You are on page 1of 38

Aspect-Based Sentiment Analysis

Aspect-Based Sentiment Analysis


Related terms:

Machine Learning, Neural Network Model, Word Embeddings, Conditional Ran-


dom Field, Deep Learning Method, Misinformation, Sentiment Analysis

View all Topics


an approach to describe what exactly is the intention of that text. So, we can say
that “I love that music” is a positive sentiment. So in an algorithmic way, we can say
that these sentiments are based on a set of polarities, i.e., positive, negative, or even
neutral. As such, one can quantify the positivity or negativity of the text.
an approach to describe what exactly is the intention of that text. So, we can say
that “I love that music” is a positive sentiment. So in an algorithmic way, we can say
that these sentiments are based on a set of polarities, i.e., positive, negative, or even
neutral. As such, one can quantify the positivity or negativity of the text.

To understand howToopinion
understand
mining how can
opinion
help, anmining
example canishelp,
given.
anInexample
a television
is given. In a television
manufacturing company,
manufacturing
supposecompany,
a businesssuppose
problema is business
discussed
problem
to analyze
is discussed
the to analyze the
sentiment from reviews.
sentiment
Theyfromwanted
reviews.
to understand
They wanted the to
contentment
understandofthe users
contentment
in of users in
product pricing andproduct
features.
pricing
A veryand
thorough
features.analysis
A very of
thorough
the reviews
analysis
was conducted
of the reviews was conducted
and the features most
and liked,
the features
featuresmostmostliked,
disliked,
features
viewsmost
on pricing,
disliked,and
views
segments
on pricing, and segments
of users with whichofpriorities
users with(forwhich
example,
priorities
certain(forusers
example,
were more
certainconcerned
users wereabout
more concerned about
the pricing than the
thequality,
pricingsomethanjust
the need
quality,
baresome
minimal
just need
features
bareinminimal
TV) were features in TV) were
identified. After this
identified.
study, theAfter
manufacturer
this study, was
the manufacturer
able to understandwas able
howtothey
understand
need how they need
to pitch what type to
of pitch
television
whattotypewhichof television
type of users.
to which
Also,type
the most
of users.
commonly
Also, the most commonly
features liked werefeatures
upgraded liked
permanently
were upgradedand thepermanently
most commonlyand thedisliked
most commonly
features disliked features
were removed. were removed.

There are differentThere


types are
of sentimental
different types
analysis.
of sentimental analysis.

• Aspect-based •sentimentAspect-based
analysis: A text
sentiment
analysis analysis:
techniqueA that
text analysis
divides the
technique that divides the
text data and defines itstext
sentiment
data andbased
defines
on its sentiment
aspects. It analyzes
based onconsumer
its aspects. It analyzes consume
feedback data by correlating
feedback
emotions
data bytocorrelating
different aspects
emotions
of atoproduct
differentor aspects of a product or
service. service.
• Fine-grained •sentimentFine-grained
analysis: Thissentiment
sentimentanalysis:
analysisThis
modelsentiment
contributes
analysis model contribute
to the advancement of polarity
to the advancement
accuracy. Sentiment
of polarity
analysis
accuracy.
can Sentiment
be done in analysis can be done in
the different categories the
discussed
different
in categories
the above section
discussed
based
in the
on above
polarity.
section based on polarity.
• Emotion detection:
• Emotion
Emotionanalysis
detection:
is the method
Emotionofanalysis
defining is and
the method
evaluating
of defining and evaluatin
the emotions conveyed thein textual
emotions data.
conveyed
Emotionindetection
textual data.
and Emotion
classification
detection and classification
are straightforward tasks
arethat
straightforward
can be completed tasks based
that canon be
thecompleted
emotions based
con- on the emotions con
veyed in the text, such as
veyed
fear,inrage,
the text,
happiness,
such assorrow,
fear, rage,
affection,
happiness,
motivation,
sorrow, affection, motivation
or neutral. or neutral.
• Intent analysis:
• The method
Intentofanalysis:
analyzingThetext
method
data toofdetermine
analyzingthe
textauthor's
data to determine the author's
intent is known as intent
intent
detection.
is known
Manyas intent
humandetection.
behaviorsMany
and actions
humanarebehaviors and actions are
motivated by motives, and
motivated
knowing byintentions
motives, andwillknowing
help youintentions
perceive these
will help you perceive these
behaviors. behaviors.

Some of the majorSome


hurdles
of the
for sentiment
major hurdles
analysis
for sentiment
are the following:
analysis are the following:

• sentences taken
• out of context,
sentences taken out of context,

• irony and sarcasm


• in sentences,
irony and sarcasm in sentences,

• sentences that• include asentences


negation,that
suchinclude
as “thisa negation,
will be his such
first and
as “this
last will be his first and last
masterpiece,” masterpiece,”
• sentences of ambiguity.
• sentences of ambiguity.
though recent work based on learning neural knowledge graph embeddings shows
an error reduction of more than 26% in semantic parsing of tweets [68]). Similar
issues arise when one is developing systems to detect stance, as in order to assess the
opinion of someone toward a given target, all relations between the entities involved
must be correctly identified and associated with the sentiment expressed. However,
a deeper linguistic analysis of text is also beneficial, if not necessary, for standard
message- or text-level sentiment analysis, as it helps to treat the issue of contextual
valence shifters mentioned in Section 3.1 by also accounting for word order and
sentence structure. To this end, the Natural Language Processing Group at Stanford
University developed a sentiment treebank.23 This treebank has been used to train a
recursive neural network built on top of grammatical structures [69], achieving an
increase of 5 percentage points on sentence polarity classification. On fine-grained
sentiment level they obtained a 9.7% improvement over a bag-of-words baseline,
and overall showed the ability to accurately capture the effects of negation and its
scope at various levels in the tree structures.
though recent work based on learning neural knowledge graph embeddings shows
an error reduction of more than 26% in semantic parsing of tweets [68]). Similar
issues arise when one is developing systems to detect stance, as in order to assess the
opinion of someone toward a given target, all relations between the entities involved
must be correctly identified and associated with the sentiment expressed. However,
a deeper linguistic analysis of text is also beneficial, if not necessary, for standard
message- or text-level sentiment analysis, as it helps to treat the issue of contextual
valence shifters mentioned in Section 3.1 by also accounting for word order and
sentence structure. To this end, the Natural Language Processing Group at Stanford
University developed a sentiment treebank.23 This treebank has been used to train a
recursive neural network built on top of grammatical structures [69], achieving an
increase of 5 percentage points on sentence polarity classification. On fine-grained
sentiment level they obtained a 9.7% improvement over a bag-of-words baseline,
and overall showed the ability to accurately capture the effects of negation and its
scope at various levels in the tree structures.
that the opinion opinion/1 is an AggregatedOpinion that includes two opinions
(opinion/2 and opinion/3) about the two aspect terms previously annotated.
that the opinion opinion/1 is an AggregatedOpinion that includes two opinions
(opinion/2 and opinion/3) about the two aspect terms previously annotated.

Listing 5. ExampleListing
of aspect
5. Example
term extraction
of aspect
andterm
aspect
extraction
term polarity
and aspect
detection
term polarity detection

Aspect-based sentiment
Aspect-based
analysis sentiment
identifies theanalysis
aspects
identifies
of a given
thetarget
aspectsentity
of a and
given
the
target entity and the
sentiment expressedsentiment
toward each
expressed
aspect.toward
Aspecteach
categories
aspect.(eg,
Aspect
food,
categories
price) identify
(eg, food, price) identify
coarser features than
coarser
aspectfeatures
terms,than
and they
aspect doterms,
not necessarily
and they do occur
notasnecessarily
terms in aoccur as terms in a
given sentence. In given
our example,
sentence.theInterms
our example,
fajitas and
thepico
terms
de fajitas
gallo refer
and to
pico
thedesame
gallo refer to the same
aspect category food,
aspect
while
category
the termfood,
service
whilerefers
the term
to theservice
aspectrefers
category
to the
service.
aspect category service.
In Listing 6 we annotate
In Listing
aspect
6 we
categories
annotateandaspect
introduce
categories
the and
use of
introduce
the property
the use of the property
marl:describesFeature
marl:describesFeature
for this purpose. for this purpose.
Listing 6. Annotation
Listing
of aspect
6. Annotation
categories
ofand
aspect
their
categories
polarity and their polarity
Bag of words: The most common statistical approach for sentiment analysis is to
treat the text content in an input document as a bag of n-grams without taking into
account the grammar and the n-grams order. The frequency of each term (n-gram)
occurring in the input is usually considered as a feature vector for the machine
learning model. Sentiment, emoticon, and emoji dictionary-based features can be
effectively combined with n-gram features and a hybrid machine learning model
can be built to learn underlying sentiment patterns. Since 2003, the majority of the
research work in sentiment analysis using statistical approaches mainly employs the
bag of words approach with notable work being done by Pang et al. [50].
Bag of words: The most common statistical approach for sentiment analysis is to
treat the text content in an input document as a bag of n-grams without taking into
account the grammar and the n-grams order. The frequency of each term (n-gram)
occurring in the input is usually considered as a feature vector for the machine
learning model. Sentiment, emoticon, and emoji dictionary-based features can be
effectively combined with n-gram features and a hybrid machine learning model
can be built to learn underlying sentiment patterns. Since 2003, the majority of the
research work in sentiment analysis using statistical approaches mainly employs the
bag of words approach with notable work being done by Pang et al. [50].

Aspect-based sentiment
Aspect-based
analysis:sentiment
In online analysis:
rating systems,
In onlinetherating
overallsystems,
rating ofthe overall rating of
an entity does not an
necessarily
entity does
explain
not necessarily
the quality explain
of the different
the qualitydimensions
of the different
of dimensions of
an entity. Consider,anfor
entity.
example,
Consider,
the restaurant
for example, dining
the domain.
restaurant Thedining
objective
domain.
of The objective of
a sentiment analysisa sentiment
system in analysis
this domainsystemis toinanalyze
this domain
the opinions
is to analyze
of people
the opinions of people
on different aspects onofdifferent
a restaurant
aspects
(e.g.,
of ambience,
a restaurant food
(e.g.,
quality,
ambience,
service
food
andquality,
price service and price
range). A typical restaurant
range). Areview
typicallooks
restaurant
like thereview
following
looks(Fig.
like the
7): following (Fig. 7):

Fig. 7. Example review


Fig. 7.
forExample
aspect-based
reviewsentiment
for aspect-based
analysis.sentiment analysis.

On analysis, one should


On analysis,
be ableone
to categorize
should be the
ablereview
to categorize
into multiple
the review
categories
into multiple categories
(Fig. 8): (Fig. 8):

Fig. 8. Example aspects


Fig. 8.and
Example
their corresponding
aspects and their
sentiments.
corresponding sentiments.

The challenge is toThe


trainchallenge
a model istotodotrain
this atype
modelof an
to interpretation.
do this type of The
an interpretation.
simplest The simplest
solution is to associate
solution
eachissentence
to associate
or aeach
phrasesentence
in a review
or a phrase
to someinaspect
a reviewandto some aspect and
the sentiment valuetheofsentiment
those aspects
valuecould
of those
be thenaspects
summedcouldup be to
then
findsummed
the overallup to find the overall
sentiment of the review.
sentimentFor this
of the
kindreview.
of anFor
approach
this kind
theoflist
anofapproach
aspects to thebelist of aspects to be
analyzed is usuallyanalyzed
predetermined.
is usually
There
predetermined.
has been some Thereresearch
has been done
someto treat
research done to treat
sentence labeling as
sentence
a separate
labeling
topic asmodeling
a separate task
topic
[51].modeling
Applying taskthe modified
[51]. Applying the modified
form of latent Dirichlet
form allocation
of latent Dirichlet
(LDA) called
allocation
local LDA,
(LDA)which
calledtreats
local each
LDA,sentence
which treats each sentence
as a document, sentences
as a document,
in an input
sentences
document in anare
input
labeled
document
to be associated
are labeled with
to be associated with
some topic. Sentiment
somemodels
topic. Sentiment
can then bemodels
appliedcanto then
learnbesentiments
applied tofrom learnthose
sentiments from those
sentences. In Ref. [51],
sentences.
a sentence
In Ref.
is assumed
[51], a sentence
to haveisaassumed
single aspect.
to haveIn aRef.
single
[52],aspect. In Ref. [52],
Pavlopoulos and Androutsopoulos
Pavlopoulos andproposedAndroutsopoulos
a methodproposed
to aggregate
a method
sentiments
to aggregate
for sentiments for
different aspects atdifferent
multipleaspects
granularities.
at multiple
Theirgranularities.
method also Theirhandlesmethod
cases alsowherehandles cases where
the same aspect is the
represented
same aspectusingis represented
different terms. using different terms.
Compositional models: The techniques discussed so far do not take into account
the way a sentiment progresses across individual components of a sentence, say,
phrases and even words. Recent work by Socher et al. explored the importance of
compositionality in sentiment analysis task [53]. In particular, a tree-bank was built
from a movie review dataset, where individual words, phrases, and sentences were
tagged with the sentiment conveyed. The input sentence/phrase is represented as a
parse tree as shown in Fig. 9.
Compositional models: The techniques discussed so far do not take into account
the way a sentiment progresses across individual components of a sentence, say,
phrases and even words. Recent work by Socher et al. explored the importance of
compositionality in sentiment analysis task [53]. In particular, a tree-bank was built
from a movie review dataset, where individual words, phrases, and sentences were
tagged with the sentiment conveyed. The input sentence/phrase is represented as a
parse tree as shown in Fig. 9.

Fig. 9. Recursive vector


Fig. 9.computation
Recursive vector
usingcomputation
compositionality
usingfunction
compositionality
[53]. function [53].

Their main focus was


Their
to main
computefocus
compositional
was to compute vector
compositional
representation
vector
for representation
phrases for phrases
and feed those vectors
and feed
into athose
classifier
vectors
to make
into a the
classifier
classifier
to make
learn the
propagation
classifier learn
of propagation of
sentiments from granular
sentiments
wordfrom
levels
granular
to phrase
word
andlevels
sentence
to phrase
levels.and
Socher
sentence
et al.levels.
[53] Socher et al. [53]
employed recursiveemployed
neural networks
recursive[41,54–57]
neural networks
to compute
[41,54–57]
compositional
to compute
vector
compositional vector
representation of phrases.
representation of phrases.

A compositional function
A compositional
takes the vector
functionrepresentation
takes the vector
of child
representation
nodes at eachof child
levelnodes at each level
and computes the and
vector
computes
representation
the vector
of arepresentation
current node in of aa bottom-up
current node fashion.
in a bottom-up fashion.
The function uses the
The vector
function
representation
uses the vector
of arepresentation
node as a featureof aof
node
a sentiment
as a feature of a sentiment
classifier to compute
classifier
the sentiment
to compute
valuethe
at sentiment
the node level.
valueSocher
at the et
node
al. [53]
level.
compute
Socher et al. [53] compute
sentiment values with
sentiment
the softmax
valuesclassifier
with the[58,59],
softmaxasclassifier
shown in [58,59],
Fig. 10.as shown in Fig. 10.
Fig. 10. SentimentFig.
value10.computation
Sentiment value
at each
computation
node [53]. at each node [53].
by identifying the most similar neighbors within the cluster of a specific user. To
implement this method, they carried out two major operation. First, they extracted
the user's preferences for the given items based on multicriteria ratings. Second, on
the basis of the preferences of the user the cluster centers were defined.
Akhtar et al. [9] presented a technique for analyzing hotel reviews and extracted some
valuable information and/or knowledge from them to assist the service providers as
well as to help customers identify the loopholes and strengths in the service sector
to improve their business performance.
Akhtar et al. [9] presented a technique for analyzing hotel reviews and extracted some
valuable information and/or knowledge from them to assist the service providers as
well as to help customers identify the loopholes and strengths in the service sector
to improve their business performance.

Yang et al. [10] presented


Yang et al.
a technique
[10] presented
consisting
a technique
of threeconsisting
main components,
of three main
namely
components, namely
aspect weight, opinion
aspect
mining,
weight,andopinion
overallmining,
rating inference.
and overallThe
rating
opinion
inference.
miningThe opinion mining
component was responsible
component forwas
extracting
responsible
only for
theextracting
key aspectsonly
andthe
opinions
key aspects
fromand opinions from
the user's reviews based
the user's
on which
reviewsit computed
based on which
a rating
it computed
for each extracted
a rating aspect.
for each extracted aspect.

Wang et al. [11] proposed


Wang et al.
an approach
[11] proposed
on solving
an approach
a problem
on solving
when a auser problem
is par-when a user is par-
ticularly new to an ticularly
environment.
new toThis
an environment.
problem is known
This problem
as a cold is start
known
problem.
as a cold start problem.
We discuss about theWe cold
discuss
startabout
problem
the cold
laterstart
in this
problem
chapter.
later
Mostin this
recommender
chapter. Most recommender
systems collect thesystems
preferences
collect
of the
users
preferences
on some attributes
of users onof some
the items.
attributes of the items.

Musat et al. [12] explained


Musat et al.
a method
[12] explained
called topic
a method
profilecalled
collaborative
topic profile
filtering
collaborative filtering
(TPCF) that solved (TPCF)
the problems
that solved
occurring
the problems
due to the
occurring
data sparsity
due toproblems
the data and
sparsity problems and
nonpersonalized ranking
nonpersonalized
methods that
ranking
led to
methods
difficultythat
in finding
led to difficulty
sufficient in finding sufficient
reliable data for making
reliable
recommendations.
data for making recommendations.

Jamroonsilp and Prompoon


Jamroonsilp [13]and
presented
Prompoon an approach
[13] presented
for ranking
an approach
items based
for ranking items based
on user's reviews. They
on user's
considered
reviews.five They
predefined
considered aspects
five predefined
for the software
aspects
items.
for the software items.
The ranking of software
The ranking
was computed
of softwareby comparing
was computed the sentences
by comparing
analyzing
the sentences
the analyzing the
different clients' ratings
different
for every
clients'
software
ratingsaspect.
for every
This
software
was performed
aspect. This
in three
was performed
phases in three phases
including gathering including
user reviews,
gathering
analyzing
user reviews,
the gathered
analyzing
reviews,
the and
gathered
carrying
reviews,
out and carrying out
the subsequent software
the subsequent
ranking. software ranking.

Zhang et al. [14] proposed


Zhang et al.
a method
[14] proposed
that utilized
a method
the aspect-level
that utilized
sentiment
the aspect-level
of users'sentiment of users'
reviews with the support
reviewsofwith
helpfulness
the support
reviews.
of helpfulness reviews.

Zheng, Shekhar, Jose,


Zheng,
andShekhar,
Rai [16] proposed
Jose, and aRai
multicriteria
[16] proposed
decision-making
a multicriteriaap-decision-making ap-
proach in the discipline
proachofineducational
the discipline
learning.
of educational
First theylearning.
integrated
First
thethey
con-integrated the con-
text-awareness andtext-awareness
multicriteria decision-making
and multicriteriaindecision-making
the recommender in the
systems
recommender systems
considering the educational
consideringdata
theaseducational
a case study.
data as a case study.

These are some of These


the work are carried
some ofout thebywork
various
carried
scientists
out byaround
variousthe
scientists
globe. around the globe.
There are thousands Thereof projects
are thousands
completedof projects
or on-going
completed
in theor
field
on-going
of MCRSsin the
with
field of MCRSs with
the aim of makingthe theaimsystem
of making
fully efficient.
the system
Leading
fully companies
efficient. Leading
are nowcompanies
using are now using
recommender systems,recommender
such as Amazon
systems,which
such uses
as Amazon
recommender
which uses
systems
recommender
to give systems to give
accurate recommendations
accurate recommendations
to its customers. Netflix
to its customers.
also uses anNetflix
MCRSalsoto provide
uses an MCRS to provide
a list of movies anda web
list ofseries
moviesto users
and web
on the
series
basis
to users
of userondetails
the basis
andofprevious
user details and previous
choices. These newchoices.
techniquesThese arenew
being
techniques
applied increasingly
are being applied
in recommender
increasingly in recommender
systems to achievesystems
maximum to achieve
accuracy.maximum accuracy.
Deep Learning for Aspect-Based Senti-
ment Analysis: A Comparative Review
Deep Learning for Aspect-Based Senti-
ment Analysis: A Comparative Review
Hai Ha Do, ... Abeer
HaiAlsadoon,
Ha Do, ...inAbeer
ExpertAlsadoon,
Systems with
in Expert
Applications,
Systems 2019
with Applications, 2019

Appendix: ListAppendix:
of Abbreviations
List of Abbreviations
ABSA: Aspect-Based
ABSA:
Sentiment
Aspect-Based
AnalysisSentiment Analysis
ACD: Aspect Category
ACD:Detection
Aspect Category Detection
CNN: Convolutional
CNN:
Neural
Convolutional
Networks Neural Networks
CRF: Conditional Random
CRF: Conditional
Field Random Field
DL: Deep LearningDL: Deep Learning
DNN: Deep NeuralDNN:
Networks
Deep Neural Networks
GRU: Gated Recurrent
GRU:Unit
Gated Recurrent Unit
LSTM: Long-ShortLSTM:
Term Memory
Long-Short Term Memory
ME: Maximum Entropy
ME: Maximum Entropy
MemNet: MemoryMemNet:
NetworksMemory Networks
NER: Name EntityNER:
Recognition
Name Entity Recognition
NLP: Natural Language
NLP: Natural
Processing
Language Processing
OTE: Opinion Target
OTE:Extraction
Opinion Target Extraction
POS: Part Of Speech
POS: Part Of Speech
RAE: Recursive Auto-encoders
RAE: Recursive Auto-encoders
RecNN: Recursive RecNN:
Neural Networks
Recursive Neural Networks
RNN: Recurrent Neural
RNN: Networks
Recurrent Neural Networks
SemEval: International
SemEval:
Workshop
International
on Semantic
Workshop
Evaluation
on Semantic Evaluation
SP: Sentiment Polarity
SP: Sentiment Polarity
SVM: Support Vector
SVM:
Machine
Support Vector Machine
WE: Word Embeddings
WE: Word Embeddings

> Read full chapter


> Read full chapter

Comprehensive
Comprehensive
analysis of
analysis
aspectof
term
aspect term
extraction
extraction
methodsmethods
using various
usingtext
various text
embeddings
embeddings
Łukasz Augustyniak,
Łukasz
... Przemysław
Augustyniak,
Kazienko,
... Przemysław
in Computer
Kazienko,
Speech
in Computer
& Language,
Speech & Language,
2021 2021
Abstract
Abstract
Recently, a variety of
Recently,
model a variety
designs andof model
methodsdesigns
haveandblossomed
methodsinhave
the context
blossomed in the context
of the sentiment analysis
of the sentiment
domain. However,
analysis domain.
there is still
However,
a lack of
there
comprehensive
is still a lack of comprehensive
studies of Aspect-based
studiesSentiment
of Aspect-based
Analysis.Sentiment
We want to Analysis.
fill thisWe
gapwant
and to
propose
fill this gap and propose
a comparison witha comparison
ablation analysis
withofablation
Aspect Term
analysis
Extraction
of Aspect using
Termvarious
Extraction
text using various text
embeddings methods. embeddings
We particularly
methods. focused
We particularly
on simple focused
architectures
on simple
basedarchitectures
on based on
long short-term memory
long short-term
(LSTM) withmemory
optional
(LSTM)
conditional
with optional
random conditional
field (CRF)random field (CRF)
enhancement using enhancement
different pre-trained
using different
word embeddings.
pre-trained wordMoreover,
embeddings.
we analyzed
Moreover, we analyzed
the influence on the theperformance
influence onofthe extending
performance
the word
of extending
vectorization
the word
step with
vectorization step with
character-based wordcharacter-based
embeddings.word The experimental
embeddings. results
The experimental
on SemEvalresults
datasetson SemEval datasets
revealed that bi-directional
revealed that
longbi-directional
short-term memory
long short-term
(BiLSTM) memory
could be(BiLSTM)
used as could be used as
a very good predictor,
a very
even
good
comparing
predictor,toeven
very comparing
sophisticated to very
and complex
sophisticated
models and complex models
using huge word embeddings
using huge word
or language
embeddings
models.or We
language
presented
models.
a comprehensive
We presented a comprehensive
analysis of various analysis
customizations
of variousof LSTM-based
customizations architecture
of LSTM-basedand word/character
architecture and word/character
embeddings that could
embeddings
be usedthat
as a could
guideline
be used
to choose
as a guideline
the best to
model
choose
version
the best
for model version for
particular user needs.
particular user needs.
1
Opinion
1 mining 2008 2487
Opinion mining276 2008
and sentiment and sentiment
analysis [27] analysis [27]

2 Mining
2 and 2004 1400
Mining and 108 2004
summarizing summarizing
customer customer
reviews [48] reviews [48]
3 Lexicon-based
3 2011 471Lexicon-based 79 2011
methods for methods for
sentiment sentiment
analysis [53] analysis [53]

4 Recognizing
4 2005 657Recognizing 55 2005
contextual contextual
polarity in polarity in
phrase-level phrase-level
sentiment sentiment
analysis [ 51] analysis [ 51]
5 Techniques
5 and 2013 203Techniques and 51 2013
applications for applications for
sentiment analy- sentiment analy-
sis: The main sis: The main
applications and applications and
challenges of one challenges of one
of the hottest re- of the hottest re-
search areas in search areas in
computer sci- computer sci-
ence [54] ence [54]

6 New
6 avenues in 2013 202New avenues in 51 2013
opinion mining opinion mining
and sentiment and sentiment
analysis [55] analysis [55]

7 Measuring
7 praise 2003 694Measuring praise50 2003
and criticism: In- and criticism: In-
ference of se- ference of se-
mantic orienta- mantic orienta-
tion from associa- tion from associa-
tion [43] tion [43]

8 Sentiment
8 2012 228Sentiment 46 2012
strength strength
detection for the detection for the
social web [56] social web [56]

9 Mining
9 the peanut 2003 617Mining the peanut44 2003
gallery: Opinion gallery: Opinion
extraction and se- extraction and se-
mantic classificat- mantic classificat-
ion of product re- ion of product re-
views [3] views [3]

10 Extracting
10 prod- 2005 517Extracting prod-
43 2005
uct features and uct features and
opinions from opinions from
reviews [ 57] reviews [ 57]
11
lexical resource lexical resource
for opinion for opinion
mining [63] mining [63]

9 SentiWordNet
9 2010 1304
SentiWordNet 186 2010
3.0: An Enhanced 3.0: An Enhanced
Lexical Resource Lexical Resource
for Sentiment for Sentiment
Analysis and Analysis and
Opinion Opinion
Mining [64] Mining [64]

10 From
10 tweets to 2010 1256
From tweets to 179 2010
polls: Linking text polls: Linking text
sentiment to pub- sentiment to pub-
lic opinion time lic opinion time
series [65] series [65]

11 Recognizing
11 2005 Recognizing
2141 178 2005
contextual contextual
polarity in polarity in
phrase-level phrase-level
sentiment sentiment
analysis [ 51] analysis [ 51]
11 Extracting
11 prod- 2007 Extracting prod-
1775 178 2007
uct features and uct features and
opinions from opinions from
reviews [ 57] reviews [ 57]
One of the top cited literature reviews is a book by Bing Liu [60]. The 167 pages
contain a wide array of topics, with chapters about document, sentence and as-
pect-based sentiment analysis. Overall the topic is approached first by introducing
the research problems of sentiment analysis and then answering them with the latest
knowledge available during the writing of the book.
orientations of the phrases it consists of, using the PMI score. The approach is
evaluated with reviews from different domains, i.e. automobile, bank, movie and
travel reviews. Movie reviews prove to be particularly challenging for the approach,
as a review of a recommendable movie can contain negative adjectives describing
incidents in the movie, e.g. violence and horror.
orientations of the phrases it consists of, using the PMI score. The approach is
evaluated with reviews from different domains, i.e. automobile, bank, movie and
travel reviews. Movie reviews prove to be particularly challenging for the approach,
as a review of a recommendable movie can contain negative adjectives describing
incidents in the movie, e.g. violence and horror.

Turney and Littman [43]


Turneyevaluate
and Littman [43]
two strategies
evaluate
for measuring
two strategies
semantic
for measuring
orientation
semantic orientation
from semantic association,
from semantic
i.e. statistically
association,
taking
i.e. statistically
into accounttaking
the context
into account
when the context when
evaluating semanticevaluating
orientation.
semantic
These approaches,
orientation. These
PMI andapproaches,
latent semantic
PMI and
analysis
latent semantic analysis
(LSA) are tested with
(LSA)
twoare
different
tested corpora,
with two with
different
LSA approach
corpora, with
beingLSA
more
approach
accurate
being more accurate
in classifying semantic
in classifying
orientation.
semantic orientation.

3.6.3 Twitter 3.6.3 Twitter


As highlighted in Fig.
As highlighted
8 (b) Twitterin
and
Fig.social
8 (b) media
Twitterhave
and gained
social media
popularity
have in
gained popularity in
sentiment analysissentiment
in recent years.
analysis
Three
in recent
top-cited
years.
papers
Threefocused
top-cited
on papers
working focused
with on working with
Twitter data. Twitter data.

Pak and Paroubek [61]


Pak and
present
Paroubek [61]
a method forpresent
automatic
a method
collecting
for automatic
of a corpus
collecting
from of a corpus from
microblogs and usemicroblogs
it to build and
a sentiment
use it to classifier.
build a sentiment
In this instance,
classifier.
theIncorpus
this instance, the corpus
is gathered from Twitter.
is gathered
The authors
from Twitter.
claim The
that authors
the approach
claim can
thatbe
theadapted
approachto can be adapted to
multiple languages,
multiple
but in their
languages,
work, itbut
is only
in their
usedwork,
withitthe
is only
English
usedlanguage.
with the English language.

Tumasjan et al. [62]Tumasjan
examine et al. [62]
around 100,000
examinetweets
aroundfrom
100,000
Twittertweets
as a predictor
from Twitter as a predictor
for election results.for
They
election
find out
results.
that They
the proportion
find out that
of mentions
the proportion
of parties
of mentions
and of parties and
prominent politicians
prominent
mirror the
politicians
electionmirror
resultsthe
quite
election
closely,
results
however
quitehaving
closely, however having
less accuracy than less
opinion
accuracy
polls.than
Another
opinion
interesting
polls. Another
findinginteresting
was that 40%finding
of the
was that 40% of the
messages were postedmessages
by thewere
4% of posted
the users.
by the 4% of the users.

Data from Twitter has


Dataalso
from
been
Twitter
usedhas
to gauge
also been
public
used
opinion
to gauge
in time
public
series
opinion
by in time series by
O’Connor, Balasubramayan,
O’Connor,Routledge
Balasubramayan,
and Smith [65].
RoutledgeWhile
andthe
Smith [65].
correlation
While
be-the correlation be-
tween sentiment measurement
tween sentimentand measurement
poll data variesand
across
poll different
data varies
data
across
sets,different
the data sets, the
broad trends are captured
broad trends
by theare
sentiment
capturedanalysis.
by the sentiment analysis.

3.6.4 Tools and lexicons


3.6.4 Tools and lexicons
Three different sentiment
Three different
analysissentiment
tools and lexicons
analysis made
tools and
it tolexicons
the top-20
madecited
it to
list.
the top-20 cited list.
Such works make itSuch
easier
works
to take
make
sentiment
it easier analysis
to take sentiment
into to use.analysis into to use.

Taboada et al. [53] Taboada


enhance et al. [53]
the existing
enhance
analysis
thetool
existing
“Semantic
analysis
Orientation
tool “Semantic
CALcu-Orientation CALcu-
lator (SO-CAL)” by lator
supplementing
(SO-CAL)” by rules
supplementing
such as intensification
rules suchand
as intensification
negation to a and negation to a
pre-existing lexiconpre-existing
based method.
lexicon
Thisbased
new method.
approachThisis then
newvalidated,
approachevaluated
is then validated, evaluated
and compared withand several
compared
existingwith
sentiment
several existing
analysissentiment
tools and approaches,
analysis toolsinand
an approaches, in an
experiment where experiment
various datawhere
sets are
various
analyzed
datawith
setsall
arethe
analyzed
approaches
with all
to the
drawapproaches to draw
conclusions. The authors used several data sets to evaluate the tool: product reviews,
social media comments, news articles and headlines.
conclusions. The authors used several data sets to evaluate the tool: product reviews,
social media comments, news articles and headlines.

In 2012, Thelwall, Buckley


In 2012,andThelwall,
Paltoglou [56]
Buckley andevaluate
Paltoglou [56]
a tool namedevaluate
“SentiStrength
a tool named “SentiStrength
2”, which is a lexicon
2”, based
which classifier
is a lexicon
usedbased
together
classifier
withused
various
together
machinewithlearning
various machine learning
algorithms. This tool
algorithms.
is evaluated
Thiswith
toola is
variety
evaluated
of data
with
sets
a variety
that feature
of data
shorts
sets that
texts,feature shorts texts,
e.g., YouTube comments
e.g., YouTube
and tweets.
comments
The tool andperforms
tweets. The
well tool
analyzing
performs
shortwell
texts
analyzing short texts
found on the socialfound
web andon the
performs
social web
fairlyand
wellperforms
context independent.
fairly well context independent.

SentiWordNet is anSentiWordNet
automaticallyisgenerated
an automatically
lexicon developed
generated by
lexicon
Baccianella,
developedEsuli
by Baccianella, Esuli
and Sebastiani [63,64]
and to
Sebastiani [63,64]
be specifically used
to beinspecifically
sentiment used
analysis.
in sentiment
The lexicon
analysis.
was The lexicon was
generated using a generated
semi-supervised
using alearning
semi-supervised
algorithmlearning
and evaluated
algorithm
by comparing
and evaluated by comparing
the generated semantic
the generated
values ofsemantic
words withvalues
the manually
of words with
annotated
the manually
sets of syn-
annotated sets of syn-
onyms. onyms.

3.6.5 Others 3.6.5 Others


Finally, we had twoFinally,
paperswethat
had
didtwo
notpapers
fit intothat
ourdid
existing
not fitcategories
into our existing
but bothcategories
of but both of
them addressed important
them addressed
topics, i.e., economic
important topics,
impacts
i.e., economic
and the domain
impacts depen-
and the domain depen-
dency of sentimentdency
analysis.
of sentiment analysis.

Widely cited paperWidely


by Ghose
cited
and
paper
Ipeirotis [58]
by Ghoseexamined
and Ipeirotis [58]
the relation
examined
of product
the relation of product
reviews to economicreviews
outcomes
to economic
and implemented
outcomes anda random
implemented
forest based
a random
classifier
forest based classifier
to predict the impact
to predict
of reviews.
the Additionally,
impact of reviews.
they found
Additionally,
that subjectivity
they foundof that
a review
subjectivity of a review
is associated with an
is associated
increased impact
with anofincreased
a review,impact
while spelling
of a review,
mistakes
whiledecrease
spelling mistakes decrease
the economic impacttheofeconomic
some reviews.
impact of some reviews.

Work by Blitzer, Dredze


Work and
by Blitzer,
Pereira [59]
Dredze
concentrates
and Pereira [59]
on theconcentrates
problem of on
context
the problem of context
dependent nature dependent
of sentiment nature
expression.
of sentiment
The authors
expression.
extendThe
a structural
authors extend
corre- a structural corre-
spondence learningspondence
algorithm,learning
which can
algorithm,
adapt between
which candomains.
adapt between
Additionally,
domains.
the Additionally, the
authors introduce authors
a measureintroduce
of domain
a measure
similarity,
of domain
to gaugesimilarity,
the transferability
to gauge of
the transferability of
classifiers from domain
classifiers
to domain.
from domain to domain.

You might also like