Sarcasm Detection using Sentiment and Semantic Features
Prateek Nagwanshi and C. E. Veni Madhavan
Department of Computer Science and Automation, Indian Institute of Science, Bangalore, India
Keywords:
Sarcasm Detection, Natural Language Processing, Opinion Mining, Computational Linguistics, Classification.
Abstract:
Sarcasm is a figure of speech used to express a strong opinion in a mild manner. It is often used to convey the
opposite sense of what is expressed. Automatic recognition of sarcasm is a complex task. Sarcasm detection
is of importance in effective opinion mining. Most sarcasm detectors use lexical and pragmatic features for
this purpose.
We incorporate statistical as well as linguistic features. Our approach considers the semantic and flipping of
sentiment as main features. We use machine learning techniques for classification of sarcastic statements. We
conduct experiments on different types of data sets, and compare our results with an existing approach in the
literature. We also present human evaluation results. We propose to augment the present encouraging results
by a new approach of integrating linguistic and cognitive aspects of text processing.
1 INTRODUCTION
Mining text, of both structured and unstructured na-
ture, involves the use of a variety of models and tech-
niques. The subtle applications of opinion, sentiment
analysis require a good handling of human figures-
of-speech. This challenging task involves inherently
the recognition of a variety of literary devices such
as metaphor, sarcasm, pun, wit. Our contention is
that a combination of machine learning methods for
large corpora analysis must be integrated with linguis-
tic and cognitive processing algorithms.
In our evolving research programme termed DI-
AMETERS (Dialog, Metaphor, Expansion, Rewrit-
ing and Summarization) we seek a descriptive and
prescriptive methodology. In this we capture the lin-
guistic and cognitive aspects of human processing of
textual utterances. The plan is to utilize the syntac-
tic and grammatical information from parse trees and
pos tags with semantic informationfrom typed depen-
dencies ([Stanford]) together with a proposed system
of cognitive tags. The cognitive tags will typically
consist of the wh-tags and the many-to-many relation-
ships, M , between the wh- question tags and the de-
pendency types. For example, the who, what tags will
be related to the types nsubj, dobj and a few others.
At present we are building this set M of mappings. In
our integrated system the standard grammatical pars-
ing of a sentence S would be followed by traversals
of the typed dependencies of S and the cognitive re-
lationship maps M to elicit a natural understanding
of S. Indeed, this stage will also involve the handling
of appropriate ontologies pertaining to world knowl-
edge. Another ongoing investigation is on the iter-
ative, unification scheme for information on linguis-
tic and cognitive parsing of two successive sentences.
This stage will identify further contextual, discourse
information from other devices of anaphora, named-
entities. In this manner certain complex Natural Lan-
guage Processing(NLP) tasks, that rely on gures-of-
speech, such as metaphors, discourse, are expected to
be handled more naturally and hence with better suc-
cess for machine processing. A further possibility is
that this approach will yield a language independent
processing methodology.
In this work we address one such complex NLP
task, namely sarcasm detection. At present we have
not invoked the full power of the proposed method-
ology. We utilize certain amount of semantic infor-
mation and proceed with a statistical classification
methodology.
Sarcasm is a figure of speech that mostly conveys
the opposite meaning to what is said. In verbal com-
munication, the effect of sarcasm is brought out us-
ing voice tone, pitch, gestures and facial expressions.
In written communication, such effects can not be
used. However in text based communication sarcasm
is used widely. It is used extensively in print media,
418
Nagwanshi P. and Veni Madhavan C..
Sarcasm Detection using Sentiment and Semantic Features.
DOI: 10.5220/0005153504180424
In Proceedings of the International Conference on Knowledge Discovery and Information Retrieval (KDIR-2014), pages 418-424
ISBN: 978-989-758-048-2
Copyright
c
2014 SCITEPRESS (Science and Technology Publications, Lda.)
social digital media channels such as e-mail, blogs,
tweets etc.
A sarcastic statement is a witty or bitter remark
that seems to admire someone or something but ac-
tually is actually used to insult or taunt. (e.g., “I
am trying to imagine you with a personality”). A
statement which contains sarcasm will generally de-
pend upon some context. Hence it is very difficult to
detect sarcasm in single sentence. In language and
literary works, different kinds of sarcasm are used:
self-deprecating, brooding, deadpan, polite, obnox-
ious, manic and raging. Humans can generally distin-
guish such subtle varieties of sarcasm. However, it is
a challenge (Gonz´alez-Ib´anez et al., 2011) to develop
a computational scheme to even distinguish between
literal and sarcastic statements.
Processing well-formed natural language sen-
tences at lexical, syntactic and, to some extend se-
mantic levels is an established science. However,han-
dling figures-of-speech that have different properties
has lagged behind because of the lack of remarkable
computational theories and models.
People often use sarcasm and irony to express
their opinions. There are many opinion mining tools.
These fail to identify the sarcastic or ironic utter-
ances. Usage of sarcasm is very common in web
content like tweets, blogs and product reviews. Users
express their feelings or reactions by using sarcasm,
irony and other linguistic devices. To understand
these opinions we have to go deep into the theory of
sarcasm. When a writer wishes to say some negative
remark about someone, he does not convey it directly,
he uses sarcasm to say “a negative thing in positive
words. (e.g., “awww i love to get cute goodnight
texts from no one”). This example shows how peo-
ple use sarcasm for conveying negative views. Words
which exhibit politeness commonly used in sarcastic
utterances. One of our aims is to capture the usage of
positive words to convey negative things.
It is a challenge to automatically interpret and
identify figurative usage of words. Our work is con-
centrates towards sarcasm. Sometimes it is difficult
for humans to identify sarcasm using human intelli-
gence, because it not so obvious. So semanticanalysis
may not be very useful. We model this by statistical
models to predict the sarcastic utterances. Some of
the sarcastic remarks came into picture because of us-
age. So we concentrate on statistical models and try
to to develop a supervised learning model which can
identify the sarcastic utterances.
2 RELATED WORK
Some major works in automatic processing of natural
language texts for detection of sarcastic utterances are
(Lakoff and Johnson, 2008; Utsumi, 2004; Tsur et al.,
2010; Reyes et al., 2012; Riloff et al., ; Gonz´alez-
Ib´anez et al., 2011).
According to (Lakoff and Johnson, 2008) people
often use sarcasm for insulting others. In sarcastic
sentences, the speaker does not explicitly mention the
negative interpretation of the sentence, so it is the re-
sponsibility of the listener to recognize speaker’s in-
tention.
(Utsumi, 2004) shows how linguistic style and
contextual features plays a vital role in processing
irony. He identifies irony on the basis of 3 types
of patterns like: “Opposition”, “Rhetorical question”,
and “Circumlocution.
Opposition is a statement in which the mean-
ing is positive but is related to a negative situation,
like:“This restaurant serves the dishes quickly.
Rhetorical questions are statements which contain
a question as an obvious fact like: “Do you know the
recipe for the dishes?”
Circumlocutions are statements weakly related to
an expectation : I think you are just going to buy the
ingredients for the recipe. According to the author
the degree of irony and sarcasm increases when the
sentence is of type opposition, rhetorical question or
circumlocution.
(Tsur et al., 2010) approached this problem by a
semi supervised algorithm which has two stages: a
pattern collection followed by a classification of sar-
castic utterances. They conduct experiments on re-
views of Amazon.com
1
. They use pattern matching
and features based on punctuation to detect sarcasm.
Each pattern is replaced by its general pattern like
[product], [company]. Classification of a new review
is based on the exact or partial match with stored pat-
terns.
(Gonz´alez-Ib´anez et al., 2011) have done a 3-
way comparison of sarcasm with positive and nega-
tive sentiment carrying tweets. They use lexical and
pragmatic features for the identification of sarcasm
in Twitter data. Lexical feature is a combination of
unigrams and dictionary based features. Pragmatic
feature contains positive emoticons(smilies) and neg-
ative emoticons(frowning faces). According to them
the auxiliary verb and the punctuation are also impor-
tant features for identifying sarcasm. They conducted
human evaluation for checking their algorithm and in
1
www.amazon.in
SarcasmDetectionusingSentimentandSemanticFeatures
419
the the end they conclude that neither the classifier
nor the human judges perform well. Our approach is
motivated by their work. They have not considered
the sentiment based features.
(Reyes et al., 2012) focus on humor and irony
processing. They compare humor and irony with dif-
ferent genres like politics, and technology. They use
different features for the identification. These consist
of ambiguity, polarity, unexpectedness and emotional
scenarios. Ambiguity is a combination of structural,
morphosyntactic and semantic layers. Structural am-
biguity can be viewed as funny situations which occur
most in the text containing humor.
(Liebrecht et al., 2013) tackle the problem by
checking the presence of markers, intensifiers, excla-
mations on twitter data. They find that the mark-
ers like “lol” and “humor” and intensifiers like “awe-
some”, “lovely” and “fantastic” etc and positive ex-
clamations like “yeah” , “yipee” and “wow” indicate
sarcasm in tweets.
(Riloff et al., ) identify the sarcastic tweets that
contain positive sentiment about an activity that is dis-
liked normally. They determine that the pattern of
positive sentiment expression followed by an activity,
which normally people do not like to do, such as work
or study will generally indicate sarcasm. The boot-
strapping process takes a seed word and collects the
negative situation phrases which are preceded by the
seed word. For example, they collect a pattern like: I
love(positive sentiment word) being ignored(negative
situation phrase). Then this procedure is applied in
the opposite direction. They use these phrases as fea-
tures for a SVM based classifier.
3 OUR ALGORITHM AND
FEATURE SET
In this paper, we present a novel supervised learning
algorithm for sarcasm identification. The algorithm
has two modules: (i) Feature-extraction, and (ii) clas-
sification.
In the first step we extract all features described in
the next section. Then we perform a SVM based clas-
sification. We evaluated our system on two data-sets:
Twitter hash-tag data set and Quotation data set.
3.1 Feature Set
Our feature set covers the aspects: lexical, syntac-
tic, semantic, pragmatic, politeness, sentiment flip-
ping aspect.
Lexical feature is based on n-grams(upto bi-
gram) which occur more than twice in the train-
ing data. We build a dictionary from these words,
then use it in a bag-of words approach.
Syntactic feature is the combination of part-of-
speech tags. We found by our studies that ad-
verbs are used often in sarcastic utterances. Pres-
ence of adjectives is also a discriminative fea-
ture. We use certain n-gram combinations of
part-of-speech tags which are very common in
sarcastic utterances. We use the Stanford Pos-
tagger (Toutanova and Manning, 2000) for gen-
erating the part-of-speech tags.
Semantic feature determines whether there exists
words which contradict or are nearly opposite of
each other.
For example:“I love being ignored.”. In the ex-
ample two nearly opposite words are used which
makes the utterance sarcastic.
We use the WordNet (Miller, 1995) for finding the
contradictory words.
Pragmatic feature determines when the senti-
ment of the sentence differs from the emoticons
or similes.
For example: “I just love to wake up early in the
morning” followed by frowning face simile. In
the example the sentiment of the sentence is sup-
posed to be positive but the use of the frowning
face simile makes it sarcastic.
Politeness
rating will determine if there are
words like “extremely”, too” used before any
positive sentiment carrying word. This feature is
very useful in discrimination because when we
sarcastically want to express something then we
say : You are extremely good”, instead of state-
ment like “you are very good”.
Flipping of sentiment will take care of the sen-
timent dissimilarity in the sentiment progression
within a sentence.
For example:“I love when I am sick and not even
able to sleep. . In this example the sentiment of
both clauses differs.
We use Senti-WordNet (Esuli and Sebastiani,
2006) for checking the sentiment values of a
word.
Our feature vector is of dimension 640. The breakup
of feature set is as follows. lexical feature(1-610),
syntactic feature(611-625), semantic feature(626-
627), flipping of sentiment(628-635), pragmatic
feature(636-638) and politeness
rating(639-640).
The lexical feature vector (dimension 610) is an
indicator vector based on a set of distinct content
words collected from the twitter data. At present we
KDIR2014-InternationalConferenceonKnowledgeDiscoveryandInformationRetrieval
420
use only unigrams, for match with the fixed set of
most commonly occurring unigram. We could use
bigrams and trigram phrases. This would expand the
feature vector size. We plan to study the efficacy of
using a more compact representation of the feature
vector. SentiWordNet data is used to mark the
positive and negative words in the sentence.
The classification accuracy was evaluated by em-
ploying the classifiers: Naive Bayes and Support vec-
tor machines(SVM). We show the results on the basis
of some sample data.
4 EXPERIMENTAL SETUP
Although some work is reported in the literature to
solve this problem, we do not have standard data set
for evaluation of performance. i.e. used to evaluate
the performance of the existing algorithms. So we are
unable to find a gold standard data set for sarcasm
identification. So we performed human evaluation of
some data and used this as gold standard. We per-
formed experiments on this data to check credibility
of our algorithm.
Several experiments were performed to evaluate
the capabilities of our algorithm. Based on the feature
set discussed in the previous section, we develop the
feature vectors corresponding to sentences. Then we
use Weka (Hall et al., 2009)which is a tool for doing
various Data Mining tasks like classification, cluster-
ing etc.
We conduct experiments on two kinds of data sets:
Experiment-1 by using tweets as training data and
Experiment-2 by using a quotation data set. Then we
conduct human evaluation of the quotation data and
try to verify the classification obtained by our algo-
rithm.
Experiment-1
Twitter Data:- A lot of opinion mining content is
available in micro-bloggingand social networking
websites: like Twitter
2
. Twitter provides the facil-
ity to users to post and read the messages from
others whom they follow. For most of the tweets
the user provides a hash-tag. This is used to iden-
tify messages on a specific topic like sarcasm. It
is used to group messages which contains similar
topics. One can search the whole tweets by just
typing the hash-tag and get the set of messages.
In Experiment-1 the classifier-1 is trained with
200 instances (100 Sarcastic tweets and 100 Pos-
itive tweets). Classifier-2 is trained with 200 in-
stances (100 Sarcastic tweets and 100 Negative
2
https://twitter.com/
tweets) A 10-fold cross validation method was
used for the test set.
1. Data Collection:- Many of the tweets with the
hash-tag, ”sarcasm“ are available in Twitter.
We used the hash-tag data as a training set. We
used the twitter streaming api
3
for collecting
the tweets which are hash-tagged as #sarcasm,
#positive, and #negative. The hash-tag is not
a reliable source of classification, so we have
manually checked some sample data.
2. Preprocessing:- We carry out an initial prepro-
cessing of the data set. The preprocessing con-
sists of following rules:
If a word starts with “http” ignore it.
If a word starts with “@” ignore it.
If a word starts with “hashtag” ignore it.
If a tweet is written in a language other than
English, then ignore it.
The abbreviations and short acronyms are re-
placed in full.
The smilies are replaced by their indicative
meaning word like “:) with “happy-smilie”
etc.
The emoticons are also replaced by their in-
dicative meaning.
3. Feature Extraction:- Features are extracted
from the training data according to our algo-
rithm.
4. Classification:- This feature set is used for the
classification purpose with the help of Weka.
5. Results:- We incorporate the features in a
graded manner by including the different fea-
ture sets one after the other. First we consider
only the unigram (lexical) features, then the
lexical and syntactic features, then lexical,
syntactic, semantic features, and finally we
introduce the pragmatic (sentiment) features.
The accuracy of the classifier after adding in-
crementally each feature is shown in Table 1.
In Table 2 we give, the classification accuracy
on the experiments between sarcasm vs posi-
tive (Sar-vs-Pos) and sarcasm vs negative (Sar-
vs-Neg).
Experiment-2
1. Standard Quotation Data: Tweets are gener-
ally used in an unstructured format. So there is
a need for cleaning up the data before processing
the sentences. Also hash-tags are not fully reliable
source to assume the class label. Hence there is a
3
https://dev.twitter.com/docs/api/streaming
SarcasmDetectionusingSentimentandSemanticFeatures
421
Table 1: Impact of features(Sar-vs-Pos): Twitter data.
Feature Accuracy
Lexical 68.65
Lexical+Syntactic 71.3
Lexical + Syntactic + Semantic 73.4
Lexical + Syntactic + Semantic + Pragmatic 74.13
Table 2: Sar-vs-Pos and Sar-vs-Neg: Twitter data.
Classifier Accuracy Precision Recall F-measure
Sar-vs-Pos 74.13 0.742 0.741 0.741
Sar-vs-Neg 74.69 0.747 0.747 0.747
need for examination of standard and structured
text like standard quotations.
In Experiment-2 Classifier-1 is trained with 200
instances (100 Sarcastic quotes and 100 Positive
quotes). The Classifier-2 is trained with 200 in-
stances (100 Sarcasm and 100 Negative quotes) A
10-fold cross validation method was used for the
test set.
2. Data Collection: Many standard quotes and
sayings are available in the web. We collect
some standard sarcastic, quotes and some posi-
tive and negative emotion statements from Inter-
net sources.
4 5 6
3. Preprocessing: The words are stored after
lemmatization and stemming.
4. Feature Extraction: Features are extracted ac-
cording to our algorithm. In standard quotations,
in general, pragmatic markers are not used: so we
ignore the pragmatic feature here.
5. Classification: Classification experiment is per-
formed with the help of the tool Weka .
6. Results: The accuracy of the classifier after
adding incrementally each feature is shown in Ta-
ble 3. In Table 4 we give, the classification accu-
racy on the experiments between sarcasm vs posi-
tive (Sar-vs-Pos) and sarcasm vs negative (Sar-vs-
Neg).
We make some observations on the data sets and
experiments. Twitter data is easily available in large
volumes. However this data, although labeled with
a #sarcasm tag, needs to be filtered (as also noted
by (Gonz´alez-Ib´anez et al., 2011)). The quotation
data is a pre-filtered collection. Hence the set of about
100 quotations used by us a reasonable representation
4
www.searchquotes.com
5
www.oocities.org/kristensquotes
6
www.coolnsmart.com
Table 3: Impact of features(Sar-vs-Pos): Quotation data.
Feature Accuracy
Lexical 71.3%
Lexical+Syntactic 73%
Lexical + Syntactic + Semantic 75.4%
Lexical + Syntactic + Semantic + Pragmatic 76.02%
Table 4: Sar-vs-Pos and Sar-vs-Neg: Quotation data.
Classifier Accuracy Precision Recall F-measure
Sar-vs-Pos 76.02 0.76 0.76 0.76
Sar-vs-Neg 77.77 0.778 0.778 0.777
of sarcastic content. We plan to investigate the robust-
ness of our approach based on different data sets. This
would also strengthen our claim on the obliviousness
of our approach to the source of data.
A 10-fold cross validation is used without loss of
generality. More parsimonious use of the data is pos-
sible. Indeed we would study this with different fea-
tures based on cognitive aspects. The cascading pro-
gression of including the features is based on the fol-
lowing principle. The lexical and syntactic features
are gathered in a straightforward manner during pars-
ing. The semantic features require processing with
WordNet synsets. This stage is more time consum-
ing. Finally, we have only considered few elementary
pragmatic features.
5 COMPARISON WITH
EXISTING APPROACH IN
LITERATURE
We compare our results with (Gonz´alez-Ib´anez et al.,
2011). We do not have access to their data set for the
classification experiments. However, we performed
experiments on a similar genre (Twitter data ) as
used in their experiments. Their model is able to get
67.83% in Sar-vs-Pos and 68.67% in Sar-vs-Neg. Our
results (about 74% accuracy) are better than the re-
sults of (Gonz´alez-Ib´anez et al., 2011).
6 SIGNIFICANCE OF THE
FEATURES
We determined the significance of features from the
results of the classification exercises. These features
play a key role in distinguishing between sarcastic
and non-sarcastic sentences. We describe these sig-
nificant features with respect to the Quotation data set.
KDIR2014-InternationalConferenceonKnowledgeDiscoveryandInformationRetrieval
422
1. verb verb:- This feature is the combination of
two syntactic entities: verb followed by verb. The
probability of first word being an auxiliary verb is
very high in our finding.
(e.g.,”If your life is all about screwing things and
getting(VBG) hammered(VBN),then congratula-
tions, you are a fool)”.
In the above example “getting hammered” has
same syntactic constructs which we want in this
feature.
2. a
j n v:- It is combination of 4 syntactic entities:
adverb followed by adjective followed by noun
followed by verb.
(e.g., If u want to look thinner, hang around peo-
ple fatter than you..”).
3. j
n a v:- It is combination of 4 syntactic entities:
adjective followed by noun followed by adverb
followed by verb.
(e.g.,“If you do not want a sarcastic answer, then
do not ask a stupid question!”).
4. politeness rating:- This feature considers the
words like: “too”,or “absolutely” etc used be-
fore a positive sentiment word. (e.g.,“You are so
clever that sometimes you don’t understand a sin-
gle word of what you are saying”.
5. pos
neg pairs:- This takes care of the case when
there are both positive and negative kinds of
words used. (e.g.,“I love being ignored”). In
this example “love” is positive sentiment show-
ing word, where “ignored” is negative sentiment
word.
In addition, We describe some of the features which
occur with nearly same probabilities in the sarcastic
sentences. They also play a vital role in classifying
sarcastic utterances.
sar
sign:- When the sentiment of the sentence
differs from the emoticons, similes or words
which shows emotion like “wow”, “yippie” etc.
a v:- This feature is for the combination of adverb
followed by verb.
aj:- This feature is for the combination of adverb
followed by adjective.
n
j v:- This feature is for the combination of noun
followed by adjective followed by verb.
7 HUMAN EVALUATION
For checking the credibility of our algorithm we con-
ducted human evaluation of the Quotation data. We
took the sentences in which most of the human judges
agreed. We acknowledge the help of volunteers(4
high school teachers and 3 postgraduate students) in
providing their input on data. There is a reasonable
agreement (i.e., at least 4 of the 8 respondents gave
scores above 6 in a scale of 0 to 10) on more than 75
of the 100 sample quotes provided for human evalua-
tion. We use this collection of 75 quotes as our gold
standard for computer experiments.
We applied our algorithm on this data and found
that our algorithm predicts 75% of these as sarcastic.
This is a reasonable agreement considering, the con-
clusions of (Gonz´alez-Ib´anezet al., 2011), that human
participants differ in recognizing sarcasm.
8 CONCLUSIONS
Use of sarcasm and other figures of speech are very
common in our daily life. However automatic pro-
cessing is a big challenge. We have introduced a new
technique for recognizing the sarcastic utterances on
the basis of semantics and lengths of sentiment pro-
gressions. We show that distinguishing between sar-
castic and a negative sentiment statement is feasible
up to 75% accuracy. We propose a set of features that
may not yet lead to the best distinguisher. We can im-
prove the performance further based on the integra-
tion of linguistic and cognitive features discussed in
the introduction section, in the course of our ongoing
work. The framework we are developing will provide
further semantic, cognitive features, such as struc-
tural or semantic distance between words or phrases
conveying opposition and incongruity.
REFERENCES
Esuli, A. and Sebastiani, F. (2006). Sentiwordnet: A pub-
licly available lexical resource for opinion mining. In
Proceedings of LREC, volume 6, pages 417–422.
Gonz´alez-Ib´anez, R., Muresan, S., and Wacholder, N.
(2011). Identifying sarcasm in twitter: a closer look.
In Proceedings of the 49th Annual Meeting of the As-
sociation for Computational Linguistics: Human Lan-
guage Technologies: short papers-Volume 2, pages
581–586. Association for Computational Linguistics.
Hall, M., Frank, E., Holmes, G., Pfahringer, B., Reutemann,
P., and Witten, I. H. (2009). The weka data min-
ing software: an update. ACM SIGKDD explorations
newsletter, 11(1):10–18.
Lakoff, G. and Johnson, M. (2008). Metaphors we live by.
University of Chicago press.
Liebrecht, C., Kunneman, F., and van den Bosch, A. (2013).
The perfect solution for detecting sarcasm in tweets#
not.
SarcasmDetectionusingSentimentandSemanticFeatures
423
Miller, G. A. (1995). Wordnet: a lexical database for en-
glish. Communications of the ACM, 38(11):39–41.
Reyes, A., Rosso, P., and Buscaldi, D. (2012). From hu-
mor recognition to irony detection: The figurative lan-
guage of social media. Data & Knowledge Engineer-
ing, 74:1–12.
Riloff, E., Qadir, A., Surve, P., De Silva, L., Gilbert, N.,
and Huang, R. Sarcasm as contrast between a positive
sentiment and negative situation.
Toutanova, K. and Manning, C. D. (2000). Enriching the
knowledge sources used in a maximum entropy part-
of-speech tagger. In Proceedings of the 2000 Joint
SIGDAT conference on Empirical methods in natural
language processing and very large corpora: held in
conjunction with the 38th Annual Meeting of the As-
sociation for Computational Linguistics-Volume 13,
pages 63–70. Association for Computational Linguis-
tics.
Tsur, O., Davidov, D., and Rappoport, A. (2010). Icwsm-
a great catchy name: Semi-supervised recognition
of sarcastic sentences in online product reviews. In
ICWSM.
Utsumi, A. (2004). Stylistic and contextual effects in irony
processing. In Proceedings of the 26th Annual Meet-
ing of the Cognitive Science Society, pages 1369–
1374.
KDIR2014-InternationalConferenceonKnowledgeDiscoveryandInformationRetrieval
424