since we piggyback on snippet
generator for entity retrieval, the only overhead of our approach at query time is that of aggregating the retrieved
entities from top 10 documents and ranking them. We measured the time taken for aggregating and ranking the entities
for two different workloads of 1000 queries each for various
entity databases like movie, product (Consumer and Electronics), and people databases. In the first workload we used
1000 head queries from the search query logs. In the second
workload, we used 1000 commerce search queries. Table 1
shows the query time overhead of our approach per query
for the two workloads. In all cases, the overhead is less than
0.1 milliseconds which is negligible compared to end-to-end
query times of web search engines (several tens of milliseconds). We observe that even for the commerce queries on the
product database where the aggregation and ranking costs
are relatively high (since most of the documents returned
for these queries contain large number of product mentions),
the overhead remains below 0.1 ms. In summary, our archi-
6.
RELATED WORK
The problem of entity search has been studied recently [13,
12, 6]. In [13], the authors propose a specialized indexing
scheme that maintains an inverted index called the “atype
index” in addition to the standard word index. The aytpe index maintains, for each entity type, the list of document IDs
that contains occurrences of one or more entities of that type
509
WWW 2009 MADRID!
Track: Search / Session: Ads and Query Expansion
along with the positions of those occurrences. Given a keyword query and the desired entity type, the system performs
a mutli-way intersection between the keyword posting lists
and the atype posting lists, computes the score of the entities in the atype posting list and return the K entities with
the highest scores. This architecture results in a bloating of
the inverted index (by a factor of 5). To reduce the bloating,
the authors propose to maintain the posting lists for only a
small number of entity types. At query time, they find the
best generalization of the desired entity type for which the
posting list exists. The query processor performs the above
intersection with the posting list of the generalized entity
type and then uses the “reachability index” to remove the
false positives. This technique increases the query time by a
factor of 1.9 due to additional intersections and the lookups
in the reachability index involved. This overhead makes it
difficult to integrate this architecture in commercial search
engines.
In [6], the authors present another entity search technique
that creates a concordance document for each entity, consisting of all the sentences in the corpus containing that entity.
They then index and search these documents using a standard IR engine like Lucene. The result of the search returns
a ranked list of entities. The advantage of this approach is
that it can leverage an existing search engine without major
modifications. However, this approach may produce large
concordance documents and hence a big index, resulting in
high search times. Furthermore, the above querying over
the entity index has to be performed in addition to the traditional querying over document indexes. Hence, this architectures can add significant overhead to web search engines
in terms of query processing time and is hence difficult to
adopt at web scale. Note that neither of the above two approaches exploit the results produced by the search engine.
Systems like KnowItAll and TextRunner [17, 9] focus on
extracting entities and relationships from web pages and
populating structured relations. In [9], the authors propose
to index the tuples in the structured relations using an inverted index and support keyword search on them. Like
the entity search approaches discussed above, this approach
does not leverage the search engine.
Several techniques have been proposed for keyword search
on databases (e.g., [5, 20, 7]). All these techniques represent
implementations of silo-ed search. As mentioned before, siloed search often results in empty or incomplete results due
to the limited amount of information in the database. The
search engine integrated approach proposed in this paper
overcomes the above limitation.
8.
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
7.
CONCLUSION
[24]
In this paper, we showed that establishing and exploiting relationships between web search results and structured
entity databases significantly enhances the effectiveness of
search on these structured databases. We proposed an architecture which effectively leverages existing search engine
components in order to efficiently implement the integrated
entity search functionality. Specifically, our techniques add
very little space and time overheads to current search engines while returning high quality results from the given
structured databases.
[25]
[26]
[27]
510
REFERENCES
http://search.live.com/products/.
http://search.live.com/xrank?form=xrank1.
http://www.google.com/products.
www.wikipedia.org.
S. Agrawal, S. Chaudhuri, and G. Das. DBXplorer: A
System for Keyword-Based Search over Relational
Databases. In IEEE ICDE Conference, 2002.
M. Bautin and S. Skiena. Concordance-Based
Entity-Oriented Search. In Web Intelligence, pages
586–592, 2007.
G. Bhalotia, A. Hulgeri, C. Nakhe, S. Chakrabarti, and
S. Sudarshan. Keyword Searching and Browsing in
Databases using BANKS. In ICDE Conf., 2002.
S. Brin and L. Page. The anatomy of a large-scale
hypertextual web search engine. Comput. Netw. ISDN
Syst., 30(1-7), 1998.
M. Cafarella, M. Banko, and O. Etzioni. Relational Web
Search. In WWW Conference, 2006.
K. Chakrabarti, S. Chaudhuri, V. Ganti, and D. Xin. An
efficient filter for approximate membership checking. In
ACM SIGMOD Conference, 2008.
S. Chakrabarti. Breaking Through the Syntax Barrier:
Searching with Entities and Relations. In PKDD
Conference, pages 9–16, 2004.
S. Chakrabarti. Dynamic Personalized Pagerank in
Entity-Relation Graphs. In WWW Conference, pages
571–580, 2007.
S. Chakrabarti, K. Puniyani, and S. Das. Optimizing
Scoring Functions and Indexes for Proximity Search in
Type-annotated Corpora. In WWW Conference, 2006.
S. Chaudhuri, V. Ganti, and D. Xin. Exploiting web search
to generate synonyms for entities. In WWW Conference,
2009.
W. Cohen and A. McCallum. Information Extraction and
Integration: an Overview. In SIGKDD Conference, 2004.
W. Cohen and S. Sarawagi. Exploiting dictionaries in
named entity extraction: combining semi-markov
extraction processes and data integration methods. In ACM
SIGKDD Conference, 2004.
O. et. al. Web-scale information extraction in knowitall. In
WWW Conference, 2004.
R. Grishman. Information Extraction: Techniques and
Challenges. In SCIE, 1997.
D. Gusfield. Algorithms on Strings, Trees and Sequences.
Cambridge University Press, 1997.
V. Hristidis and Y. Papakonstantinou. DISCOVER:
Keyword Search in Relational Databases. In 28th VLDB
Conf., 2002.
K. Järvelin and J. Kekäläinen. Cumulated gain-based
evaluation of ir techniques. ACM Trans. Inf. Syst., 20(4),
2002.
T. Joachims. Text Categorization with Support Vector
Machines: Learning with many Relevant Features. In
EMNLP Conference, 1998.
T. Joachims. Training Linear SVMs in Linear Time. In
ACM SIGKDD Conference, pages 217 – 226, 2006.
G. Navarro and V. Mäkinen. Compressed full-text indexes.
ACM Comput. Surv., 39(1), 2007.
G. Navarro and M. Raffinot. Flexible Pattern Matching in
Strings. Cambride University Press, 2002.
J. Platt. Fast Training of SVM’s Using Sequential Minimal
Optimization. In Advances in Kernel Methods: Support
Vector Machine Learning, pages 185–209. MIT Press, 1999.
V. Vapnik. Statistical Learning Theory. Whiley, 2000.
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

advertisement