MARCH/APRIL 2013 1541-1672/13/$31.00 © 2013 IEEE 15
Published by the IEEE Computer Society
Knowledge-Based approaches to
concept-level sentiment analysis
New Avenues in
Opinion Mining and
Sentiment Analysis
Erik Cambria, National University of Singapore
Björn Schuller, Technical University of Munich
Yunqing Xia, Tsinghua University
Catherine Havasi, Massachusetts Institute of Technology
The Web holds
valuable, vast,
and unstructured
information about
public opinion. Here,
the history, current
use, and future of
opinion mining and
sentiment analysis
are discussed,
along with relevant
techniques and tools.
of information were friends and special-
ized magazine or websites. Now, the “social
web” provides new tools to efciently create
and share ideas with everyone connected to
the World Wide Web. Forums, blogs, social
networks, and content-sharing services help
people share useful information. This infor-
mation is unstructured, however, and be-
cause it’s produced for human consumption,
it’s not something that’s “machine process-
able.” Capturing public opinion about social
events, political movements, company strat-
egies, marketing campaigns, and product
preferences is garnering increasing interest
from the scientic community (for the excit-
ing open challenges), and from the business
world (for the remarkable marketing fall-
outs and for possible nancial market pre-
diction). The resulting emerging elds are
opinion mining and sentiment analysis. Al-
though commonly used interchangeably to
denote the same eld of study, opinion mining
and sentiment analysis actually focus on po-
larity detection and emotion recognition,
respectively. Because the identification of
sentiment is often exploited for detecting
polarity, however, the two elds are usually
combined under the same umbrella or even
used as synonyms. Both elds use data min-
ing and natural language processing (NLP)
techniques to discover, retrieve, and distill
information and opinions from the World
Wide Web’s vast textual information.
Mining opinions and sentiments from
natural language is challenging, because
it requires a deep understanding of the ex-
plicit and implicit, regular and irregular,
and syntactical and semantic language
rules. Sentiment analysis researchers strug-
gle with NLP’s unresolved problems: co-
reference resolution, negation handling,
anaphora resolution, named-entity recogni-
tion, and word-sense disambiguation. Opin-
ion mining is a very restricted NLP problem,
O
thers’ opinions can be crucial when it’s time to make a decision or
choose among multiple options. When those choices involve valuable
resources (for example, spending time and money to buy products or services)
people often rely on their peers’ past experiences. Until recently, the main sources
IS-28-02-Cambria.indd 15 6/5/13 11:05 AM
16 www.computer.org/intelligent IEEE INTELLIGENT SYSTEMS
Knowledge-Based approaches to concept-level
sentiment analysis
because the system only needs to
understand the positive or negative
sentiments of each sentence and the
target entities or topics. Therefore,
sentiment analysis is an opportunity
for NLP researchers to make tangi-
ble progress on all fronts of NLP,
and potentially have a huge practical
impact.
Many companies use opinion min-
ing and sentiment analysis as part
of their research. For instance, com-
panies use opinion mining to create
and automatically maintain review
and opinion-aggregation websites.
Their systems continuously gather
a wide array of information from
the Web, such as product reviews,
brand perception, and political is-
sues. Other systems might also use
opinion mining and sentiment anal-
ysis as subcomponent technology to
improve customer relationship man-
agement and recommendation sys-
tems through positive and negative
customer feedback. Similarly, opinion
mining and sentiment analysis might
detect and exclude “flames” (overly
heated or antagonistic language) in
social communication and enhance
antispam systems.
Companies use sentiment analysis
to develop marketing strategies by
assessing and predicting public atti-
tudes toward their brand. Research
and development focuses on design-
ing automatic tools that crawl online
reviews and condense the infor-
mation gathered. Numerous compa-
nies already provide tools that track
public viewpoints on a large scale by
offering graphical summarizations
of trends and opinions in the blogo-
sphere. Developing opinion-tracking
systems is commercially important.
Also, several tools already exist to
help companies extract and analyze
information from blogs about large-
scale trends in customers’ opinions
about products; those tools include
SenticNet (http://sentic.net), Luminoso
(http://luminoso.com), Factiva (http://
dowjones.com/factiva), Attensity
(http://attensity.com), and Converseon
(http://converseon.com). Most existing
tools and research, however, are lim-
ited to polarity evaluation or mood
classication according to a limited
set of emotions. Such methods mainly
rely on parts of text in which people
explicitly express emotional states,
and therefore the tools can’t capture a
reviewer’s implicitly expressed opin-
ion or sentiment. To better consider
the state of this eld, we discuss here
the past, present, and future trends
of sentiment analysis by delving into
the evolution of opinion mining sys-
tems. More comprehensive surveys
on sentiment analysis can be found
elsewhere.
1–3
Common Sentiment
Analysis Tasks
The basic task of opinion mining is
polarity classication. Polarity clas-
sication occurs when a piece of text
stating an opinion on a single issue is
classied as one of two opposing sen-
timents. Reviews such as “thumbs
up” versus “thumbs down,” or “like”
versus “dislike” are examples of po-
larity classication. Polarity classi-
cations also identify pro and con ex-
pressions in online reviews and help
make the product evaluations more
credible.
Agreement detection is another
form of binary sentiment classica-
tion. Agreement detection determines
whether a pair of text documents
should receive the same or different
sentiment-related labels. After the
system identies the polarity classi-
cation, it might assign degrees of
positivity to the polarity—that is, it
might locate the opinion on a con-
tinuum between positive and nega-
tive. Also, it can classify multi-
media resources according to mood and
emotional content for purposes such
as affective human-machine interac-
tion, troll ltering, and cyber-issue
detection. If the text doesn’t contain
strong opinions or covers more than
one issue or item, new challenges
arise, such as subjectivity detection
and opinion-target identification.
Distinguishing between subjective
and objective text helps classify the
sentiment. Moreover, a piece of text
might have a polarity without neces-
sarily containing an opinion; for ex-
ample, a news article can be classied
into good or bad news without being
subjective.
Typically, a system performs sentiment
analysis over on-topic documents—
using, for example, the results of a
topic-based search engine. However,
several studies suggest that managing
these two tasks jointly might benet
overall performance. For example, a
document’s off-topic passages might
contain irrelevant affective informa-
tion and create inaccurate global-
sentiment polarity about the main
topic. Also, a document might con-
tain information on multiple top-
ics that interest the user. In such
instances, it’s important to identify
topics and separate the opinions asso-
ciated with each topic.
Evolution of Opinion Mining
Currently, opinion mining and senti-
ment analysis rely on vector extrac-
tion to represent the most salient and
important text features. We can use
this vector to classify the most relevant
features. Two commonly used features
are term frequency and presence.
Presence is a binary-valued feature
vector in which the entries indicate
only whether a term occurs (value 1)
or doesn’t (value 0). Presence forms a
more effective basis to review polar-
ity classication and reveals an inter-
esting difference: although recurrent
keywords indicate a topic, repeated
IS-28-02-Cambria.indd 16 6/5/13 11:05 AM
MARCH/APRIL 2013 www.computer.org/intelligent 17
terms might not reect the overall
sentiment.
It’s possible to add other term-based
features to the features vector. Po-
sition refers to how a token’s posi-
tion in a text unit might affect the
text’s sentiment. Further, we might
consider presence n-grams—typically
bigrams and trigrams—to be useful
features. Some methods also rely on the
distance between terms. General tex-
tual analysis uses part of speech (POS)
information (for example, nouns, ad-
jectives, adverbs, and verbs) as a basic
form of word-sense disambiguation.
Certain adjectives are good indicators
of sentiment and guide feature selection
to classify the sentiment. Also, selected
phrases chosen by pre-specied POS
patterns, usually including an adjective
or adverb, help detect sentiments.
Some researchers have developed
other text mapping techniques that
assign labels to predened categories
or real numbers representing the de-
gree of polarity. These approaches
are strictly bound by domain and
topic. Moreover, most research on
sentiment analysis focuses on text
written in English and, consequently,
most of the resources developed (such
as sentiment lexicons and corpora)
are in English. Applying this research
to other languages is a domain adap-
tation problem.
From Heuristics
to Discourse Structure
In some unsupervised learning ap-
proaches, a sentiment lexicon is gen-
erated and later used to determine the
text unit’s degree of positivity or sub-
jectivity. Creating the sentiment lexi-
con through unsupervised polarity
or subjectivity labeling of words or
phrases is crucial.
1
The sentiment lexi-
con identies a term or a phrase’s prior
polarity or prior subjectivity, which in
turn helps identify contextual polarity
or subjectivity. Early works focused
mostly on linguistic heuristics. For ex-
ample, in their work on polarity clas-
sication, Vasileios Hatzivassiloglou
and Kathleen Mc Keown discuss how
two classes of interest represent oppo-
sites.
4
These opposite constraints help
the system with label decisions.
These approaches were unable
to detect novel expression of senti-
ment. Consequently, later work fo-
cused on propagating the valence of
seed words (for which the polarity is
known) to terms that co-occur with
them in general text (or in dictionary
glosses) or to synonyms and words
that co-occur with them in other
WordNet-dened relations. For ex-
ample, Ana-Maria Popescu and Oren
Etzioni proposed an iterative collec-
tive labeling algorithm.
5
This algo-
rithm starts with a global word label
computed over a large collection of
generic topic text. Gradually the al-
gorithm redenes the label with more
specicity: rst to a specic review
corpus, then specic to a product fea-
ture, and nally to a label specic to
the context in which the word occurs.
Benjamin Snyder and Regina Barzilay
similarly explored using discourse
information to infer relationships be-
tween product attributes.
6
They de-
signed a linear classifier that would
predict whether all aspects of a prod-
uct would be given the same rating.
Then they combined the prediction
with individual-aspect classifiers,
which would minimize loss function.
For opinionated documents, such
as product reviews, regression tech-
niques are often used to predict the
degree of positivity of opinions. Re-
gression techniques implicitly model
similar relationships between classes
that correspond to points on a scale,
such as the number of stars that a re-
viewer gives.
1
Modeling discourse
structure, such as twists and turns
in a document, leads to more effec-
tive sentiment labeling. In earlier
research, Bo Pang and Lillian Lee
attempted to partially address this
problem by incorporating location in-
formation into the feature set.
7
More recent studies emphasize the
importance of position in sentiment
summarization. For example, the in-
cipits of articles in topic-based sum-
marization usually indicate the text’s
sentiment. However, the last n sen-
tences of a product review often best
summarize the document’s overall
sentiment—almost as well as the n
(automatically computed) of most sub-
jective sentences.
7
Mahesh Joshi and
Carolyn Penstein-Rosé, for example,
explored how to use features based on
syntactic dependency relations to im-
prove opinion-mining performance.
8
They converted a transformation of
dependency-relation triples into com-
posite back-off features that general-
ize better than the regular, lexicon-
based, dependency-relation features.
From Coarse- to
Fine-Grained Analysis
We see opinion mining and senti-
ment analysis research evolving in
both technique sophistication and
analysis depth. Early on, Bo Pang and
her colleagues classied entire docu-
ments by overall positive or negative
polarity, and also by rating scores
of reviews.
9,10
These documents were
mainly supervised, manually labeled
samples, such as movie or product re-
views explicitly indicating an overall
positive or negative opinion.
Opinions and sentiments don’t oc-
cur only at the document level, nor
are they limited to a single valence or
target. One document might contain
positive and negative opinions to-
ward one or more topics. Hence, later
work adopted a segment-level opin-
ion analysis that used graph-based
techniques to distinguish sentimen-
tal from unsentimental sections. Pang
and Lee used segment-level opinion
IS-28-02-Cambria.indd 17 6/5/13 11:05 AM
18 www.computer.org/intelligent IEEE INTELLIGENT SYSTEMS
Knowledge-Based approaches to concept-level
sentiment analysis
analysis in their work to segment sec-
tions of a document by subjectiv-
ity. In another study, Peter Turney
classied items based on xed, syn-
tactic phrases used for expressing
opinions.
11
Finally, Jaap Kamps and
his colleagues classified items by
bootstrapping—using a small set of
seed opinion words and a knowledge
base such as WordNet.
12
In another work, Ellen Riloff and
Janyce Weibe reduced text-analysis
granularity to the sentence level by
using the presence of opinion-bearing
lexical items (single words or n-
grams) to detect subjective sen-
tences.
13
Soo-Min Kim and Eduard
Hovy, instead, used semantic frames
that identied sentimental topics (or
targets).
14
Reviewers tend to adhere
to being either subjective or objective,
and that creates continuity among
adjacent sentences. Hence, other re-
searchers collectively classify docu-
ments by assigning preferences for
pairs of nearby sentences.
10
Even sentence-level approaches of-
ten fail to discover sentiments about
an entity and/or its aspects. To cor-
rect that, other researchers adopted
an aspect-level approach, wherein an
opinion consists of targets and the
sentiments associated with them.
15–17
For example, the sentence “the new
iPhone 5’s screen size is amazing, but
its battery life is short” evaluates two
aspects (opinion targets): the screen
size and battery life of the same en-
tity. The sentiment about the iPhone
5’s screen size is positive, but the sen-
timent about its battery life is nega-
tive. Based on this level of analysis,
we can produce a structured opinion
summary about an entity and its as-
pects, and can draw more accurate
statistics about those aspects.
From Keywords to Concepts
We can study the evolution of senti-
ment analysis research by the analytical
tokens, or building blocks, and the
implicit information associated with
those tokens. We can group the
existing approaches into four main
categories: keyword spotting, lexi-
cal afnity, statistical methods, and
concept-based techniques.
Keyword spotting. Although the most
naïve approach, keyword spotting’s
accessibility and economy make it
popular. This approach classies text
by affect categories based on the pres-
ence of unambiguous affect words
such as happy, sad, afraid, and bored.
For example, Clark Elliott’s Affective
Reasoner watches for 198 affect key-
words (such as distressed or enraged),
affect intensity modiers (such as ex-
tremely, somewhat, or mildly), and a
handful of cue phrases (such as did
that and wanted to).
18
Other popular
sources of affect words are Andrew
Ortony and his colleagues’ Affec-
tive Lexicon,
19
which groups terms
into affective categories, and Janyce
Wiebe and her colleagues’ linguistic
annotation scheme.
20
Keyword spotting is weak in two
areas: it can’t reliably recognize affect-
negated words, and it relies on sur-
face features. Although keyword spot-
ting can correctly classify the sentence
“today was a happy day” as being af-
fectively positive, it is likely to assign
the same classication to a sentence
like “today wasn’t a happy day at all.”
Also, keyword spotting relies on the
presence of obvious affect words that
are only surface features of the prose.
Sometimes, a sentence conveys affect
through underlying meaning rather
than affect adjectives. For example,
the text “My husband just led for di-
vorce and he wants to take custody of
my children away from me” evokes
strong emotions, but uses no affect
keywords, and therefore is ineffec-
tive. Lexical afnity is slightly more
sophisticated than keyword spotting.
Lexical affinity. This approach not
only detects obvious affect words, it
also assigns arbitrary words a probable
“afnity” to particular emotions. For
example, lexical affinity might as-
sign the word “accident” a 75-percent
probability of indicating a negative
affect, as in “car accident” or “hurt
by accident.” This approach usu-
ally trains probability from linguistic
corpora.
21–23
Although it often out-
performs pure keyword spotting, there
are two main problems with this
approach. First, negated sentences
(I avoided an accident) and sentences
with other meanings (I met my girl-
friend by accident) trick lexical afn-
ity, because they operate solely on the
word level. Second, lexical affinity
probabilities are often biased toward
text of a particular genre, dictated by
the linguistic corpora’s source. This
makes it difficult to develop a re-
usable, domain-independent model.
Statistical methods. This approach,
which includes Bayesian inference
and support vector machines, is pop-
ular for affect text classication. Re-
searchers use statistical methods on
projects such as Pang’s movie review
classier and many others.
9,10,15,24
By
feeding a machine-learning algorithm
a large training corpus of affectively
annotated texts, the system might
not only learn the affective valence of
affect keywords (as in the keyword-
spotting approach), but also take into
account the valence of other arbitrary
keywords (similar to lexical afnity),
punctuation, and word co-occurrence
frequencies.
Generally, statistical methods are
semantically weak, which means that
individually—with the exception of
obvious affect keywords—a sta-
tistical model’s other lexical or co-
occurrence elements have little predic-
tive value. As a result, statistical text
classiers only work well when they
IS-28-02-Cambria.indd 18 6/5/13 11:05 AM
MARCH/APRIL 2013 www.computer.org/intelligent 19
receive sufficiently large text input.
So, while these methods might be
able to affectively classify a user’s text
on the page level or paragraph level,
they don’t work well on smaller text
units such as sentences or clauses.
Concept-based approaches. These
methods use Web ontologies or
semantic networks to accomplish se-
mantic text analysis.
25–27
This helps
the system grasp the conceptual and
affective information associated with
natural language opinions. By relying
on large semantic knowledge bases,
such approaches step away from
blindly using keywords and word co-
occurrence counts, and instead rely
on the implicit meaning/features as-
sociated with natural language con-
cepts. Superior to purely syntactical
techniques, concept-based approaches
can detect subtly expressed senti-
ments. Concept-based approaches
can analyze multi-word expressions
that don’t explicitly convey emotion,
but are related to concepts that do.
The concept-based approach relies
heavily on the depth and breadth of
the knowledge bases it uses. Without
a comprehensive resource that encom-
passes human knowledge, an opinion-
mining system will have difficulty
grasping the semantics of natural lan-
guage text. Moreover, the typicality
of knowledge bases—that is, the fact
that they contain only typical informa-
tion associated with concepts—limits
their capability to handle semantic
nuances. Their xed/at representa-
tion, nally, places bounds on infer-
ences of semantic and affective fea-
tures associated with concepts.
Multimodal
Sentiment Analysis
New sources of opinion mining and
sentiment analysis abound. Webcams
installed in smartphones, touchpads,
or other devices let users post opinions
in an audio or audiovisual format rather
than in text. For a rough idea of the
amount of material, consider that You-
Tube users upload two days’ worth of
video material to its website every min-
ute. Aside from converting spoken lan-
guage to written text for analysis, the
audiovisual format provides an oppor-
tunity to mine opinions and sentiment.
Many new areas might be useful in
opinion mining, such as facial expres-
sion, body movement, or a video blog-
ger’s choice of music or color lters.
Affect analysis, a related eld, ad-
dresses the use of linguistic, acous-
tic, and (potentially) video informa-
tion. This eld focuses on a broader
set of emotions or the estimation of
continuous emotion primitives; for
example, valence can be related to
sentiment. In one study, research-
ers provide recent surveys on spoken
and written-language-based analy-
sis; in another study, researchers ex-
plore further multimodal combina-
tions.
28,29
There’s almost no research
that focuses on multimodal sentiment
and opinion analysis. Stephan Raaij-
makers and his colleagues fuse acous-
tic and linguistic information, but
that information is based on the tran-
script of the spoken content rather
than on automatic speech recognition
output.
30
In addition to this research,
Louis-Philippe Morency and his col-
leagues combine acoustic, textual,
and video features to assess opinion
polarity in 47 YouTube videos.
31
They
demonstrate signicant improvement
in leave-one-video-out evaluation us-
ing Hidden Markov Models for clas-
sication. The authors identied po-
larized words, smiles, gazes, pauses,
and voice pitch as relevant features.
Again, the researchers relied on tran-
scripts to analyze the text and not the
actual spoken word.
Multimodal sentiment analysis hasn’t
been fully explored, but holds great
promise as an application. For example,
it might be extremely valuable when a
textual transcript is unavailable, and
we need a performance point of view
for synergy effects and fail-safeness.
In the latter respect, it will be particu-
larly interesting to see further modali-
ties involved—such as physiological
and brain signals, along with the use
of contextual knowledge. We’ll then
need to investigate analyses of robust-
ness against disturbances in individual
(or all) modalities alongside audio-
visual condence estimation.
Discussion
Gradually, sentiment analysis re-
search is distinguishing itself as a sep-
arate eld, falling between NLP and
natural language understanding. Un-
like standard syntactical NLP tasks,
such as summarization and auto-
categorization, opinion mining mainly
focuses on semantic inferences and
affective information associated with
natural language, and doesn’t require
a deep understanding of text. We en-
vision sentiment analysis research
moving toward content-, concept-,
and context-based analysis of natu-
ral language text, supported by time-
efcient parsing techniques suitable
for big social data analysis.
32
Collecting opinions on the Web will
still require processing at the content/
syntactic level, filtering out unopin-
ionated user-generated content (sub-
jectivity detection) and evaluating the
trustworthiness of the opinion and its
source. By contrast, concept/semantic
analysis infers semantic and affective
information associated with natural
language opinions, and hence, enables
a comparative fine-grained feature-
based sentiment analysis. Rather than
gathering isolated opinions about a
whole item, users generally prefer to
compare specic features of differ-
ent products (for example, the iPhone 5
versus the Galaxy S3 touchscreen)
or even sub-features (comparing the
IS-28-02-Cambria.indd 19 6/5/13 11:05 AM