Knowledge-Based Approaches to
Concept-Level Sentiment Analysis
Enhanced SenticNet
with Affective Labels
for Concept-Based
Opinion Mining
Soujanya Poria, Jadavpur University
Alexander Gelbukh, National Polytechnic Institute
Amir Hussain, University of Stirling
Newton Howard, Massachusetts Institute of Technology
SenticNet 1.0 is one
of the most widely
used, publicly
available resources
Dipankar Das, National Institute of Technology
Sivaji Bandyopadhyay, Jadavpur University
T
he Internet contains important information on its users’ opinions
and sentiments. The extraction of such unstructured Web data
for concept-based
is known as opinion mining and sentiment analysis, a recent and explo-
opinion mining.
sively growing research field widely used for marketing, customer service,
The presented
financial market prediction, and other
purposes.
Researchers have developed several lexical resources—including WordNet-Affect
( W NA) 1 and S entiWord Net 2 —for different opinion-mining tasks, but most are
incomplete and noisy (see the “Lexical
Resources” sidebar). In particular, SenticNet3
is a concept-based resource containing
5,732 single- or multiword concepts along
with a quantitative polarity score in the
range from −1 to +1; example concepts and
scores include “aggravation” (−0.925), “accomplish goal” (+0.967), “and December”
(+0.111).
methodology
enriches SenticNet
concepts with
affective information
by assigning an
emotion label.
2
However, it’s often desirable to have a
more complete resource containing affective labels for the concepts as well as polarity scores. Currently, the main lexical
resource for emotion detection in natural
language texts is WNA, which assigns emotion labels to words—for example, “wrath”
(anger), “nausea” (disgust), and “triumph”
(joy). WNA, however, has a small number
of words and gives only qualitative affective
informationan emotion labelbut not
quantitative information on the emotion’s
intensity.
Our goal was to create a resource resulting from automatically merging SenticNet
1541-1672/13/$31.00 © 2013 IEEE
IEEE INTELLIGENT SYSTEMS
Published by the IEEE Computer Society
IS-28-02-Gelbukh.indd 2
5/22/13 5:12 PM
Lexical Resources
A
ll known approaches to opinion mining crucially
depend on the availability of adequate lexical resources that provide emotion-related information.
Several methods that employ semiautomatic building have
been suggested, both for English1 and other languages. 2,3
SentiWordNet or WordNet-Affect (WNA) are the most
widely used resources. They contain a rather small number
of words and are mostly limited to affective information for
single words.
Recent research 4 shows that concept-based sentiment
analysis and opinion mining outperform word-based methods. This concept-focused approach relies on polarity and affective information for commonsense knowledge concepts—
for example, to accomplish a goal, experience a bad feeling,
and celebrate a special occasion—which are often used to
express viewpoints and affect.
SenticNet is a publicly available, affective, commonsense resource for sentic computing,5 a new paradigm that
exploits artificial intelligence, the Semantic Web, and af­
fective computing techniques to better recognize, interpret, and process natural language opinions over the Web.
SenticNet was developed through the ensemble applica­
tion of graph-mining and dimensionality-reduction techniques over multiple commonsense knowledge bases.6
It has been exploited to develop applications in fields such
and W NA—that is, our resource
would contain both SenticNet polarity scores and WNA emotion categories for the concepts. Because WNA’s
vocabulary is almost a subset of that
of SenticNet, our task was to automatically extend the emotion labels
from WNA to the rest of SenticNet’s
vocabulary. The obtained resource
is an extended version of WNA, and
users can apply our methodology to
extend similar resources that provide
emotion labels for concepts.
Here, we describe a methodology to
automatically assign emotion labels
to SenticNet concepts. We trained a
classifier on the subset of SenticNet
concepts present in WNA. As classification features, we used several concept similarity measures, as well as
various psychological features available in an emotion-related corpus.
Of the various classifiers we tried, we
obtained the highest accuracy using
the classifier called Support Vector
Machine (SVM). This work lays the
foundation for a new concept-level
march/april 2013
IS-28-02-Gelbukh.indd 3
as social media marketing, human-computer interaction, and
e-health.
Although SenticNet is much larger than WNA, it doesn’t
provide the specific emotion labels for the concepts. In our
work, we fill this gap by extending WNA’s labels to other
SenticNet concepts.
References
1. S. Mohammad and P.D. Turney, “Emotions Evoked by Common
Words and Phrases: Using Mechanical Turk to Create an Emotion Lexicon,” Proc. NAACL-HLT Workshop Computational Approaches to Analysis and Generation of Emotion in Text, Association for Computational Linguistics, 2010, pp. 26–34.
2. P. Arora, A. Bakliwal, and V. Varma, “Hindi Subjective Lexicon
Generation Using WordNet Graph Traversal,” Int’l J. Computational Linguistics and Applications, vol. 3, no. 1, 2012, pp. 25–39.
3. A. Wawer, “Extracting Emotive Patterns for Languages with
Rich Morphology,” Int’l J. Computational Linguistics and Applications, vol. 3, no. 1, 2012, pp. 11–24.
4. E. Cambria and A. Hussain, Sentic Computing: Techniques, Tools,
and Applications, Springer, 2012.
5. R. Picard, Affective Computing, MIT Press, 1997.
6. E. Cambria, D. Olsher, and K. Kwok, “Sentic Activation: A
Two-Level Affective Common Sense Reasoning Framework,”
Proc. Assoc. Advancement of Artificial Intelligence (AAAI),
AAAI, 2012, pp. 186–192.
opinion-mining resource, and presents new features and measures for
generating such resources.
Our Method
There are many uses for WNA, ranging from social, economic, and commercial uses to health care, political,
and governmental uses. For example, companies need to know what
emotions customers express in product reviews so that they can improve
their products, which translates into
better income for the businesses and
better quality of life for the consumers. A political party or government
needs to know the emotions prevailing in the press related to its actions. The improvements in decision
making that result from such analysis represent a more efficient democracy. This would help provide democracy in real time, in contrast to using
elections as a means of punishing
bad government or giving credit to a
party’s promises. In all such applications, a larger resource will give more
www.computer.org/intelligent
precise and reliable results, because
more words in the analyzed texts will
contribute more emotion labels to the
statistics.
We find it interesting to view our
obtained resource as SenticNet augmented with affective labels. Such a
resource will give rise to a range of
novel applications combining the polarity information from SenticNet
with the affective information that
we added. In this way, a company
could obtain information on which
products or features customers like
and don’t like (the polarity), and on
the specific emotions customers feel
relating to the products (the affect)—
for example, are they surprised, angered, or joyful?
Further, we can use polarity information as a degree measure of the
corresponding emotion. Although
both “sulk” and “offend” have the
same emotion label of anger in WNA,
“sulk” has polarity −0.303 in SenticNet and “offend” has −0.990. Therefore, “offend” is stronger in terms of
3
5/22/13 5:12 PM
Knowledge-Based Approaches to Concept-Level
Sentiment Analysis
the anger affect than “sulk.” This information is important for weighted
emotion detection. For example, consider the following review of a mobile
phone: “the keyboard is comfortable
[+0.120, joy] and the interface is amicable [+0.214, joy], but the color is
queasy [−0.464, disgust].” Although
the review contains two joy words
and only one disgust word, weighting
the labels by the polarity score indicates that the customer’s main emotion was disgust.
Lexical Resources
Our work’s primary aim is to assign
WNA emotion labels to SenticNet’s
concepts. For this, we use a supervised
machine-learning approach. The intuition behind this approach is that
words that have similar features are
similar in their use and meaning and,
in particular, are related with similar
emotions. This lets us extend WNA’s
emotion labels for known seed words
to words absent from WNA that share
features with the seed words.
To achieve this, we must select linguistically relevant features, which
we extracted from various relevant
text corpora and dictionaries. In this
section, we describe the corresponding lexical resources. In addition,
we used standard resources such as
WordNet 3.0.
SenticNet
As a target lexicon and source of polarity information for our polaritybased, concept-similarity measure,
we used SenticNet 1.0 (http://sentic.
net/senticnet-1.0.zip), which contains
5,732 concepts, of which 2,690 are
multiword concepts (such as “animate flesh dead person”, “ban Harry
Potter”, and “why happen”). Of the
5,732 SenticNet concepts, 3,303 exist in WordNet 3.0 and 2,429 do not.
Of the latter set, most are multiword
concepts such as “access Internet” or
4
IS-28-02-Gelbukh.indd 4
“make mistake,” except for 68 singleword concepts such as “against” or
“telemarketer.”
WordNet-Affect Emotion Lists
As a training corpus, we used
t h e W N A l i s t s (w w w. c s e .u nt .
e d u / ~ r a d a /a f f e c t i v e t e x t / d a t a /
WordNetAffect­E motionLists.tar.gz)
provided as part of the SensEval
2007 data. This dataset consists of
six word lists corresponding to Ekman’s six basic emotions: anger, disgust, fear, joy, sadness, and surprise.4
Other work provides an overview of
different sets of basic emotions proposed in the literature.5
The dataset contains 606 synsets,
of which all but two are assigned exactly one label each. If the synsets
are broken down into individual concepts, the dataset contains 1,536 concepts. Only 63 concepts are multiword expressions, for example, with
hostility or jump for joy. All but 72
concepts (93 percent) are present in
the SenticNet vocabulary.
ISEAR Dataset
As a source of various features and
corpus-based similarity measures between concepts, we used the International Survey of Emotion Antecedents and Reactions (ISEAR) dataset
(see www.affective-sciences.org/system/
files/page/2636/ ISEAR.zip and www.
affective-sciences.org/researchmaterial).6
The survey, conducted in the 1990s
across 37 countries, consists of short
texts called statements, obtained from
approximately 3,000 respondents who
were instructed to describe a situation in which they felt a particular
emotion. A statement contains 2.37
sentences on average. The dataset
contains 7,666 statements, 18,146 sentences, and 449,060 running words.
For each statement, the dataset
contains 40 numeric or categorical
parameters that give various types
www.computer.org/intelligent
of information, such as respondent
age, emotion felt in the situation described, emotion intensity, and so on.
The majority of these parameters are
numerical scores with a few discrete
values (usually three or four) expressing different parameter degrees.
Preprocessing
and Tokenizing
Many of the features we used were
based on occurrences of the concepts
in the ISEAR dataset. For locating
concepts in the text, we used preprocessing tools from Rapidminer’s text
plug-in (http://rapid-i.com/content/view/
181/190), and for lemmatizing we
used the WordNet lemmatizer. Of
the 5,732 concepts contained in Sentic­
Net, we found 2,729 at least once
in ISEAR, either directly or after
lemmatizing.
Features Used
for Classification
For each S ent icNet concept , we
extracted from ISEAR statistical
features of its occurrences and cooccurrences with other SenticNet concepts in ISEAR statements (we treated
multiword concepts as a single token).
We used two feature types: those
based on the ISEAR dataset parameters and those based on various similarity measures between concepts.
Features Based
on ISEAR Parameters
Some of the 40 parameters provided
for each ISEAR statement—such as
the respondent’s ID parameter—aren’t
informative for our goals. We used
the following parameters:
• background data related to the
respondent (age, gender, religion,
father’s occupation, mother’s occupation, and country);
• general data related to the emotion
felt in the situation described in the
IEEE INTELLIGENT SYSTEMS
5/22/13 5:12 PM
statement (intensity, timing, and
longevity);
• physiological data (ergotropic arousals, trophotropic arousals, and felt
change in temperature);
• expressive behavior data (movement,
nonverbal activity, and paralinguistic activity); and
• emotion felt in the situation described in the statement.
Even for values expressing a parameter’s degree or intensity, we used each
degree value as an independent feature, and the frequency of occurrences
of a given concept under a given para­
meter value as the feature value.
Namely, for each occurrence of each
concept in an ISEAR dataset statement, we extracted the corresponding
parameters from the ISEAR dataset.
Then we aggregated the data for multiple occurrences of the same concept
in the whole corpus in a feature vector
for that concept as the corresponding
frequencies.
Features Based
on Similarity Measures
Another type of classification feature we
used was the similarity values between
concepts. We used two types of similarity measures: those based on lexical
resources, such as SenticNet and WordNet, and those based on concept cooccurrence. The intuition behind similarity-based features is that if the distances from two data points in Euclidian space to a number of other points
are similar, then it’s probable that the
two points are close to each other.
Given a concept, we treated as independent dimensions of its feature
vector the value of each similarity
measure, which we obtained by comparing the given concept with the
other concepts in the vocabulary.
SenticNet score-based similar it y.
We define the distance between two
march/april 2013
IS-28-02-Gelbukh.indd 5
SenticNet concepts a and b as DSN(a, b) =
|p(a) - p(b)|, where p(⋅) is the polarity
specified for these concepts in SenticNet. The similarity is the inverse of
the distance: SimSN (a, b) = 1/DSN(a, b).
the corpus; and p(x,y) is the probability for a sentence to contain both
x and y—that is, the normalized
number of sentences that contain
both x and y.
WordNet distance-based similarity.
Emotional affinit y. We define the
emotional affinity between two concepts a and b in the same way as PMI
but at the level of complete ISEAR
statements—that is, p(x) in Equation 1 is defined as the corresponding
number of statements instead of sentences, normalized by the total number of statements.
PMI often reflects syntactic relatedness of the wordsfor example,
it’s high for a verb and its typical object, or for parts of a frequently found
multiword expression. However,
emotional affinity incorporates a
wider notion of relatedness within the
same real-world situation, as well as
the notions of synonymy and rephrasing. Because ISEAR statements have
strong emotional content and are related with one emotion each, it’s probable that when two words co-occur in
the same ISEAR statement, then the
emotion with which one of them is related is the same as the emotion with
which the other one is related.
We used English WordNet 3.0 to measure the semantic distance between
two words. WordNet::Similarity (www.
d.umn.edu/tpederse/similarity.html)
is an open source package developed at the University of Minnesota
for calculating nine different lexical
similarity measures between pairs of
word senses. This gave us nine different similarity measures.
Because WordNet similarity is defined for specific word senses, for every two concepts found in WordNet,
we defined the corresponding similarity as the maximum similarity between all senses of the first concept
and all senses of the second concept.
For the concepts not found in
WordNet, we set the similarity between such a concept and any other
concept to a random value. An alternative would be to use 0 or some other
not-found category, but this deteriorated the results by making all notfound concepts appear similar to each
other because of the numerous coinciding values in their feature vectors.
Point-wise mutual information. Point-
wise mutual information (PMI) is
a similarity measure based on cooccurrences of two concepts within
the same sentence. For concepts a
and b, it is defined as
SimPMI = log
p(a, b)
,
p(a)p(b) (1)
where p(x) is the probability for a
sentence in the corpus to contain the
concept x—that is, the number of
sentences where x occurs normalized
by the total number of sentences in
www.computer.org/intelligent
ISEAR text-distance similarity. We
used positional information of the
concept tokens in the ISEAR statements to measure the similarity between concepts. For this, we calculated the average minimum distance
between the pairs of tokens of SenticNet concepts that co-occurred in the
ISEAR dataset statements. We defined
the similarity as the inverse of the distance. If the concepts didn’t co-occur
in any statement, then we considered
the similarity between them to be 0.
Classification Procedure
We cast the task as a six-way categorization task, in which we assigned
5
5/22/13 5:12 PM
Knowledge-Based Approaches to Concept-Level
Sentiment Analysis
Table 1. Accuracy obtained with different feature sets, datasets, and classifiers.
Feature set
SenticNet + Word
Net + ISEAR
Total dataset
SenticNet ∩ ISEAR
Total data size
Labeled dataset
2,729
SenticNet + WordNet
SenticNet
5,732
SenticNet ∩ ISEAR ∩ WNA
Labeled data size
1,202
5,732
SenticNet ∩ WNA
1,436
Naive Bayes (%)
71.20
51.23
53.21
Multilayer perceptron (%)
74.12
55.54
52.12
Support Vector Machine (%)
88.64
57.77
59.23
exactly one of the six WNA emotion
labels to each considered concept. We
conducted two sets of experiments.
I n one ex p er i ment we to ok i nto
account the features that relied on
ISEAR, and in the other we didn’t
use those features.
In the experiments that relied on
features based on occurrences or cooccurrences of concepts in the ISEAR
statements, we only used the 2,729
S enticNet concepts fou nd i n t he
ISEAR data (which thus had valid
ISEAR-based features) in further processing, and hence assigned them the
emotion labels. In contrast, in the experiments without ISEAR-based features, we assigned the labels to all
5,732 SenticNet concepts.
Therefore, we constructed feature
vectors for 2,729 and 5,732 SenticNet concepts, respectively. As training and test data, we used the intersection between the corresponding
set of concepts and the WNA vocabulary (for which we had the gold
standard emotion labels); this intersection consisted of 1,202 and 1,436
concepts, respectively.
To evaluate, we randomly divided
the corresponding set of available labeled data into training and test data,
using 66.7 percent of the set for training and 33.3 percent for testing. For
constructing the final resource, we used
100 percent of available labeled data
for training. Finally, we used various
machine-learning algorithms trained
on the training data to derive the labels
for the test and unlabeled data.
6
IS-28-02-Gelbukh.indd 6
Evaluation and Discussion
As a gold standard data for evaluation, we used WNA concepts. We
considered a label to be assigned correctly to a concept if the WNA data
assigned the same label to the same
concept (in a few cases, when the
WNA assigned two labels to a concept, we considered our label assignment correct if our assigned label was
one of those two labels).
Table 1 summarizes our experiments
with different feature sets, datasets,
and classifiers. Using the features derived from SenticNet, WordNet, and
ISEAR, we included in those experiments only the 2,729 concepts in the
intersection of SenticNet and ISEAR
because they had valid ISEAR-based
features. When using feature vectors
not containing features that rely on
the ISEAR data—that is, consisting
only of the SenticNet similarity and
the nine WordNet::Similarity measures between the given concept and
all other concepts in the dataset—
all 5,732 SenticNet concepts were
processed.
Accordingly, as a labeled dataset.
we used the intersection of the total
dataset with WNA; for ISEAR-based
experiments, this set contained 1,202
labeled concepts, and for experiments
not relying on ISEAR, the set contained 1,436 concepts. The SVM classifier showed the best performance,
obtaining an accuracy of 88.64 percent with the ISEAR-based features.
Table 2 shows the corresponding confusion matrix.
www.computer.org/intelligent
As Table 1 shows, ISEAR-based
features are important for accurate
classification in our task. Also, increasing the training dataset improves
accuracy. We thus focus on our results for the corpus of 2,729 concepts obtained with the ISEAR features in the discussion that follows.
Although we experimented with different subsets of features, we observed that similarity-based features
performed better than the ISEAR
parameter-based features.7 Indeed,
similarity measures identify whether
two concepts have similar properties
and thus should be placed in the same
category.
SenticNet-based similarity had a
positive impact. The use of all available features gave the best results
both when we used only similaritybased features and when we took into
account ISEAR parameter-based features.7 This is probably because of
the balancing of the two somewhat
contradicting and complementary information sources: ISEAR parameterbased features provided affective information that helped to choose the
label for those concepts for which this
information was available, and similarity measures indicated whether
different concepts are to be assigned
the same or different labels.
Polarity and Label Agreement
Even when the emotion label assigned
by the algorithm doesn’t coincide exactly with the one present in the goldstandard data, it can share important
properties with the correct label and
thus can be considered correct in a relaxed sense.
Considering joy and surprise as
positive emotions and the rest as
negative—and ignoring the confusion
within the areas marked with dotted
lines in Table 2—we observe a 95.71
percent agreement with the gold standard WNA data on whether the emotion
IEEE INTELLIGENT SYSTEMS
5/22/13 5:12 PM
Table 2. Confusion matrix on the test set of 396 labeled concepts.
Actual
labels
Predicted labels
Surprise
Surprise
Joy
Sadness
Anger
Fear
Disgust
Precision
(%)
Recall
(%)
39
4
-
-
-
1
85
89
Joy
3
91
2
3
1
1
91
90
Sadness
-
3
52
1
-
-
85
93
Anger
2
2
3
72
4
2
90
85
Fear
-
-
3
1
63
2
91
91
Disgust
2
-
1
3
1
34
85
85
is positive or negative. Table 3 gives
examples of agreement and disagreement in label polarity; two labels
agree if both are positive or both are
negative.
When we used the entire set of
1,202 labeled examples as both training and test data, agreement was 96.8
percent. Table 4 shows agreement between our results and the gold standard in affective polarity sign (negative or positive emotion) per emotion
label.
Agreement with
the Hourglass Model
The Hourglass of Emotions8 reinterprets Plutchik’s model by organizing
primary emotions around four dimensions whose different levels of activation make up the total emotional
state of the mind (see Table 5).
Based upon the Hourglass model,
SenticNet 29 aims to assign one primary emotion and one secondary
emotion to each SenticNet concept.
If we map the six emotion labels
used in our research (shown in Table 5
in boldface) to the affective dimensions of the Hourglass model, then
joy and sadness are mapped to the
same dimension, as are anger and
fear. Ignoring these differencesthat
is, ignoring the confusion within the
areas marked with a dashed line in
Table 2we have a 91.16 percent
agreement with the gold standard in
identifying the affective dimensions,
or 93.5 percent when we used all labeled data as both the training and
test sets.
march/april 2013
IS-28-02-Gelbukh.indd 7
Table 3. Examples of agreement and disagreement in label polarity.
Concept
Gold standard (polarity)
Our label (polarity)
Frustration
Anger (negative)
Fear (negative)
Agreement
Yes
Offensive
Disgust (negative)
Anger (negative)
Yes
Triumph
Joy (positive)
Surprise (positive)
Yes
Favor
Joy (positive)
Anger (negative)
No
Bored
Sadness (negative)
Fear (negative)
Yes
Wonder
Surprise (positive)
Joy (positive)
Yes
Table 4. Our results vs. the gold standard (WNA) in affective polarity
sign (negative or positive) per emotion label.
Polarity
Emotion label
Positive
Joy
Agree
Disagree
Total
422
8
430
97
8
105
Anger
198
3
201
Disgust
148
10
158
Fear
123
8
131
Sadness
175
2
177
Total: 1,163
Total: 39
Total: 1,202
Surprise
Negative
Table 5. Primary emotions used in The Hourglass of Emotions. 8
Sentic level
Pleasantness
Attention
Sensitivity
Aptitude
1
Ecstasy
Vigilance
Rage
Admiration
2
Joy
Anticipation
Anger
Trust
3
Serenity
Interest
Annoyance
Acceptance
4
Pensiveness
Distraction
Apprehension
Boredom
5
Sadness
Surprise
Fear
Disgust
6
Grief
Amazement
Terror
Loathing
Agreement with
the SenticNet polarity score
A possible way to indirectly evaluate
our algorithm on non-WNA data is
comparing the polarity of our assigned
www.computer.org/intelligent
labels (considering joy and surprise
as positive and other labels as negative) with the sign of the polarity score given in SenticNet. Table 6
gives some examples.
7
5/22/13 5:12 PM
Knowledge-Based Approaches to Concept-Level
Sentiment Analysis
Table 6. The polarity of our assigned labels vs. the sign used
in SenticNet’s polarity score.
Concept
SenticNet score
Our label (polarity)
Agreement
Better grade
−0.185
Joy (positive)
No
Collect information
−0.309
Joy (positive)
No
Commonsense
+0.588
Surprise (positive)
Yes
Dislike
−0.917
Disgust (negative)
Yes
Efficiency
−0.136
Surprise (positive)
No
Rescue
+0.963
Joy (positive)
Yes
Table 7. Examples of SenticNet concept polarities and labels.
Concept
Polarity
Label
Annoyed
−0.755
Anger
Birthday
+0.309
Joy
December
+0.111
Joy
Feel guilty
−0.540
Sadness
Make mistake
−0.439
Sadness
Weekend
+0.234
Joy
Table 8. Distribution of SenticNet
concepts per emotion label.
Label
Concepts
Anger
421
Disgust
334
Fear
303
Joy
718
Sadness
426
Surprise
527
Total: 2,729
As Table 6 shows, we obtained a
95.30 percent agreement. As in the
cases of the concepts “better grade”
or “efficiency,” sometimes the disagreement can be attributed to a possible problem with the SenticNet score
rather than with our algorithm.
Developed
Resource Statistics
We obtained two resources: one with
the use of ISEAR data and the other
without it. The former set is smaller
(2,729 versus all 5,732 SenticNet
concepts) but more accurate. Table 1
shows the lower bounds on the accuracy of each resource: 88.64 percent
8
IS-28-02-Gelbukh.indd 8
and 59.23 percent, respectively. We
believe that the resource obtained
with the ISEAR dataset is accurate
enough, and still big enough, for
practical applications.
Table 7 shows exa mples f rom
ISEAR’s 2,729 SenticNet concepts
along with their polarity scores and
emotion labels. The label on “annoyed” coincides with the one given
in WNA, while other concepts aren’t
present in WNA, and thus we can’t
directly validate their labels. However, they appear quite plausible:
birthday would fit under joy, weekend for most people is joy, as well as
December (because of Christmas and
other holidays). All three concepts
also have positive labels in SenticNet.
The two concepts with negative polarity, “feel guilty” and “make mistake,” were also plausibly associated
with sadness.
Table 8 shows the distribution
of the concepts per emotion label.
Note that the most numerous classes
are the two positive emotions: joy and
surprise. Probably this can be attributed to the selection of concepts in
the SenticNet vocabulary: SenticNet
www.computer.org/intelligent
contains 3,576 concepts with positive polarity and only 2,221 concepts
with negative polarity, so it seems
natural that the two positive emotions are more numerous in this vocabulary, with joy being more general
and thus more frequent than surprise.
We should also point out that ISEAR
contains nearly the same number of
statements (1,093 to 1,096) for each
of the seven emotions it considered,
of which only one (joy) is positive;
thus there are many more emotionally negative statements in the ISEAR
dataset than emotionally positive.
W
e developed the world’s largest freely available dictionary for opinion mining and sentiment
analysis that contains both quantitative polarit y scores and qualitative affective labels. Namely, we
automatically and accurately supplied SenticNet with affective WNAcompatible labels using a machinelearning algorithm. Our resource,
along with associated programs and
other relevant data, is available for
academic use from www.gelbukh.
com/resources/emo-senticnet.
Our work opens up multiple directions for future research, such as using other types of monolingual or
multilingual 10 corpora as a source
of features to improve the accuracy
or to label more concepts, as well as
the use of more elaborated classification techniques in combination with
syntactic and psychological clues to
improve accuracy. It will be interesting to see how ISEAR information on
gender, country, and so on affects the
results.
Acknowledgments
The work was partially supported by the
governments of India, Mexico, and the UK,
with grants from CONACYT-DST India
122030 (“Answer Validation through Textual Entailment”); Mexico’s CONACYT
IEEE INTELLIGENT SYSTEMS
5/22/13 5:12 PM
The Authors
50206-H, ICYT PICCO10-120, and SIPIPN 20131702; and the UK’s Engineering
and Physical Sciences Research Council
(EPSRC) grant EP/G501750/1 (“Common
Sense Computing to Enable the Development of Next-Generation Semantic Web
Applications”).
References
1.C. Strapparava and A. Valitutti,
“WordNet-Affect: An Affective Extension of WordNet,” Proc. 4th Int’l Conf.
Language Resources and Evaluation,
European Language Resources Association, 2004, pp. 1083–1086.
2.A. Esuli and F. Sebastiani, “SentiWordNet: A Publicly Available Lexical Resource for Opinion Mining,” Proc. Int’l
Conf. Language Resources and Evaluation, European Language Resources
Association, 2006, pp. 417–422.
3.E. Cambria et al., “SenticNet: A Publicly Available Semantic Resource for
Opinion Mining,” Proc. AAAI Common Sense Knowledge, Assoc. Advancement of Artificial Intelligence (AAAI),
2010, pp. 14–18.
4.P. Ekman, “Facial Expression and Emotion, Am. Psychologist, vol. 48, no. 4,
1993, pp. 384–392.
5.A. Ortony and T.J. Turner, “What’s
Basic About Basic Emotions?” Psychological Rev., vol. 97, no. 3, 1990,
pp. 315–331.
6.K.R. Scherer, “What Are Emotions?
And How Can They Be Measured?”
Social Science Information, vol. 44,
no. 4, 2005, pp. 693–727.
march/april 2013
IS-28-02-Gelbukh.indd 9
Soujanya Poria is a graduate student at the Department of Computer Science and Engineering of Jadavpur University, India. His research interests include natural language
processing and machine learning. Poria has a BE in computer science and engineering
from Jadavpur University. Find his contact details at http://soujanyaporia.in.
Alexander Gelbukh is a research professor at the Computing Research Center of the
National Polytechnic Institute, Mexico. His research interests include natural language processing. Gelbukh has a PhD in computer science from the All-Russian Institute of Scientific and Technical Information. Find his contact details at http://nlp.cic.ipn.
mx/~gelbukh.
Amir Hussain is a professor of computing science at the University of Stirling, UK. His
research interests include computational intelligence and cognitive computing technology and applications. Hussain has a PhD in electronic and electrical engineering from the
University of Strathclyde in Glasgow. Find his contact details at www.cs.stir.ac.uk/~ahu.
Newton Howard is one of the directors of the Synthetic Intelligence Project, a resident
scientist at the Massachusetts Institute of Technology, and a national security advisor to
several US government organizations. Howard has a PhD in cognitive informatics and
mathematics from La Sorbonne. Find his contact details at http://web.mit.edu/nhmit/
www/index.html.
Dipankar Das is an assistant professor at the Department of Computer Science and Engineering, National Institute of Technology (NIT), Meghalaya, India. His research focuses
on natural language processing. Dar has a PhD in computer science from Jadavpur University. Find his contact details at http://dasdipankar.com.
Sivaji Bandyopadhyay is a professor at the Department of Computer Science and En-
gineering of Jadavpur University, India. His research interests include natural language
processing and machine learning. Bandyopadhyay has a PhD in computer science from
Jadavpur University. Find his contact details at www.sivajibandyopadhyay.com.
7.S. Poria et al., “Merging SenticNet
and WordNet-Affect Emotion Lists for
Sentiment Analysis,” Proc. 11th Int’l
Conf. Signal Processing, IEEE, 2012,
pp. 1251–1255.
8.E. Cambria, A. Livingstone, and A. Hussain, The Hourglass of Emotions, LNCS
7403, Springer, 2012, pp. 144–157.
9.E. Cambria, C. Havasi, and A. Hussain,
“SenticNet 2: A Semantic and Affective Resource for Opinion Mining and
Sentiment Analysis,” Proc. Florida
www.computer.org/intelligent
Artificial Intelligence Research Soc. Conf.
(FLAIRS), AAAI, 2012, pp. 202–207.
10.G. Sidorov et al., “A New Combined
Lexical and Statistical Based Sentence
Level Alignment Algorithm for Parallel
Texts,” Int’l J. Computational Linguistics and Applications, vol. 2, nos. 1–2,
2011, pp. 257–263.
Selected CS articles and columns
are also available for free at
http://ComputingNow.computer.org.
9
5/22/13 5:12 PM