Opinion Mining Meets Decision Making:
Towards Opinion Engineering
Klemens Schnattinger
1
and Heike Walterscheid
2
1
Computer Science, Baden-Wuerttemberg Cooperative State University Loerrach,
Hangstrasse 46-50, 79539 Loerrach, Germany
2
Economics, Baden-Wuerttemberg Cooperative State University Loerrach, Hangstrasse 46-50, 79539 Loerrach, Germany
Keywords: Opinion Mining, Decision Making, Natural Language Processing, Description Logics.
Abstract: We introduce a methodology for opinion mining based on recent approaches for natural language
processing and machine learning. To select and rank the relevant opinions, decision making based on
weighted description logics is introduced. Therefore, we propose an architecture called OMA (Opinion
Mining Architecture) that integrates these approaches of our methodology in a common framework. First
results of a study on opinion mining with OMA in the financial sector are presented.
1 INTRODUCTION
The work reported is part of the project OMA
aiming at the development of an opinion mining and
evaluation system for real-world domains. The
methodological challenge is two-fold. The opinion
mining task is that the textual sources must be pre-
processed and analysed as well as the opinion
evaluation task is to put the opinions in an order.
To address these problems, we concentrate on
foundations of Natural Language Processing (NLP)
in combination with machine Learning (ML) (Sun,
Luo and Chen, 2017) and on Weighted Description
Logics, an extension of “classical” Description
Logics (DL) with utility theory for the calculation of
quantitative preference relations (Acar et al., 2017).
Hence, we combine these techniques in a common
architecture, called the Opinion Mining Architecture
(OMA). In addition, we present data from a first
empirical evaluation of OMA. Qualitative measures
are the subject of future research to focus more on
validity and causality of sentiments.
2 OPINION MINING
2.1 Opinion and Opinion Mining
Usually, the term opinion is defined as “the personal
view that someone has about something”
(Dictionary, 2002). Formally, an opinion is defined
as follows (Liu, 2012):
where
denotes the th entity,
the th aspect of the th
entity,
the th opinion holder,
the time when
the opinion is expressed,
the opinion towards
the th aspect of the th entity from opinion holder
at time
.
For example, in “The screen of this tablet is
good”, the components
,
and
can be
identified: screen is an aspect of the entity tablet.
Additionally, a positive sentiment is expressed.
and
are not given, that is, the five components
are not always necessary to express an opinion.
To perform opinion mining, machine learning
approaches are meaningful. Classifiers are used and
are trained with known texts to identify their
sentiment orientation. For the task of identifying the
opinion holder, detecting opinion expressions, and
identifying the target or aspect of the opinion,
corpora with annotated opinion or sentiment scores
are necessary but difficult to get.
In contrast, lexicon approaches identify the
sentiment of text purely without a training set
according to given sentiment lexicons. A sentiment
lexicon is a dictionary of sentiment words and
phrases, contains a sentiment orientation and a
strength for each sentiment entry, which is expressed
through a sentiment score. Lexicons use less
resources, because they don’t use annotated corpora.
In addition, such a sentiment lexicon can be
Schnattinger K. and Walterscheid H.
Opinion Mining Meets Decision Making: Towards Opinion Engineering.
DOI: 10.5220/0006576403340341
In Proceedings of the 9th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (KDIR 2017), pages 334-341
ISBN: 978-989-758-271-4
Copyright
c
2017 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
integrated into machine learning approaches. Thus,
performance can be significantly increased.
2.2 Natural Language Processing
To perform opinion mining the reviewed texts must
be pre-processed. For this purpose, the following
processes are usually carried out for structuring the
text and for extracting features:
Tokenization decomposes a sentence or
document into tokens. Tokens represents words or
phrases. For English or German, the decomposition
of words is easy with spaces, but some additional
expertise should be kept in mind, such as opinion
phrases and named entities. Words, such as “the”,
“a” only provide little information. Thus,
tokenization must remove these words, which are
called stop words.
POS tagging is a technique that analyses the
lexical information of a word for determining their
POS tag (e.g. adjective or noun). POS tagging is a
so-called sequential labelling problem. Conditional
Random Fields (CRFs) (Lafferty, McCallum and
Pereira, 2001) and Markov models (Sutton and
McCallum, 2011) are applied to this problem. On
the one hand, adjectives can represent opinion
words. On the other hand, entities and aspects of
opinion mining can be expressed with nouns or
combination of nouns.
Parsing is a technique that provides syntactic
information. Among other things, it analyses the
grammatical structures of a given sentence and
generates a tree with the corresponding relationship
of different so-called constituents as “a group of
words treated by a syntactic rule as a unit” (Carnie,
2010). Unlike POS tagging, parsing determines
richer structural information. It can be used
especially for fine-grained opinion mining (Socher,
Bauer and Manning, 2013).
2.3 Machine Learning
For opinion mining gaining features from texts is
important. Thus, text features are discussed,
including n-gram features with weighting schemes,
syntactic features and semantic features.
An n-gram is a set of n adjacent items.
Additionally, the number of times an item appears in
the text is denoted. In opinion mining, double-digit
weights of unigram and bigram are widely accepted.
Instead of binary weights, other schemes can be used
(Paltoglou and Thelwall, 2010).
Syntactic features include POS tags and
syntactic information. These features either build up
a feature space for machine learning approaches
(Joshi and Penstein-Rosé, 2009), or generate rules
for e.g. entities and aspects in fine-grained opinion
mining (Gindl, Weichselbraun and Scharl, 2013).
Semantic features are conjunctions which
specifies negation, increase, and decrease of a
sentiment. Negation turns the sentiment orientation
into the opposite. Increase and decrease also
influence the strength of sentiment, respectively, are
useful for opinion mining (Taboada et al., 2011).
Opinion mining is usually divided into three
levels: document level, sentence level, and fine-
grained level. The task of the document level
opinion mining determines sentiment orientation of
an entire document. The objective of the document
level opinion mining is identifying the
in
Recent techniques for document
level opinion mining are among others:
Supervised approaches: Usual classifiers in
machine learning, such as a Naïve Bayes or Support
Vector Machines, are used. The features considered
are, among others, n-gram, POS tags, position
information (Pang, Lee and Vaithyanathan, 2002)
and semantic features (Kennedy and Inkpen, 2006).
Probabilistic generative models: Generative
models such as joint sentiment topic model (Lin and
He, 2009) are proposed which use a Markov chain.
Unsupervised lexicon-based approaches:
Averaged sentiment orientation is used to suggest
the overall sentiment orientation of an entire
document (Turney, 2002). To improve the results
e.g. discourse structure-based weighting scheme
(Bhatia, Ji and Eisenstein, 2015) are proposed.
In opinion mining at the sentence level,
sentiment orientation is determined for each
sentence in the document. However, not all the
detailed information of opinions is collected such as
opinion target and opinion holder. For example,
“The screen of this tablet is good.” expresses a
positive sentiment orientation to aspect “screen” of
entity “tablet”. Recent techniques for sentence level
opinion mining are among others:
Supervised approaches: Again, Naïve Bayes
classifiers are used to determine subjectivity of
sentences (Yu and Hatzivassiloglou, 2003) and
CRFs for the dependencies between sentences (Yang
and Cardie, 2015).
Unsupervised approaches: For subjectivity
classification in sentences graph-based (Pang and
Lee, 2004), as well as lexicon-based approaches
(Kim and Hovy, 2004) exists.
The problems with the fine-grained level opinion
mining can’t be traced with traditional classification
techniques. Several variations are suggested
including aspect level opinion mining (Cambria et
al., 2013) that aims to discover aspects or entities of
opinion mining and the corresponding sentiment
orientation. Thus, it is split into two sub-tasks:
opinion target extraction and sentiment
classification. Recent techniques for fine-grained
level opinion mining are:
Unsupervised approaches: Association mining
algorithm for aspect detection and linguistic
knowledge (Popescu, 2005) and part-whole patterns
(Zhang et al., 2010) are considered. For aspects
extraction (Qiu et al., 2009) propose propagation
algorithms. Additionally, rule-based methods are
also suitable (Gindl, Weichselbraun and Scharl,
2013).
Probabilistic generative models: For aspects
detection (Brody and Elhadad, 2010) and sentiment
detection (Lazaridou, Titov and Sporleder, 2013) so
called Latent Dirichlet Allocation (LDA) topic
models are adopted.
2.4 Comparative Opinion Mining
A comparative opinion is defined as a relationship of
similarities or differences between two entities.
Comparative opinion mining takes these entities and
preferences of opinion holders into account. From
comparative sentences, compared entities,
comparative words and aspects can be extracted. For
instance, in “Tablet X’s screen is better than tablet
Y.”, “tablet X” and “tablet Y” are the compared
entities, “better” is the comparative word and
“screen” is the compared aspect. Because the word
“better” expresses the preference, “tablet X” is
preferred. However, many comparative words, e.g.,
“longer”, express different positive or negative
sentiment orientations in different contexts.
A rule-based method for this kind of sentence
decomposes this problem into two sub-tasks (Jindal
and Liu, 2006): comparative sentence identification
and comparative relation extraction. Class
Sequential Rules (CSRs) with class labels (i.e.,
“comparative” or “noncomparative”) and Label
Sequential Rules (LSRs) applied on comparative
sentences help solving these tasks, respectively.
Another method divides comparative sentences
into two categories: opinionated comparatives and
comparatives with context-dependent opinions
(Ganapathibhotla and Liu, 2008). In the first case,
comparative words are used. In the second, external
information is needed.
3 DECISION MAKING
3.1 Preference and Utility
Preferences are an important variable in the study of
decisions such as in mathematical economics, social
choice theory and opinion mining. To keep it simple
in the beginning preferences will be “modelled as a
binary relation over the set of choices” (Kaci, 2011).
A set of choices for a rational agent as homo
oeconomicus (Mill, 1836) which has the preference
relation are named and
is read “
is at
least as good as
” where
. Furthermore,
at is a complete, reflexive and transitive relation.
There are two preference relations for :
for any ,
This is read: is better than .
This is read: the agent is indifferent between
and .
A utility function maps a choice to a real
number representing the degree of request. The
representation theorems formally are defined as
follows (Fishburn, 1969):
Given the choices a utility function,
represents
For instance, if and
, this leads to
since 5 < 20. This means, the choices
and are values of a single
attribute (Acar et al., 2017).
Normally, due to framing or irrationality e.g.
decisions are more complex (Tversky and
Kahneman, 1981). Therefore, choices are formalized
as values or elements of attributes. For instance, if
we will buy a car, not only the price will be of
interest, but also its colour, and even more.
Formally, the set of attributes is denoted by .
Then,
refer to a specific attribute in where
. With these preliminaries, we can
formalize the set of choices made by the cartesian
product over the set of attributes. This set of choices
is denoted by where
. Now, the
utility function has been expanded:
is
the (multi-attribute) utility function which represents
,
The size of the is
, the assumption that u is
additive helps to significantly reduce the complexity.
A typical additive function is
where
.
Now, we can formulate an optimization task,
namely that a rational agent should make the choice
with the maximum utility:
where matches to maximal elements in
with respect to the utility function (and therefore
means w.r.t. the preference relation ).
3.2 Description Logics
The signatures of description logics (Baader et al.,
2003) can be given as a triple
, where
denotes the set of atomic concepts,
the set of role
names and
the set of atomic individuals.
We denote concepts or classes by and , roles
by and , and individuals as and . Concept
descriptions are defined in a common way from
as , , and if and are concept
descriptions. Further, and exist if
and is a concept description. The top concept is
an abbreviation for and for .
For the semantic we need an interpretation for
the presented syntax. An interpretation is a pair
where the domain
is a set that can’t be
empty, and
is a so-called interpretation function.
This function maps to every concept name a set
and to every role name a binary relation
. The function also defines:
ex
In DLs, we distinguish between terminological
knowledge (so-called TBox) and assertional
knowledge (so-called ABox). A TBox is a set of
concept inclusions which has the semantics
and a concept definition is if
. An ABox is a set of concept
assertions where
and
, as well as role assertions where
and
.
In the following we will consider only a
coherent TBox . This means that all concepts in
are satisfiable. The usual interpretation function is
used for the notion satisfiable (Baader et al., 2003)
and write . We say that an ABox entails an
assertion α (and write ), if every model of
also satisfies. An ABox is called consistent
with a TBox if there exists an interpretation that
satisfies and . We then call the pair
a knowledge base. Further, is satisfiable if is
consistent w.r.t. . In the remainder, we will use the
instance check. Thus, for a knowledge base and
an assertion , one can check whether holds.
A concrete domain is defined as a pair
.
is the domain of and pred()
is the set of predicate names of . The following
assumptions have been applied:
and
for each with arity n there is
. According to (Baader et al., 2003), functional
roles are denoted with lower case letters, for
example with . In description logics with concrete
domains,
is partitioned into a set of functional
roles and one of ordinary roles. A role is functional
if for every and it is necessary
that Functional roles are explained
as partial functions from
to
. A concrete
domain is closed under negation (denoted by
). For
this reason, a logical formula can be calculated
which are in the so-called negation normal form
(NNF). A formula is in NNF when the negation
operators are only used between atomic statements.
3.3 Weighted Description Logics
We will introduce an ontological approach to
decision making. This approach can be considered
as a generic framework, the so-called DL decision
base (Acar et al., 2017). We use an a priori
preference relation over attributes (called the
ontological classes). Thereby, an a posteriori
preference relation over choices (called ontological
individuals) can be derived. Formally, a priori utility
function over (the set of attributes) is defined
( ). Additionally, a utility function u
defined over choices, which uses logical entailment,
extends the utility function U to the subset of
attributes. The utility function u was used because a
choice was defined as an individual and its outcome
as a set of concepts. Another reason is that can
take various forms, e.g., . Modelling
attributes has two steps:
1. Each attribute is modelled by a concept.
2. For every value of an attribute a new
(sub)concept has been introduced.
For instance, if colour is an attribute to be
modelled, it is simply represented by the concept
(i.e., ). A colour can be
regarded as a value, as if it were a concept of its
own. If blue is a value of the attribute colour, the
attribute set is simply extended by adding the
concept Blue, as a sub-concept of . It should
be noted, that an axiom has been introduced to
guarantee the disjointedness. (e.g. )
and that this procedure results in a binary term
vector for , because an individual c (as a choice) is
either a member of the concept or not.
Given a total preference relation (i.e.,
) over
an ordered set of not necessarily atomic attributes ,
and a function that represents (i.e.,
iff
for
). The
function asigns an a priori weight to each concept
. Therefore, one can say, that “ makes the
description logic weighted”. The utility of a concept
is denoted by . The following applies:
The greater the utility of an attribute the more the
attribute is preferable. Furthermore, the attribute set
can be divided into two subsets:
desirable denotes the set of attributes with
non-negative weights, denoted
, and
undesirable
, i.e., iff
and
with
This means that any attribute that is not in
(not desirable) must lie in
and is therefore
undesirable. In addition, it should be noted that an
attribute with weight zero can be interpreted as
desirable with no utility.
As mentioned above, a choice is an individual
. denotes the finite set of choices. To
determine a preference relation (a posteriori) over
(i.e.,
), which respects
, a utility function
is introduced. indicates the utility of
a choice relative to the attribute set . Also, a
utility function over attributes as an aggregator is
introduced. For simplicity, the symbol is used for
both choices and attributes whenever it is evident
from the context.
The -utility is a particular and is defined as
and
and is
called the sigma utility of a choice .
triggers a preference relation over i.e.,
iff
. Each choice
corresponds to a set of attributes, which is logically
entailed e.g., . Due to the criterion
Additivity, each selection corresponds to a result.
Putting things (DL, and ) together, a generic
UBox (so-called Utility Box) is defined as a pair
, where is a utility function over and
is the utility function over . Also, a decision base
can be defined as a triple where
is a consistent knowledge base, is a TBox
and is an ABox,
is the set of choices, and
is an UBox. Note: provides
assertional information about the choices and
terminological information about the agent ability to
reason over choices.
Example:
We want to buy a tablet computer. Two alternatives
are considered, which fit the original purpose. The
buyer’s decision base ( , choices
, and attributes mentioned in ) are
given. The language uses discrete domains. The
domain is used and
with
and
. The partition
of domain
is
and
with
with andsuch that .
Further predicates are defined similar. Note:
is closed under negation. This means that
we can invert the predicates in an obvious way like
. The other partition is
defined as follows:
. The
remaining predicate names and functional roles are
also defined (basic predicate names and functional
roles like
are not given here):
( ),
Considering the agent is more interested in a
tablet with a keyboard than in an upper class or
inexpensive tablet. The utilities can be calculated by
and
. Thus,
.
Figure 1: The Opinion Mining Architecture OMA.
4 OMA - THE OPINION MINING
ARCHITECTURE
The Opinion Mining Architecture (OMA) we
propose is strongly basing on the approaches of
natural language processing and machine learning
presented in Section 0 as well as on decision making
with weighted description logics presented in section
3. We separate OMA into opinion mining (pre-
processing, analysing texts, filtering out opinions)
and decision support (evaluating extracted opinions)
according to SYNDIKATE (Hahn and Schnattinger,
1997). OMA serves the generation of opinions from
texts like news, employee and public participation,
expressions of opinions, political conversations, etc.
(see step 1 in Figure 1). The representation of the
underlying domain (TBox) as well as the opinions
expressed as assertions (ABox) use a description
logic model (see step 2 in Figure 1). The TBox
contains concepts which represents artefacts like
compliance, rule, judgment, idea, sentiment,
opinion, etc. The ABox contains assertions. In terms
of content, it consists of opinions that are extracted
from the sources of text. Whenever an opinion is
stored in the ABox, different types of machine
learning and natural language processing models
carried out an evaluation (see step 3 in Figure 1).
These models are presented in a so-called MBox
(methodology box). The evaluation provides a
ranking of the opinions according to their utility.
These weighted opinions are stored in the UBox (see
section 3.3). Note: Not every opinion can be
weighted and therefore does not appear in the UBox.
In view of OMA architecture, we intend to build
a model for opinion mining in various domains such
as sentiment mining for the financial sector. The
results of a first attempt to determine sentiments for
Deutsche Bank, Commerzbank, Volksbank and
Sparkasse during the introduction of account
management fees in spring 2017 has shown that
OMA can deliver conclusive results. Starting from
measured sentiment score for each of these banks,
the sentiment scores for those banks fell, which have
announced the introduction of a fee for account
management in April 2017. As you can see in Figure
2 sentiment scores for the Sparkasse and Volksbank
ran relatively uniformly from January to March
2017. In April, the score declined due to the
announcement of account management fees. One
month later in May, after first account fees were
reported on the account statement, the score fell
significantly. For Deutsche Bank and Commerzbank
such behavior couldn’t be observed, since these
banks charge account fees for a long time already.
Interestingly, this result could have been
achieved by the fact that a supervised learning
method had to be used to improve the results of the
score calculation in addition to the pre-processing
techniques of NLP, such as stop word lists and
tokenization. Therefore, we used a Naïve Bayes
classifier at document level and trained him with
several hundred tweets. To select the right tweets,
we use a bag-of-word model with unigrams. As a
technological platform, we used OpenNLP.
Figure 2: Sentiment scores for Sparkasse, Volksbank, Deutsche Bank and Commerzbank from January to May 2017.
5 CONCLUSION & FUTURE
WORK
We have presented a methodology for opinion
mining together with decision making based on
machine learning, natural language processing
methods for emerging opinions and weighted
description logics. We were also able to present an
initial evaluation showing that OMA can deliver
good results.
May the approaches of opinion mining depend
on specific domains, the principles underlying the
ordering of opinions are to be generalized.
Nevertheless, as weighted assertions are ubiquitous,
one may easily envisage assertions with other
content, e.g. data from IoT devices that provide
incorrect values due to electronic fluctuations. The
extension of OMA to data from IoT is also part for
our project. From a formal perspective, we will
introduce the methods mentioned in Section 2.3,
such as supervised approaches with semantic
features to get more information about the opinions
causal nexus. Finally, we want to compare these
approaches in a comprehensive evaluation and make
recommendations for one or the other approach.
ACKNOWLEDGEMENTS
We would like to thank our colleague Prof. Dr.
Jürgen Schenk for cooperation in the project OMA.
Klemens and Heike are winners of the Dr. Karl
Helmut Eberle Foundation's award on the study
“Digitization and Knowledge Transformation”.
REFERENCES
Acar, E., Fink, M., Meilicke, C., Thome, C. and
Stuckenschmidt, H. (2017) 'Multi-attribute Decision
Making with Weighted Description Logics', IFCoLog:
Journal of Logics and its Applications, 4, pp. 1973-
1995.
Baader, F., McGuinness, D., Narci, D. and Patel-
Schneider, P. (2003) The Description Logic
Handbook: Theory, Imlementation, and Applications,
Cambridge University Press.
Bhatia, P., Ji, Y. and Eisenstein, J. (2015) 'Better
Document-level Sentiment Analysis from RST
Discourse Parsing', EMNLP'15: Proc. 2015 Conf. on
Empirical Methods in Natural Language Processing,
2212-2218.
Brody, S. and Elhadad, N. (2010) 'An Unsupervised
Aspect-Sentiment Model for Online Reviews', HLT-
NAACL'10: Human Language Technologies: The
2010 Annual Conf. North American Chapter of the
Association for Computational Linguistics, 804–812.
Cambria, E., Schuller, B., Liu, B., Wang, H. and Havasi,
C. (2013) 'Statistical Approaches to Concept-Level
Sentiment Analysis', IEEE Intelligent Systems 28, pp.
6-9.
Carnie, A. (2010) Constituent Structure, Second Edition
edition, Oxford University Press.
Dictionary (2002) The American Heritage® Idioms
Dictionary, [Online], Available:
www.dictionary.com/browse/opinion [1 Jul 2017].
Fishburn, P. (1969) Utility Theory for Decision Making,
New York, London, Sydney, Toronto: John Wiley &
Sons, Inc.
Ganapathibhotla, M. and Liu, B. (2008) 'Mining Opinions
in Comparative Sentences', COLING'08: Proc. 22nd
Intl. Conf. on Computational Linguistics, 241–248.
Gindl, S., Weichselbraun, A. and Scharl, A. (2013) 'Rule-
based Opinion Target and Aspect Extraction to
Acquire Affective Knowledge', WWW'13: Proc. 22nd
Intl. Conf. on World Wide Web, New York, NW, 557-
564.
Hahn, U. and Schnattinger, K. (1997) 'Deep Knowledge
Discovery from Natural Language Texts', KDD'97:
Proc. 3rd Intl. Conf. on Knowledge Discovery and
Data Mining, Newport Beach, CA, 175-178.
Jindal, N. and Liu, B. (2006) 'Mining Comparative
Sentences and Relations', AAAI'06: Proc. 21st
National Conf. on Artificial Intelligence, 1331–1336.
Joshi, M. and Penstein-Rosé, C. (2009) 'Generalizing
Dependency Features for Opinion Mining', ACL'09:
Proc. 47th Annual Meeting of the Association for
Computational Linguistics and the 4th Intl. Joint Conf.
on Natural Language Processing, 313-316.
Kaci, S. (2011) Working with Preferences. Less is More,
Heidelberg, Dordrecht, London, New York: Springer.
Kennedy, A. and Inkpen, D. (2006) 'Sentiment
Classification of Movie Reviews Using Contextual
Valence Shifters', Computational Intelligence 22, pp.
110-125.
Kim, S.-M. and Hovy, (2004) 'Determining the Sentiment
of Opinions', COLING'04: Proc. 20th Intl. Conf. on
Computational Linguistics, 1367–1373.
Lafferty, J., McCallum, A. and Pereira, F. (2001)
'Conditional Random Fields: Probabilistic Models for
Segmenting and Labeling Sequence Data', ICML'01:
Proc. 8th Intl. Conf. on Machine Learning, 282-289.
Lazaridou, A., Titov, I. and Sporleder, C. (2013) 'A
Bayesian Model for Joint Unsupervised Induction of
Sentiment, Aspect and Discourse Representations',
ACL'13: Proc. 51st Annual Meeting of the Association
for Computational Linguistics, 1630–1639.
Lin, C. and He, Y. (2009) 'Joint Sentiment/Topic Model
for Sentiment Analysis', CIKM'09: Proc. 8th ACM
Conf. on Information and Knowledge Management,
New York, NY, 375-384.
Liu, B. (2012) Sentiment Analysis and Opinion Mining,
Morgan & Claypool Publishers.
Mill, J.S. (1836) On the Definition of political economy
and on the method of philosophical investigation in
that science, London and Westminster.
Paltoglou, G. and Thelwall, M. (2010) 'A Study of
Information Retrieval Weighting Schemes for
Sentiment Analysis', ACL'10: Proc. 48th Annual
Meeting of the Association for Computational
Linguistics, 1386-1395.
Pang, B. and Lee, L. (2004) 'A Sentimental Education:
Sentiment Analysis Using Subjectivity Summarization
Based on Minimum Cuts', ACL'04: Proc. 42nd Annual
Meeting of Association for Computational Linguistics,
271–278.
Pang, B., Lee, L. and Vaithyanathan, S. (2002) 'Thumbs
up? Sentiment Classification using Machine Learning
Techniques', EMNLP'02: Proc. Conf. on Empirical
Methods in Natural Language Processing, 79–86.
Popescu, A.-M. (2005) 'Extracting Product Features and
Opinions from Reviews', HLT'05: Proc. Conf. on
Human Language Technology and Empirical Methods
in Natural Language Processing, 339–346.
Qiu, G., Liu, B., Bu, J. and Chen, C. (2009) 'Expanding
Domain Sentiment Lexicon through Double
Propagation', IJCAI'09: Proc. 21st Intl. Joint Conf. on
Artificial Intelligence, 1199–1204.
Socher, R., Bauer, J. and Manning, C. (2013) 'Parsing with
Compositional Vector Grammars', ACL'13: Proc. 51st
Annual Meeting on Association for Computational
Linguistics, 455-465.
Sun, S., Luo, C. and Chen, J. (2017) 'A Review of Natural
Language Processing Techniques for Opinion Mining
Systems, Volume 36', Information Fusion, pp. 10-25.
Sutton, C. and McCallum, A. (2011) 'An Introduction to
Conditional Random Fields', Foundations and Trends
in Machine Learning 4, pp. 267-373.
Taboada, M., Brooke, J., Tofiloski, M., Voll, K. and
Stede, M. (2011) 'Lexicon-Based Methods for
Sentiment Analysis', Computational Linguistics 37,
pp. 267-307.
Turney, P. (2002) 'Thumbs Up or Thumbs Down?
Semantic Orientation Applied to Unsupervised
Classification of Reviews', ACL'02: Proc. 40th Annual
Meeting on Association for Computational
Linguistics, 417-424.
Tversky, A. and Kahneman, D. (1981) 'The Framing of
Decisions and the Psychology of Choice', Science,
Vol. 211, No. 4481, pp. 453-458.
Yang, B. and Cardie, C. (2015) 'Context-aware Learning
for Sentence-level Sentiment Analysis with Posterior
Regularization', ACL'14: Proc. 42nd Annual Meeting
of the Association for Computational Linguistics, 325-
336.
Yu, H. and Hatzivassiloglou, V. (2003) 'Towards
Answering Opinion Questions: Separating Facts from
Opinions and Identifying the Polarity of Opinion
Sentences', EMNLP'03: Proc. 2003 Conf. on
Empirical Methods in Natural Language Processing,
129-136.
Zhang, L., Liu, B., Lim, S. and O'Brien-Strain, E. (2010)
'Extracting and Ranking Product Features in Opinion
Documents', COLING'10: Proc. 23rd Intl. Conf. on
Computational Linguistics, 1462–1470.
