Sentiment Analysis Approaches based on Granularity Levels
Benaissa Azzeddine Rachid, Harbaoui Azza and Ben Ghezala Henda
University of Manouba, RIADI Laboratory, ENSI School, La Manouba, Tunisia
Keywords:
Opinion Mining, Sentiment Analysis, Machine Learning, Lexicon based, Ontology based, Granularity Level.
Abstract:
The evolution of web 2.0 has enabled the emergence of social media where users can post, share and discuss
their opinions about products, events, peoples and organizations. This increase of the user generated content
(UGC) has allowed the publication of several works during the last decade in the scientific community working
on sentiment analysis. Sentiment analysis, also known as opinion mining is the field of extraction and analysis
of opinions, feelings and attitudes of users on the web. In this paper, we provide an overview of the field of
sentiment analysis by discussing the workflow of mining opinions in different granularity levels and covering
common and recent approaches and techniques used to solve tasks related to sentiment analysis process at
every level.
1 INTRODUCTION
Knowing what people think has always been a very
important information to make a decision. For this
reason we often seek out the opinions of others. A
few years ago, when a person needed opinions, he /
she asked family and friends. Even organizations had
to conduct surveys, opinion polls and focus groups to
collect public or consumer opinions. Those days are
gone. At the present time, people express their opini-
ons on social media platforms like Twitter, Facebook,
and others and e-commerce sites like Amazon. Col-
lection and analysis of this huge volume of opiniona-
ted data are thus needed. Sentiment analysis (SA) is
the field specialized in such tasks.
Sentiment analysis, also called opinion mining, is
the field of study that analyzes peoples opinions, sen-
timents, evaluations, appraisals, attitudes, and emoti-
ons towards entities such as products, services, orga-
nizations, individuals, issues, events, topics, and their
attributes as mentioned in (Liu, 2012).
This domain is also known in the literature as opinion
mining, sentiment mining, opinion extraction, sub-
jectivity analysis, etc. However, in this paper, we will
limit ourselves to the use of sentiment analysis or opi-
nion mining as they represent the most used keywords
in journal publications and in conference proceedings
based on (Ahlgren, 2016). In the last decade senti-
ment analysis has gained popularity. (Mantyla et al.,
2018) mentioned that the number of papers published
in the field of sentiment analysis is 6996 papers. This
incredible increase makes opinion mining one of the
active and growing search areas.
In this paper, we will cite recent researche techni-
ques used in sentiment analysis based on granularity
level (document, sentence and aspect). The rest of
the paper is organized as follows: section 2 introdu-
ces some of the papers that defined sentiment analy-
sis based on granularity level. In section 3 Coarse-
grained-level (document and sentence) is highlighted.
The aspect-level is covered in section 4. Other levels
of SA are discussed in section 5. In Section 6 a com-
parison of the approaches of SA is being dealt with,
and finally Section 7 concludes the paper.
2 GRANULARITY BASED
SENTIMENT ANALYSIS
Sentiment Analysis is closely related to the field of
Natural Language Processing (Sun et al., 2017), it is a
big suitcase of NLP problems (Cambria et al., 2017).
It is also studied in Information Retrieval and Data
Mining.(Hemmatian and Karim Sohrabi, 2017) and
(Bhatia et al., 2018) consider opinion mining as a sub-
field of the Web Content Mining process in the field
of Web Mining.
Some papers that dealt with the tasks of opinion
mining in a granularity level manner are presented
above.
(Missen et al., 2012) reviewed the field of opinion
mining from word to document level in a very detailed
manner. They also highlighted the importance of so-
324
Rachid, B., Azza, H. and Henda, B.
Sentiment Analysis Approaches based on Granularity Levels.
DOI: 10.5220/0007187603240331
In Proceedings of the 14th International Conference on Web Information Systems and Technologies (WEBIST 2018), pages 324-331
ISBN: 978-989-758-324-7
Copyright © 2018 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved

cial networks for opinion mining tasks.
(Feldman, 2013) presented a general architecture
for sentiment analysis systems. The input of the sy-
stem is a corpus that will be converted to text and pre-
processed using linguistic tools. The resulted text will
be annotated by a Document Analysis Module. The
annotation may be attributed to whole documents, to
sentences or to fine grained entities (aspects).
(Liu, 2012) explored in his review/book the pro-
blems and the objectives of sentiment analysis. He
used previous definition defining the opinion as a
quintuple (ei, aij, sijkl, hk, tl ), where ei is the name
of an entity, aij is an aspect of ei, sijkl is the sentiment
on aspect aij of entity ei, hk is the opinion holder, and
tl is the time when the opinion is expressed by hk. Ba-
sed on this definition, the objective of opinion mining
is defined as determining all the quintuples of given
texts.
As mentioned earlier, the next section will talk
about sentiment analysis tasks based on granularity
level. We first begin by coarse-grained level (Docu-
ments and sentences) and go deep in the hierarchy as
the complexity increases to fine grained level (aspect).
3 COARSE-GRAINED LEVEL
SENTIMENT ANALYSIS
3.1 Document-level Sentiment Analysis
In (Ravi and Ravi, 2015) 159 articles were distribu-
ted based on the granularity of sentiment analysis, 73
articles appeared in the document level which makes
it the most studied topic in the field. Document-level
sentiment classification, as known in the literature, is
considered the simplest sentiment analysis task. The
task of opinion mining at this level is to identify opi-
nionated documents and classify them according to
their polarities.
Authors in (Missen et al., 2012) mentioned that
most of the researchers at this level follow a two-
step approach: Topic Relevance Retrieval and Opi-
nion Finding step.
The document is considered as a basic information
unit which includes multiple sentences. Based on the
quintuple introduced in the first section, the task of
opinion mining is to determine the overall sentiment
of the opinion holder about the entity described in the
document. This approach helps the users in decision
making by providing a summary of total number of
positive and negative documents.
3.2 Sentence-level Sentiment Analysis
Just like document level opinion mining, sentence le-
vel opinion mining is also a classification problem.
Sentences are regarded as short documents which ma-
kes the classification the same for both levels. Most
of the researches at document level dont perform a
three class classification (positive, negative, and neu-
tral). However, at the sentence level, the neutral class
cannot be ignored because sentences may express no
opinion or sentiment. Thus, the purpose at this level
is to classify each sentence in an opinion document as
positive, negative or neutral opinion or sentiment.
Sentiment sentence classification is generally per-
formed in two classes of classification problem. The
first determines whether the sentence is expressing an
opinion (sentiment) or not and the second classify the
sentences as positive, negative or neutral. The first
step in the process is known in the literature as Sub-
jectivity Classification. It aims to distinguish opini-
ons( subjective sentences) from facts (objective sen-
tences) (Chaturvedi et al., 2018). Subjective senten-
ces can express some personal feelings, views judg-
ments or beliefs(Liu, 2015) that might vary from per-
son to person, whereas, objective sentences express
factual information which remains valid for all indi-
viduals. Because of that some researchers prefer to
classify sentences as opinionated or non-opinionated.
The second step is called sentence sentiment
classification. After classifying the sentences as
being subjective (opinionated) or objective (non-
opinionated), sentence sentiment classification aims
to classify them as positive, negative or neutral. An
assumption that generally researchers make at this le-
vel of analysis is that a sentence expresses a single
sentiment. Thus, sentences that express more than
one sentiment are treated differently. More complex
sentences (interrogative, comparatives, conditionale
and sarcastic sentences) also need advanced techni-
ques.
3.3 Machine Learning Approaches
3.3.1 Supervised Learning
Supervised learning supposes that there are multiple
classes to which a document can be classified. The
process of learning is carried out using the data of trai-
ning available for each class. The training set is used
by a classifier to learn the different characteristics of
documents. Learning task is done by using classi-
fication algorithms either probabilistic (naive Bayes,
Bayesian Neutral Network, Maximum entropy) or
non-probabilistic (Support Vector Machine, Artifi-
Sentiment Analysis Approaches based on Granularity Levels
325

cial Neural Network, K-nearest Neighbor, Rule Ba-
sed, Decision Tree) (Hemmatian and Karim Sohrabi,
2017).The performances of the classifier are validated
using a test data. At the end of this process, every do-
cument should be tagged with its appropriate category
(class).
Like most supervised learning approaches, feature
engineering is the key to build a good sentiment ana-
lysis classifier. The most common used features are
N-gram (Terms and their frequency), syntactic featu-
res (Part Of Speech, Syntactic Dependency) and se-
mantic features (Sentiment words and phrases, senti-
ment shifters).
3.3.2 Unsupervised Learning
Unlike supervised learning, that considers the target
value (label), unsupervised learning process does not
provide any label data. Unsupervised classification
belongs to semantic orientation approach. It aims
to determine the semantic orientation of the phra-
ses within the document. The algorithm described
by (Turney, 2002) is totally unsupervised. He used
syntactic pattern as a sequence of Part Of Speech
tags. The algorithm consists of three steps. First,
two consecutive words are extracted if their POS tags
are conform to certain constraints. Then, it estimates
the polarity of adjectives and adverbs present in opi-
nion review by calculating their proximity using the
pointwise mutual information (PMI) method. PMI
(P,W) measures the statistical dependence between
the phrase P and the word W based on their co-
occurrence in a given corpus or over the Web (Feld-
man, 2013). Turney used two words for his approach
excellent and poor. Finally, the overall polarity of the
review is then deduced by aggregating the polarity of
the adjectives and adverbs that compose it, and the
review is classified as positive or negative.
In the last few years, deep learning has gained po-
pularity in many fields and had shown valuable re-
sults. It is a powerful machine learning technique as
mentionned in the recent survey (Zhang et al., 2018).
Sentiment analysis is one of the fields that recently
has been attracted to deep learning techniques. It has
been shown that document and sentence representati-
ons can be very useful for SA tasks. For that purpose,
deep learning techniques such as word embeddings,
Long Short Term memory, reccurent, reccursive, con-
volutional neural networks were applied to sentiment
analysis classification.
3.4 Lexicon based Approaches
Another approach to do sentiment analysis in docu-
ment level, which can be seen as an unsupervised ap-
proach (Liu, 2012), is the lexicon-based approach. It
consists of using a collection of known and precompi-
led sentiment terms tagged with their semantic orien-
tation called sentiment lexicon (polarity or opinion
lexicon). These terms are used to express the positive
or negative feelings. The terms that make the lexicon
are generally adjectives and adverbs, but names and
verbs should also be considered.
The aim of using such a lexicon is to determine the
overall sentiment of a given text based on the assump-
tion that the collective polarity of a sentence or docu-
ments is the sum of polarities of the individual phrases
or words. The document is classified as positive if the
sum is positive, negative if the sum is negative and
neutral if the sum is equal to zero.
Lexicon-based approach is divided into two main
methods: corpus-based and dictionary-based. Dictio-
nary based approach will use an existing dictionary,
which is a collection of opinion words along with
their positive or negative sentiment strength (Ravi and
Ravi, 2015). Corpus based approach relies on the
probability of occurrence of a sentiment word in con-
junction with positive or negative set of words by per-
forming a research on very huge amount of text (Ravi
and Ravi, 2015).
The process may also include intensification and
negation called sentiment shifters. Negations are used
to reverse the semantic polarity of a particular term,
while intensifiers are used to change the degree to
which a term is positive or negative as mentioned in
(Alistair and Diana, ).
4 ASPECT-LEVEL SENTIMENT
ANALYSIS
Polarity classification of opinion text at document and
sentence level is helpful in many cases but it does not
provide all the necessary details because they do not
discover what exactly people liked and did not like.
Generally, documents are made of several passages
of opinions of different semantic categories. Thus,
classification at coarse-level does not identify senti-
ments or opinion targets. For example, being posi-
tive/negative of the sentiments about an entity in a
text document, do not mean that the author is being
positive/negative about all the aspects of the expres-
sed entity.
Due to the need of a finer grain analysis, aspect-
level sentiment analysis represents a key step. Aspect-
level sentiment analysis (previously called feature-
based sentiment analysis) describes that an opinion
consists of a sentiment and a target. The objective of
the analysis at this level is to discover the specific tar-
WEBIST 2018 - 14th International Conference on Web Information Systems and Technologies
326

gets and then specify their sentiment polarities. Using
the quintuple definition (ei, aij, sijkl, hk, tl ) (section
2), aspect-level sentiment analysis aims to locate the
first three components. Therefore the analysis is di-
vided into two tasks: Aspect extraction and Aspect
sentiment classification.
The first task is also called opinion target extraction
(Liu, 2015), because it concentrates on the extraction
of both entities and their aspects. Entities appoint
to products names, services, events, etc, and as-
pects,which can be expressed implicitly or explicitly,
generally identify the attributes and components of
entities.
The second step, similar to the identification of the
polarity of opinions at coarse granularity, associates a
polarity with the various extracted opinion targets.
The extraction of remaining components of the quin-
tuple are studied as sub-tasks of aspect-level senti-
ment analysis called opinion holder extraction and
time extraction. The extraction of all quintuples pre-
sent in a document is helpful to produce a summary
of opinions about entities and their aspects.
Such a summary is known as aspect-based summary
(or feature-based summary) (Hu and Liu, 2004).
4.1 Machine Learning Approaches
4.1.1 Supervised Learning
Supervised learning approaches for aspect-level sen-
timent analysis uses the same machine learning algo-
rithms for coarse-level analysis. The difference be-
tween the two levels (coarse and grained) resides in
the features used for the learning. The features used
for both document and sentence levels are not appli-
cable for aspect-level because the key problem is that
they are target independent, whereas, the core concept
of the aspect-level sentiment analysis is the identifi-
cation of opinion target. Researchers study this chal-
lenging problem either by generating a set of target
dependent feature or by determining an application
scope of sentiments that cover the target entity/aspect
in a sentence.
And as supervised learning approaches, machine
learning algorithms need a huge annotated data for
training. In this case a collection of annotated aspects
and non-aspects data is needed.
4.1.2 Unsupervised Learning
Although supervised approaches for aspect-level sho-
wed good results, it is hard to provide a huge as-
pects and non-aspects annotated data collection be-
cause manually labeling data is costly and time con-
suming.
For that reason, several researches have studied the
task using unsupervised approaches.
Same as document and sentence levels, deep lear-
ning techniques were also applied to aspect-level by
generating target and context representations, or by
identification of important sentiment words for targets
(Zhang et al., 2018).
4.2 Lexicon based Approaches
(Liu, 2015) stated that lexicon based approach for
aspect-level sentiment analysis is based on three pil-
lars: (1) a sentiment lexicon containing sentiment
words, phrases, idioms and composition rules, (2) a
set of rules to handle sentiment shifters,the ”but” clau-
ses and other types of sentences, (3) a sentiment ag-
gregation function or a set of sentiment and target re-
lationships acquired from parse trees.
4.3 Ontology based Approaches
An ontology is an explicit, machine-readable speci-
fication of a shared conceptualization (Studer et al.,
1998). Ontologies provide a formal representation
of knowledge since it models the terms in a speci-
fic domain and captures the semantic relation between
these terms.
The usage of such relation is very important in
aspect-level sentiment analysis especially in product
reviews because in such reviews, products are gene-
rally qualified by their aspects. This hierarchical rela-
tion between products and their aspects can be captu-
red using ontological approaches.
Table 1 below reviews some recent articles in sen-
timent analysis and opinion mining. The articles were
collected from academic research sites and organized
according to granularity level and the approaches pre-
sented previously.
5 OTHER LEVELS OF SA
Sentiment analysis is basically studied at the three le-
vels mentioned previously (document, sentence and
aspect), but these levels are not the only ones. A vari-
ety of researchers dealt with the problem using anot-
her levels such as word-level , clause-level, phrase-
level and concept-level.
Sentiment Analysis Approaches based on Granularity Levels
327

Table 1: Summary of articles in sentiment analysis and opinion mining.
Level Approach Technique Studied issue DataSet Year Reference
Document Level
Supervised le-
arning
Naive Bayes,
SVM, Maxi-
mum Entropy,
Stochastic
Gradient
Descent
Movie review classi-
fication
IMDb movie
review dataset
2016 (Tripathy
et al., 2016)
Unsupervised
learning
K-means
clustering
Mood swing analy-
zer
Facebook
messages
2015 (Kalyani
et al., 2015)
Deep learning Deep Me-
mory Net-
work, Long
Short Term
Memory
Document classifica-
tion considering user
and products
IMDb an Yelp 2017 (Dou, 2017)
Deep learning Convolutional
NN, LSTM
Dual prediction of
word and document
sentiments
Movie review,
IMDb, Twit-
ter
2018 (Lee et al.,
2018)
Lexicon
based
Dictionary
based
sentiment classifi-
cation system for
social media genres
(SmartSA)
Twitter, Digg
and MySpace
samples
2016 (Muhammad
et al., 2016)
Sentence Level
Supervised le-
arning
Conditional
Random
Fields
Context-aware ap-
proach for learning
sentiment
Customer
Review and
Multi-domain
Amazon
datasets
2014 (Yang and
Cardie, 2014)
Supervised le-
arning
Joint Frame-
work
Segmentation and
sentence polarity
prediction
Tweet Sem-
Eval 2013
dataset and
Rottentoma-
toes dataset
2015 (Tang et al.,
2015)
Deep learning Recursive
neural net-
work, LSTM
Increasing
phrase/sentence
representation
Stanford
Sentiment
Treebank,
Movie review
dataset
2017 (Huang et al.,
2017)
Deep learning Recursive
neural net-
work
Software libraries
recommendation and
negative results
Stackover
Flow, Mobile
app reviews,
JIRA
2018 (Lin et al.,
2018)
Lexicon
based
Dixtionary
based
Sentiment classifica-
tion of twitter messa-
ges
Stanford
Twitter
Sentiment,
SemEval
2013
2014 (Musto et al.,
2014)
Aspect Level
Supervised le-
arning
Neural net-
work, word
embeddings
and Composi-
tional vector
models
Aspect rating and
weight detection
TripAdvisor
Hotel reviews
2018 (Pham and
Le, 2018)
WEBIST 2018 - 14th International Conference on Web Information Systems and Technologies
328

Table 1: Continued.
Level Approach Technique Studied issue DataSet Year Reference
Aspect Level
Supervised le-
arning
NB, SVM, KNN,
Decision Tree, Bayes
Network
Three ABSA sub-
tasks
Arabic Ho-
tels’ reviews
SemEval-
2016: Task-5
2018 (Al-Smadi
et al., 2018)
Deep learning LSTM Chinese aspect term
sentiment classifica-
tion
SemEval2014
and four Chi-
nese reviews
datasets
2018 (Peng et al.,
2018)
Deep learning (Sentic) LSTM Targeted Aspect-
Based SA
SentiHood,
SemEval
2015
2018 (Ma et al.,
2018)
Unsupervised
learning
Enriched LDA Aspect extraction English and
Persian pro-
duct reveiws
2017 (Shams and
Baraani-
Dastjerdi,
2017)
Unsupervised
learning
Topic modeling
(combined with
word embedding and
ME classifier)
System for Aspect
Based Sentiment
Analysis (ABSA)
SemEval
2016 Task 5
dataset
2017 (Pablos et al.,
2017)
Lexicon
based
Corpus based A media monitoring
system about the opi-
nion mining in politi-
cal field
Arabic jour-
nalistic text
2017 (Najar and
Mesfar, 2017)
Lexicon
based
Dictionary based and
syntactic dependency
Automating training
data labeling
Twitter data-
set (mobile
phones)
2017 (Rathan et al.,
2017)
Lexicon
based
Corpus based Usage of Chinese ra-
dical parts for SA
IPEEN and
TripAdvisor
restaurent
reviews
2018 (Chao and
Yang, 2018)
Ontology ba-
sed
Retrieval and analy-
sis of social media
content
Twitter mes-
sages
2015 (Thakor and
Sasi, 2015)
Ontology ba-
sed
Detection of adoles-
cent depression sig-
nals
Twitter and
social media
channels
2017 (Jung et al.,
2017)
6 COMPARISON OF
APPROACHES
A comparison of Machine learning and lexicon base
approaches is presented in Table 3 as they represent
the two main approaches for sentiment classification
at the document, sentence and aspect levels, whereas
ontology based approaches are specially used at as-
pect level.
Lexicon based approaches are domain indepen-
dent and do not need labelled data. It is a strong
advantage over machine learning approaches that are
dependent to the domain which means that a classi-
fier trained in a certain domain will show weak re-
sults if it is used for a different domain. Another big
inconvenient for ML methods is the need of labelled
data, which requires human participation and annota-
tion that could be expensive and time consuming.
7 CONCLUSION
Sentiment Analysis is gaining more and more popu-
larity nowadays and that is because we all need each
others opinions and point of views. Opinion mining
has been studied in different domains and languages
Sentiment Analysis Approaches based on Granularity Levels
329

Table 2: Comparison of machine learning and lexicon based approaches.
Approaches Advantages Inconvenient
Machine Learning
• Unnecessity of dictiona-
ries
• High accuracy of classi-
fication
• High precision and
adaptability
• Dependent to the dom-
ain
• Slow time
• Needs human participa-
tion and labelled data
Lexicon based
• Does not need labelled
data
• Domain independent
• Fast time
• Needs strong linguistic
resources
• Low accuracy
• Requires dictionaries
that covers lot opinion
words
and has showed to be very effective and benefic in fi-
nance, politics, e-commerce, but it can also be used in
health, cybersecurity and point of view discovery.
In this paper, we presentend the field of sentiment
analysis or opinion mining by covering utilized ap-
proaches based on three levels of granularity (docu-
ment, sentence and aspect). Another important le-
vel related to sentiment analysis is the concept level
which is being dealt with frequently. Concept-based
approaches to sentiment analysis focus on a seman-
tic analysis of text through the use of web ontologies
or semantic networks (Cambria, 2013). This makes
sentiment analysis at concept-level exciting and chal-
lenging and need more researches because of the lack
of sentiment ontologies. As it can be remarked in
the recent papers reviewed previously, researchers are
following deep learning approaches to deal with SA
tasks and challenges, and there is much more to be
done using deep learning approaches.
Most of the researchers in SA are analyzing pe-
oples opinions on social networks, e-commerce sites
and other platforms where people can share there opi-
nions, whereas they can also express their opinions
outside the digital world. Graffiti are a way for pe-
ople to express there opinions in an anonymous man-
ner. Sentiment analysis can be applied to Graffiti for
discovering several caracteristics and traits of society.
REFERENCES
Ahlgren, O. (2016). Research on sentiment analysis: The
first decade. In 2016 IEEE 16th International Con-
ference on Data Mining Workshops (ICDMW), pages
890–899.
Al-Smadi, M., Al-Ayyoub, M., Jararweh, Y., and Qa-
wasmeh, O. (2018). Enhancing aspect-based senti-
ment analysis of arabic hotels reviews using morpho-
logical, syntactic and semantic features. Information
Processing & Management.
Alistair, K. and Diana, I. Sentiment classification of movie
reviews using contextual valence shifters. Computati-
onal Intelligence, 22(2):110–125.
Bhatia, S., Sharma, M., and Bhatia, K. K. (2018). Sen-
timent Analysis and Mining of Opinions, pages 503–
523. Springer International Publishing, Cham.
Cambria, E. (2013). An introduction to concept-level sen-
timent analysis. In Castro, F., Gelbukh, A., and
Gonz
´
alez, M., editors, Advances in Soft Computing
and Its Applications, pages 478–483, Berlin, Heidel-
berg. Springer Berlin Heidelberg.
Cambria, E., Poria, S., Gelbukh, A., and Thelwall, M.
(2017). Sentiment analysis is a big suitcase. IEEE
Intelligent Systems, 32(6):74–80.
Chao, A. F. and Yang, H.-L. (2018). Using chinese
radical parts for sentiment analysis and domain-
dependent seed set extraction. Comput. Speech Lang.,
47(C):194–213.
Chaturvedi, I., Cambria, E., Welsch, R. E., and Herrera, F.
(2018). Distinguishing between facts and opinions for
sentiment analysis: Survey and challenges. Informa-
tion Fusion, 44:65 – 77.
Dou, Z.-Y. (2017). Capturing user and product information
for document level sentiment analysis with deep me-
mory network. In Proceedings of the 2017 Conference
on Empirical Methods in Natural Language Proces-
sing, pages 521–526. Association for Computational
Linguistics.
Feldman, R. (2013). Techniques and applications for senti-
ment analysis. Commun. ACM, 56(4):82–89.
WEBIST 2018 - 14th International Conference on Web Information Systems and Technologies
330

Hemmatian, F. and Karim Sohrabi, M. (2017). A survey on
classification techniques for opinion mining and sen-
timent analysis.
Hu, M. and Liu, B. (2004). Mining and summari-
zing customer reviews. In Proceedings of the Tenth
ACM SIGKDD International Conference on Know-
ledge Discovery and Data Mining, KDD ’04, pages
168–177, New York, NY, USA. ACM.
Huang, M., Qian, Q., and Zhu, X. (2017). Encoding syn-
tactic knowledge in neural networks for sentiment
classification. ACM Trans. Inf. Syst., 35(3):26:1–
26:27.
Jung, H., Park, H.-A., and Song, T.-M. (2017). Ontology-
based approach to social data sentiment analysis: De-
tection of adolescent depression signals. 19:e259.
Kalyani, Gupta, E., Rathee, G., Kumar, P., and Singh Chau-
han, D. (2015). Mood swing analyser: A dynamic
sentiment detection approach. 85:149–157.
Lee, D., Ju, H., Yu, H., Park, J.-m., and Kim, K.-Y. (2018).
Dualsentinet: Dual prediction of word and document
sentiments using shared word embedding. In Procee-
dings of the 12th International Conference on Ubiqui-
tous Information Management and Communication,
IMCOM ’18, pages 34:1–34:9, New York, NY, USA.
ACM.
Lin, B., Zampetti, F., Bavota, G., Di Penta, M., Lanza, M.,
and Oliveto, R. (2018). Sentiment analysis for soft-
ware engineering: How far can we go? In Procee-
dings of 40th International Conference on Software
Engineering.
Liu, B. (2012). Sentiment Analysis and Opinion Mining.
Morgan & Claypool Publishers.
Liu, B. (2015). Sentiment Analysis: Mining Opinions, Sen-
timents, and Emotions. Cambridge University Press.
Ma, Y., Peng, H., and Cambria, E. (2018). Targeted
aspect-based sentiment analysis via embedding com-
monsense knowledge into an attentive lstm. In AAAI.
Mantyla, M. V., Graziotin, D., and Kuutila, M. (2018). The
evolution of sentiment analysisa review of research to-
pics, venues, and top cited papers. Computer Science
Review, 27:16 – 32.
Missen, M. M. S., Boughanem, M., and Cabanac, G.
(2012). Opinion mining: reviewed from word to do-
cument level. Social Network Analysis and Mining,
3:107–125.
Muhammad, A., Wiratunga, N., and Lothian, R. (2016).
Contextual sentiment analysis for social media gen-
res. Knowledge-Based Systems, 108:92 – 101. New
Avenues in Knowledge Bases for Natural Language
Processing.
Musto, C., Semeraro, G., and Polignano, M. (2014). A
comparison of lexicon-based approaches for senti-
ment analysis of microblog. 1314:59–68.
Najar, D. and Mesfar, S. (2017). Opinion mining and senti-
ment analysis for arabic on-line texts: application on
the political domain. 20:1–11.
Pablos, A. G., Cuadros, M., and Rigau, G. (2017).
W2VLDA: almost unsupervised system for aspect ba-
sed sentiment analysis. CoRR, abs/1705.07687.
Peng, H., Ma, Y., Li, Y., and Cambria, E. (2018). Learning
multi-grained aspect target sequence for chinese sen-
timent analysis. Knowledge-Based Systems, 148:167
– 176.
Pham, D.-H. and Le, A.-C. (2018). Learning multiple
layers of knowledge representation for aspect based
sentiment analysis. Data & Knowledge Engineering,
114:26 – 39. Special Issue on Knowledge and Systems
Engineering (KSE 2016).
Rathan, M., Hulipalled, V., K R, V., and Patnaik, L. (2017).
Consumer insight mining: Aspect based twitter opi-
nion mining of mobile phone reviews.
Ravi, K. and Ravi, V. (2015). A survey on opinion mining
and sentiment analysis. Know.-Based Syst., 89(C):14–
46.
Shams, M. and Baraani-Dastjerdi, A. (2017). Enriched lda
(elda): Combination of latent dirichlet allocation with
word co-occurrence analysis for aspect extraction. Ex-
pert Systems with Applications, 80:136 – 146.
Studer, R., Benjamins, V., and Fensel, D. (1998). Know-
ledge engineering: Principles and methods. Data &
Knowledge Engineering, 25(1):161 – 197.
Sun, S., Luo, C., and Chen, J. (2017). A review of natu-
ral language processing techniques for opinion mining
systems. Information Fusion, 36:10 – 25.
Tang, D., Qin, B., Wei, F., Dong, L., Liu, T., and Zhou,
M. (2015). A joint segmentation and classification
framework for sentence level sentiment classification.
IEEE/ACM Transactions on Audio, Speech, and Lan-
guage Processing, 23(11):1750–1761.
Thakor, P. and Sasi, S. (2015). Ontology-based sentiment
analysis process for social media content. Procedia
Computer Science, 53:199 – 207. INNS Conference
on Big Data 2015 Program San Francisco, CA, USA
8-10 August 2015.
Tripathy, A., Agrawal, A., and Rath, S. K. (2016). Clas-
sification of sentiment reviews using n-gram machine
learning approach. Expert Systems with Applications,
57:117 – 126.
Turney, P. D. (2002). Thumbs up or thumbs down?: Se-
mantic orientation applied to unsupervised classifica-
tion of reviews. In Proceedings of the 40th Annual
Meeting on Association for Computational Linguis-
tics, ACL ’02, pages 417–424, Stroudsburg, PA, USA.
Association for Computational Linguistics.
Yang, B. and Cardie, C. (2014). Context-aware learning for
sentence-level sentiment analysis with posterior regu-
larization. In Proceedings of the 52nd Annual Meeting
of the Association for Computational Linguistics (Vo-
lume 1: Long Papers), pages 325–335. Association
for Computational Linguistics.
Zhang, L., Wang, S., and Liu, B. (2018). Deep learning for
sentiment analysis: A survey. Wiley Interdisciplinary
Reviews: Data Mining and Knowledge Discovery.
Sentiment Analysis Approaches based on Granularity Levels
331