Targeting Sentiment Expressions through Supervised

Targeting Sentiment Expressions through Supervised
Targeting Sentiment Expressions through Supervised Ranking of Linguistic
Jason S. Kessler
Nicolas Nicolov
Dept. of Computer Science
Indiana University
Bloomington, IN 47405, U.S.A.
[email protected]
J.D. Power and Associates, McGraw-Hill
4888 Pearl East Circle
Boulder, CO 80301, U.S.A.
nicolas [email protected]
User generated content is extremely valuable for mining
market intelligence because it is unsolicited. We study
the problem of analyzing users’ sentiment and opinion
in their blog, message board, etc. posts with respect
to topics expressed as a search query. In the scenario
we consider the matches of the search query terms are
expanded through coreference and meronymy to produce a set of mentions. The mentions are contextually
evaluated for sentiment and their scores are aggregated
(using a data structure we introduce call the sentiment
propagation graph) to produce an aggregate score for
the input entity. An extremely crucial part in the contextual evaluation of individual mentions is finding which
sentiment expressions are semantically related to (target) which mentions — this is the focus of our paper.
We present an approach where potential target mentions
for a sentiment expression are ranked using supervised
machine learning (Support Vector Machines) where the
main features are the syntactic configurations (typed dependency paths) connecting the sentiment expression
and the mention. We have created a large English corpus of product discussions blogs annotated with semantic types of mentions, coreference, meronymy and sentiment targets. The corpus proves that coreference and
meronymy are not marginal phenomena but are really
central to determining the overall sentiment for the toplevel entity. We evaluate a number of techniques for
sentiment targeting and present results which we believe push the current state-of-the-art.
1. Introduction
User generated content is extremely valuable for mining
market intelligence because it is unsolicited. The focus of
this paper is on analyzing comments by users about companies, their products and services. We retrieve a large set
of documents, including blogs and message board postings
that mention a company name, its products and its services.
After we filter spam documents (Nicolov and Salvetti 2007;
Kolari, Finin, and Joshi 2006) we apply a combination of
data-driven clustering, and manual specifications of clusters
which result in cluster descriptions which are essentially a
very large IR search query. There are alternative ways of
c 2009, Association for the Advancement of Artificial
Copyright Intelligence ( All rights reserved.
classifying documents to a topic. We are interested in the
demographics and sentiment on a certain topic or subtopic.
The IR query would retrieve relevant documents on a topic
together with matching key phrases. These matching key
phrases serve as input for the sentiment analysis component
which we focus on in this paper.
Conceptually we view the input phrases recovered from
the IR search (we call these anchors) as identifying few of
the mentions of an entity in a document. We expand the input set of mentions through co-reference resolution (Ng and
Cardie 2002; Luo et al. 2004; NIST Speech Group 2006).
Because often when people express sentiment, they express
sentiment toward parts and features of a product (Hu and
Liu 2006) we further expand the current set of mentions
with additional ones reachable from the current ones through
meronymy relations. We further look for coreferential mentions to the newly introduced ones and then seek additional
mentions connected to them through meronymy, etc. Essentially we find the transitive closure of the coreference and
meronymy relations on the initial set of mentions.
Next, we wish to determine which of these mentions are
being evaluated and the polarity of these evaluations. Much
of the evaluation will be rooted in words or phrases which
directly evaluate these mentions. We refer to these as sentiment expressions (SEs) and to the mentions they evaluate as their targets. The polarity and intensity of these SEs
can be found through their semantic properties and by taking into account other elements such as negators and intensifiers that target them. The evaluations rendered by these
SE’s to their target mentions propagate through coreference
and meronymy chains, and aggregately form what we call
entity-level sentiment.
The focus of our paper is on the problem of linking sentiment expressions to the mentions they target. We present
an approach where potential target mentions of an SE are
ranked using supervised machine learning (Support Vector
Machines) where the main features are the syntactic configurations (typed dependency paths) connecting the SE and
the mention. We have created a large English corpus of
blog entries containing critical discussions of products and
includes annotations of semantically typed mentions, their
coreference and meronymy relationships, and sentiment expressions linked to their targets. The corpus proves that
coreference and meronymy are not marginal phenomena but
are really central to determining the overall sentiment for the
top-level entity. We evaluate a number of techniques for sentiment targeting and present results which we believe push
the current state-of-the-art.
2. Related work
The problem of finding targets of objects similar to our sentiment expressions has received considerable recent attention.
In our previous work (Nicolov, Salvetti, and Ivanova
2008) we have used token proximity as a proxy for semantic relatedness. While the technique is fast, however, error
analysis has revealed cases where sentiment expressions are
in proximity of a key phrase match but they are not semantically related. It is exactly these cases that motivated us to
develop the current approach of identifying the targets of the
sentiment expressions.
Hu and Liu (2004), working to find opinions about movie
reviews, provide one of the earliest approaches to the problem of targeted sentiment analysis. Having automatically
annotated adjectival opinion terms, these terms are linked to
adjacent features they modify. Our approach does not require mention/SE adjacency or limit the syntactic types of
sentiment expressions.
Bloom07 consider the task of linking targets of attitudes,
within the appraisal expression framework. Attitudes are
similar to our SEs. An attitude links to a target when two
conditions are met: the typed dependency path between the
two elements matches one of 41 hand-crafted paths, and no
path with a higher rank in the path list links the attitude to another target. Accuracy in the middle 70% range is recorded
for this task over 100 sentences of manually annotated data.
We test a machine learning based approach over a much
larger annotated corpus. Similar to Bloom, Garg, and Argamon, Popescu and Etzioni (2005) use 10 “syntactic dependency rule templates” over a dependency tree to relate identified product features to potential opinion words. Kessler
(2008) also uses a database of typed dependency paths to
find intrasentential subjective relationships. Unlike our approach, Kessler finds relationships that denote the declaration of belief and only considers relationships that link expressions to declarative finite clauses.
Zhuang, Jing, and Zhu (2006), working on the same problem as Hu and Liu (2004), automatically induce a list of
dependency path templates (decorated with part-of-speech
tags) by collecting frequently occurring paths between opinion words and features of movies which co-occur in the
same sentence. Zhuang, Jing, and Zhu do not directly evaluate the linking component of their system, instead showing that their combined opinion-feature linker and opinionterm/feature annotator outperform that of Hu and Liu.
Bethard et al. (2004) consider the task of finding targets
of opinion denoting verbs that are propositional. In contrast,
we find targets of all sentiment expressions, regardless of
their constituent, part-of-speech, or length in tokens. However, we only consider sentiment expressions which target
explicit mentions.
Kim and Hovy (2006) link opinion-bearing adjectives and
verbs (equivalent to our sentiment expressions) to topics (our
mentions) using automatic semantic role labeling (SRL) as
an alternative to using purely syntactic relationships. SRL is
the identification of which semantic roles (e.g., Agent, Patient) some constituents take and the linking of these constituents to predicates. When linked, these constituents are
called frame elements. Their hypothesis is that certain frame
elements are always sources or targets of their predicates.
This is in contrast to syntactic relations which are not sufficient to determine source or target relations. Kim and Hovy
give the examples:
He condemned [ topic the lawyer].
[ topic Her letter] upset me.
Her [ target letter] upset me.
[ topic Her letter and attitude] upset me.
Her [ target letter] and [ topic attitude] upset me.
(1-a-b) demonstrating that, although condemned and upset
are both verbs, they have different arguments as topics– condemned (1-a) linking to its direct object and upset (1-b) to its
subject. However, both topics are in the stimulus frame element. These variations indicate that global syntactic patternbased approaches such as Bloom, Garg, and Argamon are
inadequate. While we do not employ a semantic role labeler
in our study, we do wish to capture these relations. Another
difference is that we do not consider mentions to always be
at the chunk level. For example, (1-c) shows how we would
annotate Kim and Hovy’s (1-b). (1-d) (our annotation would
be (1-e)) shows how annotating below the chunk level increases the difficulty of the task by allowing multiple targets
in the same frame element when coordinates are separated.
Kobayashi et al. (2006) presents an approach to targeting
attitudes in Japanese. Their study is over a corpus of product
reviews annotated with, among other things, linked aspectevaluation. Their “aspects” are equivalent to our mentions
while their “evaluations” are equivalent to our sentiment expressions. They identify these linkages by ranking candidate
aspects by their likelihood of being linked to by a given evaluation. Using the method presented in Iida et al. (2003),
the ranking is computed by training a classifier to determine which of two candidate aspects is more likely to be
linked. The winner of each comparison goes to be classified
against the next unseen aspect until a global winner is found.
The classifier uses features derived from a shallow syntactic
parse as well as the semantic type of each candidate aspect.
We explore a similar approach using similar features though
our corpus is in English. The results of a study comparing
dependency parser performance among different languages
suggests that Japanese could be correctly parsed more reliably than the 13 other studied languages (Nivre 2008). This
leads us to believe that English, not sharing Japanese’s casemarkers, may be more difficult to parse and present more of
a challenge.
Ruppenhofer, Somasundaran, and Wiebe (2008) discusses difficulties in finding sources and targets of subjective expressions. Subjective expressions, which include expressions of belief, speculation, emotion, and evaluation, encompass more than the sentiment expressions which target
mentions of physical entities we consider. Their treatment
of targets is more general than ours, allowing for a target
to be whatever the subjective expression is about. Wilson
(2008) presents the MPQA 2.0 corpus which, among other
things, contains annotations of subjective expressions and
their targets. Section 9 compares the size of the corpus we
developed with MPQA 2.0.
Stoyanov and Cardie (2008a) present a similar corpus to
MPQA 2.0 (also based on MPQA 1.2.) which they call the
MPQA TOPIC corpus. This is annotated with, among other
things, topic spans of opinions that are sentiment-bearing,
which are equivalent to what we call targets. They use the
term “target” to refer to the content of the opinion itself,
for instance the complement clause of a speech event verb.
Implicit topics are also annotated. Stoyanov and Cardie
(2008b) present an approach, based on this corpus, for finding coreferent topics.
3. The larger problem
The focus of this paper is on the problem of finding the
targets of individual expressions of sentiment. In this section we discuss how this task fits into the larger problem of
opinion mining and show how it can be subdivided into approachable tasks.
The larger goal in annotating this corpus is to determine
how sentiment is expressed toward entities (e.g., cars or
cameras) that are being evaluated. We annotate the corpus
under the hypothesis that one can recover this sentiment by
determining which sentiment expressions target mentions of
the entity itself and the entity’s parts and features. Once the
contextual polarity of each of these sentiment expressions is
found, the polar sentiment can be propagated to the top-level
entity. Based on the sentiment toward mentions of parts and
features, and their parts and features, etc... we can determine
part and feature level sentiment, and propagate sentiments
up to their parent parts, eventually coming to an aggregate,
document level sentiment toward a relevant, parent entity.
This approach, if done correctly, will ensure that sentiment
directed toward irrelevant entities will not be considered.
We consider sentiment expressions to be any evaluative
expression that targets a mention of an entity. These exclude
mentions that target entities not explicitly mentioned or toward events. A mention is defined to be any concrete object
or feature of an object that could be evaluated. For instance,
consider the following invented car review. Sentiment expressions are italicized while mentions are underlined.
I like my new Honda Civic.
It gets great gas mileage, has a powerful engine,
and a nice interior!
Sadly, the seats and upholstery are not durable.
I also looked at the Toyota Corolla, but its
engine seemed sluggish.
My mechanic told me yesterday the Civic had
good handling.
Figure 1 describes the meronymy, coreference, and sentiment targeting relations that directly or indirectly affect the
Honda Civic entity. Mentions and sentiment related to the
Corolla entity are not pictured because they are not con-
nected to the graph. Each entity is represented by a rectangle. Each mention of an entity is drawn as an egg within that
entity’s box. Contextual sentiment toward entities as well as
their referent objects are listed at the top of each box. Partof and feature-of relations are drawn between entity boxes.
Sentiment expressions and their modifiers occur outside the
entity boxes and link to the mentions they target. Immediate
opinion holders are treated like any other entity, except they
link to the sentiment expressions which they hold.
While having such a structured document representation
in place makes the task of finding the sentiment toward the
top-level entity considerably easier, it opens up many avenues for opinion mining. For example, queries posed to a
question-answering system such as “What do people not like
about product X?” or “What other features do users who
dislike the camera’s zoom lens feel strongly about?” can
be trivially answered with this data structure. Framing the
opinion mining problem in this manner allows for the use of
a wide array of previous work in computational linguistics.
There have been studies on the problems of finding opinion holders (Kim and Hovy 2006; Bethard et al. 2004),
the contextual polarities of sentiment expressions (Choi
and Cardie 2008; Wilson, Wiebe, and Hoffmann 2005),
and coreference resolution among topics (Stoyanov and
Cardie 2008b) and meronymic relations (Girju, Badulescu,
and Moldovan 2006) between mentions. We will consider
approaches to these important problems in future work.
Polanyi and Zaenen (2006) discuss the problem of determining valences (what we call contextual polarity or sentiment)
of mentions and opinion bearing terms. The issue of determining how sentiment propagates up the subtopic-topic
chain is also discussed. For the remainder of the paper, we
will concentrate on the problem of determining targets of
sentiment expressions.
4. Data
To perform this study, we manually collected 194 blog entries that contained a paragraph or longer evaluative passage
about a car or a digital camera. See Figure 1 for detailed
information about the corpus composition.
The data was tokenized, sentence-split, part-of-speech
tagged with SVMTool (Giménez and Márquez 2004) and
parsed with the Stanford dependency parser (de Marneffe,
MacCartney, and Manning 2006).
Table 1: Corpus size statistics.
4.1. Annotation process
Our annotation philosophy was to allow the annotators to
annotate expressions of sentiment as they intuitively viewed
them. However, each sentiment expression annotation was
to cover the smallest amount of text possible while still
obeying token boundaries.
Figure 1: The sentiment propagation graph of example 2
Mention annotations were performed over dozens of seferences.
mantic types, a subset of which originated in the ACE guideTo evaluate inter-annotator agreement, we adopted the
lines (NIST Speech Group 2006). These include everything
agr metric from Wiebe, Wilson, and Cardie (2005) which
from people to geo-political entities to times, rates, ages to
measures the recall of one annotator’s work using the other’s
locations and facilities. Each mention was annotated with a
annotations as a gold standard. The positions of the two ansemantic type. We included custom, domain specific types
notators are reversed, and the harmonic mean of the two preincluding Car, SUV, CarPart, Truck, Camera and Cameracision values is taken, rendering a single performance metPart. Pronouns that referred to any of these entities were
ric. Formally
|A ∩ B|
marked with their referent’s semantic type.
agr(A||B) =
Some verbs whose nominalizations are features of objects
(e.g., “handles” and “handling”) are also annotated as menwhere A and B are the annotation sets produced by differtions. For example
ent annotators on the same document, and A ∩ B is the set
of overlapping annotations. Table 2 shows promising agree(3)
a. The acceleration is good.
ment statistics. The harmonic mean agr reported for exb. It accelerates well.
pressive subjective elements in (Wiebe, Wilson, and Cardie
(3-a-b) are clearly paraphrases. Both express positive sen2005) of 72% and is 82% agreement for direct subjective
timent toward an object’s acceleration. Some denominal
expressions and speech events. Wilson (2008) reports 85%
verbs may also be marked. For instance, “zoom” in (4-b)
agreement for target annotation. Our scores are similar.
will be annotated as a mention because it refers to the funcNote that contextual polarity of sentiment expressions, intion of a concrete entity.
tensifiers, committers, shifters, and neutralizers were also
annotated. In addition, opinion holder information, compar(4)
a. The camera has good [ target zoom].
isons, coreference, and meronymic relations were also anb. The camera let me [ target zoom] in on the
notated. We will discuss these annotations in more detail in
police car perfectly.
future work.
We intend to make the corpus available for joint projects
It should be noted that sentiment expressions can have muland
possibly wider distribution in the future.
tiple targets. In (5-a), this occurs through conjunctive coordination. In (5-b), it is a semantic property of the sentiment
Coffee and Red Bull are essential to graduate
student life.
[ target Bill] bragged about his [ target Mustang].
In (5-b), bragged targets Bill with negative sentiment and
targets Mustang with positive sentiment held by Bill.
Six car documents were used as annotator training documents and annotated by four annotators. Five more were
marked by three, and 25 by two. One camera document
was annotated by every annotator as a training exercise for
the domain. The remainder of the camera documents were
annotated by only one person. Conflicts in the documents
annotated multiple times were addressed in annotator con-
5. Problem statement
We define the problem of targeting sentiment expressions as
follows. Assume a document is annotated with mentions and
sentiment expressions as described above. Further, assume
we have access to the types of mentions that have been annotated. The task is to identify which mention or mentions a
sentiment expression targets. Given the annotation scheme,
we can assume that every sentiment expression has a target.
Example (6) illustrates the difficulty and importance of
this problem.
Unlike the sluggish Chevy, the Audi R8 performed like a dream.
The Dodge Ram, which I purchased from the
Washington, D.C. Craigslist used auto section,
A vs. B
A1 vs. A2
A1 vs. A3
A1 vs. A4
A2 vs. A4
A2 vs. A3
A3 vs. A4
# Docs
Sent. Exps.
Table 2: Interannotator agreement. Mean refers to harmonic mean. Target agreement is assessed only when sentiment expressions match. We do not take into account mention types when calculating agreement.
turned out to be excellent.
in (6-a), the negative sentiment expression sluggish should
target Chevy. However, it is only three tokens away from
Audi R8, which is being praised. A proximity based approach that would assign sentiment to all mentions within a
window would likely cause sluggish to be considered as targeting. As a simple alternative, an approach that would simply select a sentiment expression’s closest mention would
In (6-b), the sentiment expression excellent and its target,
Dodge Ram, are at least 16 tokens away. This presents two
problems for an n-token window approach. First, the target
is so far away from the sentiment expression that it would
likely fall outside of the window. Second, there are four
mentions in the relative clause that are closer than the target
which are not targeted by the sentiment expression. A system using a shallow parse to evaluate targets for sentiment
expression would be at an advantage. It would prevent the
mentions within the relative clause from linking to a constituent in the main verb phrase.
In Example (1), we have seen that a shallow parse is
not sufficient for determining targeting. Lexico-semantic
knowledge is required.
6. Statistics about the problem:
Why the task is hard
So far, we have presented a few examples that present
challenges for existing methods. The following statistics
suggest that some of the assumptions on which these methods are built may be problematic.
51% of targets appeared to the left of the sentiment expression and 49% appeared to the right. This means that
assuming a target will appear on one side more often than
another will not improve performance.
While the median number of tokens between a sentiment
expression and its target is 2, the mean is 6.21. Making
matters worse, 41% of sentiment expressions have a mention they do not target at least as close to them as a target.
30% have one that is closer. This indicates that a proximity
based targeting algorithm will have a difficult time achieving
a high precision.
91% of targets are in the same sentence as their sentiment
expression. This indicates that high performance is possible without having to perform difficult inter-sentential target
70% of noun phrase targets are common noun phrases,
14% are pronouns, and 16% are proper names. While this
does not have direct implications for target classification, it
does motivate the need for robust semantic type detection
and coreference.
1,002 unique shortest paths along dependency parses connect sentiment expressions to intrasentential targets. 49% of
targets are linked to their sentiment expression by one of
the ten most frequent paths. 58% of targets follow one of
the top 20 paths. 65% of targets follow one of the top 40
paths, and 67% are in the top 50. 73% are in the top 100.
This means that even having a large (> 100) set of patterns
has a substantial ceiling on recall. Table 3 shows how some
of the most frequent paths have very poor precision if they
are treated as targeting whenever they appear between a sentiment expression and a candidate target. For instance, the
path ↑prep,↑pobj is the fourth most frequently targeting pattern in the corpus. However, it only leads to a correct target
32% of the time.
7. Approaches to the problem
When solving this problem, we do not account for intersentential SE-mention linkages. We evaluate four approaches to finding targets on the car and camera domains
both separately and concatenated.
Our first approach is the proximity baseline. We call this
approach Proximity. We select the candidate mention closest to the sentiment expression. Ties are broken by picking
the mention to the right. Ideally, an approach that properly
uses syntactic information will do better than this approach.
The next three approaches make use of dependency parse
trees. However, before the trees can be used they have to
be transformed to accommodate mention and sentiment expression annotations. Following Kessler (2008), we make
each annotation its own node in the tree. The terms covered
by each annotation and the arcs between them are subsumed
by the annotation node. Incoming and outgoing arcs to subsumed nodes are migrated to the annotation nodes. This has
the effect of treating annotations as single words.
The second, Heuristic Syntax, is a baseline syntactic approach. This approach takes advantage of the dependency
parse tree of the sentence where the sentiment expression occurs. The parse tree is first transformed by merging all nodes
that occur in mentions or sentiment expression together into
mention or sentiment expression-level. Next, every node
that governs or is governed by a sentiment expression and is
a mention is considered to be a target. Six of the 20 most frequently occurring relations are longer than one dependency
hop. Also, many single hop relations are unreliable (e.g.,
↑dep and ↓dep have lower than 50% precision).
# Targeting
...changed to beefier stock
It was loud
I hated the 3 speed transmission
↑ prep,↑ pobj
↓ advmod
↓ nn
↑ dep
I’m not letting go of the Cressida
it sure did run
convenience items
Fast, powerful 3S-GTE engine
↑ pobj
↓ rcmod
↑ cop
I really like the base car
’72 Gremlin that needs a lot of work.
it’s only one year old
↓ amod
↑ nmod
↑ dobj
Table 3: Ten most frequently targeting paths. The paths follow the dependency links (labeled with type and direction) from
# times targeting
sentiment expressions to targets. Precision is computed by # times leading to a mention .
The third approach is to employ the pattern list from
Bloom, Garg, and Argamon to predict targets. We call this
approach Bloom. The list of 42 hand-crafted dependency
parse patterns is ordered with the intention of having the first
pattern found be labeled as targeting.
The fourth approach is RankSVM. The objective is to
learn a model that ranks mentions occurring in the same sentence as a sentiment expression such that the ones which are
highest ranked are likely to be targeted. Because we allow
sentiment expressions to target multiple mentions, we also
want to allow multiple mentions to be tied for as the highest
ranking candidate mentions. Casting this as a bipartite ranking problem (Freund et al. 2003) addresses these concerns.
In this formalism, the instance set is partitioned into two
sets: one that is ranked higher and one that is ranked lower.
Here, each sentiment expression has an instance set where
each instance corresponds to a mention in the same sentence. The instances corresponding to mentions which are
targets are given a rank of 1 while the others are given a rank
of 0. We use SVMLight’s implementation of RankSVM
(Joachims 2002) to train a model over these ranked instances. RankSVM trains a preference function with the objective of scoring higher ranking instances higher than lower
ranking instances. The default parameters are used. The feature vectors are formed based on the syntactic and semantic
relationship between the sentiment expression and candidate
mention. This approach allows for multiple targets to be
ranked equally. See Table 4 for the features used. We evaluated this approach using 10-fold cross validation and default
SVMLight parameters.
A benefit of using this approach over Kobayashi et al.’s
instance-to-instance competition is that RankSVM can give
many instances the same score, allowing for multiple targets
to be annotated per sentiment expression. The competition
ensures exactly one target is found and thus cannot handle
cases such as conjunctively coordinated targets.
8. Results
We performed experiments with the approaches mentioned
above. The results are shown in Table 5.
As expected, the Proximity approach performed worse
than Hueristic Syntax and RankSVM.
The biggest surprise was the poor recall of Bloom. Many
of the over 40 patterns used did not connect any sentiment
expression to its target. In contrast to Bloom, Heuristic Syntax did extremely well. Eight of the top ten paths are singletons and thus picked up by Heuristic Syntax. In fact, the performance of Bloom was worse than the Proximity approach
which did not use syntax at all.
Clearly, RankSVM outperformed all other approaches we
evaluated. It won all categories except recall among adjectives in the camera corpus. Adjectives almost always target
the nouns they modify and thus tend to employ short and
regular linking paths. This case is perfectly suited to the
Heuristic Syntax approach, leaving it the winner in this category.
Both Heuristic Syntax and RankSVM could target multiple mentions per sentiment expression. This is reflected in
the larger number of targets each annotated.
We did not notice significant differences in results across
domains. Treating both domains as part of the same data
set generally reduced the performance of RankSVM. This
may indicate that not only general syntactic properties of
the problem were learned in each domain but that the semantic properties of domain specific sentiment expressions
were also learned by the specialized classifiers.
While we did not test our data on the same corpus as used
in Kim and Hovy, we did include statistics for sentiment expressions headed by verbs and adjectives. Like Kim and
Hovy, we found that adjectives were easier to parse than
verbs. With their system achieving an F-score of 66.5%
for verbs and 70.3% for adjectives, it is likely the tournament model outperforms their semantic role labeler-based
system. This is particularly significant because the tournament approach does not use any external lexical resources or
unannotated data.
9. Discussion
In this paper we do not address how sentiment expressions
are identified. We assume there is a component which performs this task. Many systems use a lexicon of sentiment
expressions that have collected in an offline process (e.g.,
(Stone et al. 1966; Turney 2002; Nicolov, Salvetti, and
Ivanova 2008)).
Our results are lower than have been reported in experiments by Bloom, Garg, and Argamon. This is due to a
number of factors. For one, we were working on blog data,
which is notoriously difficult to parse. Furthermore, we had
Feature Name
Lexical Distance
Lexical Path
Lexical Stem Path
Lexical POS Path
Dependency Path
Sent Exps in Path
Mentions in Path
Mention type
POS Relation
Stem Dependency Path
# of tokens separating the mention from the SE
The tokens between the mention and SE
Stems of Lexical Path
Parts of speech of Lexical Path
Shortest dep. path between SE and mention
# of sentiment expressions along Dependency Path
# of mentions along Dependency Path
Semantic type of mention
Parts of speech of the SE and mention heads
Stem of SE concatenated to Dependency Path
to drive the
to drive the
Table 4: Features used in the ranking classifier. “Dep. path” refers to typed-dependency path. Examples are based on the
sentence, “I like to drive the car,” evaluating like and car.
Data Set
both all
car all
camera all
# targets
5742 3663 0.638 0.547
1157 618 0.534 0.458
2824 2031 0.719 0.664
3131 1973 0.630 0.533
699 397 0.568 0.510
1385 979 0.707 0.659
2615 1693 0.647 0.566
465 228 0.490 0.399
1437 1050 0.731 0.667
Heuristic Syntax
4586 3234 0.705 0.483
1277 697 0.546 0.516
2316 1929 0.833 0.631
2399 1658 0.691 0.448
750 418 0.557 0.537
1083 892 0.824 0.601
2186 1579 0.722 0.528
531 283 0.533 0.496
1233 1037 0.841 0.659
3717 2314 0.623 0.346
807 398 0.493 0.295
2270 1591 0.701 0.520
1902 1160 0.610 0.314
481 251 0.522 0.322
1072 730 0.681 0.492
1819 1160 0.638 0.388
331 152 0.459 0.266
1197 861 0.719 0.547
5849 4377 0.748 0.654
1178 866 0.735 0.641
2860 2329 0.814 0.761
3206 2357 0.735 0.637
710 533 0.751 0.684
1408 1135 0.806 0.764
2643 2009 0.760 0.671
468 332 0.709 0.581
1452 1186 0.817 0.753
Table 5: Results of target classification experiments. Results in the top partition are annotated over all documents. Those in
the lower partitions are from the camera and car corpora. N is # of targets classified, C is # correct targets, P is precision, R is
recall, and F is F-Score.
a looser constraint on what, syntactically, could constitute a
sentiment expression or mention.
While we do not address the problem of identifying sentiment expressions or determining their polarity, there are
instances where knowing the target of a potential sentiment
expression is essential to these decisions. These tend to occur in polar facts or utterances that do not contain subjective
language but imply sentiment. Consider,
I broke the handle.
I broke the poorly made handle.
In (7-a), the author is criticizing himself for breaking something. However, in (7-b), the author is effectively heaping
criticism on the handle.
Another challenging aspect of our corpus was that we allowed for sentiment expressions to target multiple mentions.
This problem does not appear to be addressed in similar
work.1 We plan to work on including conjunctive coordinates of the retrieved target as immediate future work.
Another aspect of future work is to create features that are
more amenable to mention-to-mention comparison.
The corpus can also be used to train mention detection,
coreference, and meronymy systems.
We know of two other publicly available corpora that contain opinion-related information in English that include targets of opinions.
The first was presented in Hu and Liu (2004), in which
Agarwal, Prabhakar, and Chakrabarty (2008) performs sentence level sentiment analysis, using conjunctions to determine
which clauses contribute most to the overall polarity of the sentence. This work does not address targeted sentiment analysis.
the topic of each sentence is annotated and its contextual
sentiment value is given. The sentences are drawn from on
line reviews of five consumer electronics devices. It contains
113 documents spanning 4,555 sentences and 81,855 tokens.
While our corpus is larger and contains much richer annotations, it does not contain annotations for implicit sentiment
expressions which are indirectly covered by their approach.
The second is the subset of the MPQA v2.0 corpus containing target annotations (Wilson 2008). The documents
are mostly news articles. It contains 461 documents spanning 80,706 sentences and 216,080 tokens. It contains
10,315 subjective expressions (annotated with links) that
link to 8,798 targets. These subjective expressions are annotated with “attitude types” indicating what type of subjectivity they invoked. 5,127 of these subjective expressions
convey sentiment.
We are continuing the annotation effort adding documents
in diverse domains. By the end of 2009 we expect to have
annotated more than 300,000 tokens. We hope to make the
corpus freely available for research purposes.
10. Conclusions
In this paper we have considered the problem of finding the
semantic relation between sentiment expressions and their
target mentions. This problem arises in the context of sentiment analysis for (top-level) entities identified with a few
mentions (in our case these input mentions come about as
matches from an IR query). We have motivated that it is crucial to augment the initial set of mentions through the transitive closure of coreference and meronymy relations. We
have annotated a large English corpus with product discussions containing semantic types of mentions, coreference,
meronymy and sentiment target relations. We have considered the sentiment targeting using supervised machine
learning (ranking mentions with SVMs) and have also implemented previously introduced techniques to compare our
approach. Our approach is language independent and can be
applied to other languages given availability of a parser and
an annotated corpus. We have improved on the state-of-theart doing experiments on a larger dataset (in particular, we
are referring to the techniques suggested by Bloom, Garg,
and Argamon (2007) which we implemented. Yet, the task
remains quite challenging. We continue to work actively
on this problem — we aim to explore a more sophisticated
set of features. We are also growing the sentiment corpus
adding documents in diverse domains.
We would like to thank Prof. Martha Palmer and Prof. Jim Martin from Colorado University, Dr. Miriam Eckert, Steliana Ivanova,
and Ron Woodward from J.D. Power and Associates, Prof. Michael
Gasser from Indiana University, and Jon Elsas from Carnegie Mellon for help with this research.
Agarwal, R.; Prabhakar, T. V.; and Chakrabarty, S. 2008. “I know
what you feel”: Analyzing the role of conjunctions in automatic
sentiment analysis. In GoTAL.
Bethard, S.; Yu, H.; Thornton, A.; Hatzivassiloglou, V.; and Jurafsky, D. 2004. Automatic Extraction of Opinion Propositions
and their Holders. In 2004 AAAI Spring Symposium on Exploring
Attitude and Affect in Text.
Bloom, K.; Garg, N.; and Argamon, S. 2007. Extracting Appraisal Expressions. In NAACL-HTL.
Choi, Y., and Cardie, C. 2008. Learning with compositional semantics as structural inference for subsentential sentiment analysis. In EMNLP.
de Marneffe, M.; MacCartney, B.; and Manning, C. 2006. Generating typed dependency parses from phrase structure parses. In
Freund, Y.; Iyer, R.; Schapire, R. E.; and Singer, Y. 2003. An
efficient boosting algorithm for combining preferences. J. Mach.
Learn. Res. 4.
Giménez, J., and Márquez, L. 2004. Svmtool: A general pos
tagger generator based on support vector machines. In LREC.
Girju, R.; Badulescu, A.; and Moldovan, D. 2006. Automatic
discovery of part-whole relations. Comput. Linguist. 32(1).
Hu, M., and Liu, B. 2004. Mining and Summarizing Customer
Reviews. In KDD.
Hu, M., and Liu, B. 2006. Opinion Extraction and Summarization
on the Web. In AAAI-2006 Nectar Paper Track.
Iida, R.; Inui, K.; Takamura, H.; and Matsumoto, Y. 2003. Incorporating Contextual Cues in Trainable Models for Coreference
Resolution. In EACL Workshop on the Computational Treatment
of Anaphora.
Joachims, T. 2002. Optimizing search engines using clickthrough
data. In KDD.
Kessler, J. S. 2008. Polling the Blogosphere: a Rule-Based
Approach to Belief Classification. In ICWSM.
Kim, S.-M., and Hovy, E. 2006. Extracting Opinions, Opinion
Holders, and Topics Expressed in Online News Media Text. In
ACL Workshop on Sentiment and Subjectivity in Text.
Kobayashi, N.; Iida, R.; Inui, K.; and Matsumoto, Y. 2006. Opinion Mining on the Web by Extracting Subject-Attribute-Value
Relations. In AAAI-CAAW.
Kolari, P.; Finin, T.; and Joshi, A. 2006. SVMs for the Blogosphere: Blog Identification and Splog Detection. In Computational Approaches to Analyzing Weblogs.
Luo, X.; Ittycheriah, A.; Jing, H.; Kambhatla, N.; and Roukos, S.
2004. A mention-synchronous coreference resolution algorithm
based on the bell tree. In ACL.
Ng, V., and Cardie, C. 2002. Improving machine learning approaches to coreference resolution. In ACL.
Nicolov, N., and Salvetti, F. 2007. Efficient Spam Analysis for
Weblogs through URL Segmentation. In RANLP, volume 292 of
Current Issues in Linguistic Theory (CILT).
Nicolov, N.; Salvetti, F.; and Ivanova, S. 2008. Sentiment Analysis: Does Coreference Matter? In AISB 2008 Convention Communication, Interaction and Social Intelligence.
NIST Speech Group. 2006. The ace 2006 evaluation plan: Evaluation of the detection and recognition of ace entities, values, temporal expressions, relations, and events.
Nivre, J. 2008. Algorithms for Deterministic Incremental Dependency Parsing. Comput. Linguist. 34(4).
Polanyi, L., and Zaenen, A. 2006. Contextual Valence Shifters. In
Computing Attitude and Affect in Text: Theory and Applications.
Popescu, A.-M., and Etzioni, O. 2005. Extracting product features and opinions from reviews. In HLT-EMNLP.
Ruppenhofer, J.; Somasundaran, S.; and Wiebe, J. 2008. Finding
the sources and targets of subjective expressions. In LREC.
Stone, P. J.; Dunphy, D. C.; Smith, M. S.; and Ogilvie, D. M.
1966. The General Inquirer: A Computer Approach to Content
Analysis. MIT Press.
Stoyanov, V., and Cardie, C. 2008a. Annotating Topics of Opinions. In LREC.
Stoyanov, V., and Cardie, C. 2008b. Topic Identification for FineGrained Opinion Analysis. In COLING.
Turney, P. D. 2002. Thumbs up or thumbs down?: semantic
orientation applied to unsupervised classification of reviews. In
Wiebe, J.; Wilson, T.; and Cardie, C. 2005. Annotating Expressions of Opinions and Emotions in Language. In LREC.
Wilson, T.; Wiebe, J.; and Hoffmann, P. 2005. Recognizing
contextual polarity in phrase-level sentiment analysis. In HLTEMNLP.
Wilson, T. A. 2008. Fine-grained Subjectivity and Sentiment
Analysis: Recognizing the Intensity, Polarity, and Attitudes of Private States. Ph.D. Dissertation, University of Pittsburgh.
Zhuang, L.; Jing, F.; and Zhu, X.-Y. 2006. Movie review mining
and summarization. In CIKM.
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