IEEE Computational Intelligence Magazine - February 2020 - 65

Emotion analysis [1] in natural language processing (NLP) targets to automatically extract the emotional state of
a user through his/her writing (tweets,
post, blogs, etc.). Ekman [2] studied the
human emotion behavior in details and
categorized them into six basic human
emotions. According to him, basic emotions are anger, fear, surprise, sadness,
joy and disgust. Comparatively, the aim
of sentiment analysis is to predict the
polarity orientation (e.g., positive, negative, neutral or conflict) in the user-written
texts [3]. Coarse-grained sentiment analysis (document or sentence level) usually
ignores critical information toward a
target. In fine-grained sentiment analysis
[4], we can emphasize on a target without losing any critical information.
Sentiment and emotions are closely
related. Emotions are usually shorter in
duration, whereas sentiments are more
stable and valid for a longer period of
time [5]. Sentiments are also normally
expressed toward a target entity, whereas
emotions are not always target-centric [6].
Table I depicts example scenarios for
both the problems. In the first example,
emotion 'joy' is derived from the phrase
'died from laughter' which is also very intense. However, the emotion associated
with the second example which contains
similar phrase 'died from cancer' is 'sadness'.
In such scenario predicting the correct
emotion is often very challenging and
non-trivial.The third and fourth examples
reflect emotion classes 'fear' and 'anger'
derived from the respective phrases 'Still
salty' and 'revenge'. In sentiment analysis
problem, the first example expresses 'posi-

tive' sentiment whereas the second example has 'negative' sentiment for their
respective targets, i.e., WTS and Lloyds.
In general, emotion analysis and sentiment analysis classify a text into one of
the predefined classes (e.g., joy, fear, etc.
for emotion and positive, negative, etc. for
sentiment). However, predicted opinion
or sentiment class of a text does not
carry the finer information such as the
exact state of mood or opinion of a user.
Level or intensity of the expressed emotions or sentiments often differs on a
case-to-case basis within a single class.
For example, some emotions are comparatively gentler than the others (e.g.,
'not good' versus 'terrible'). Emotion
expressed by both the phrases is anger,
however, the phrase 'not good' expresses
relatively mild emotion, whereas the
phrase 'terrible' is much severe.
Similarly, both phrases 'its fine' and 'its
awesome' carry positive sentiment but
express different level of sentiments.
Sentiment of the latter case is strong,
whereas the sentiment of the earlier case
is comparatively weak. Thus, measuring
the degree of emotion is of paramount
importance in analyzing the finer-level
details of the expressed emotions and
sentiments. Such analysis has wide realworld applications such as big social data
analysis for business intelligence [9],
stock market prediction [10], healthcare
[11], recommendation systems [12], etc.
In this paper, we propose a multi-layer
perceptron (MLP) based ensemble technique for solving two different problems,
i.e., emotion analysis and fine-grained sentiment analysis. We aim to identify the

intensities of emotions and sentiments,
respectively for the two tasks. For emotion
analysis, we employ generic tweets, whereas for sentiment analysis our target domain
is financial text. At first, we develop a support vector regression (SVR) [13] based
feature-driven system and three deep
learning systems, namely a convolutional
neural network (CNN) [14], a long shortterm memory (LSTM) network [15] and a
gated recurrent unit (GRU) network [16]
for the intensity prediction. In the second
step, we combine the outputs of these
systems via the MLP network. The final
output obtained from this combined
model is better as compared to the individual models. We further perform a series of
normalization heuristics to minimize the
noise. The normalized text has a higher
degree of readability than un-normalized
text, thus making it a better candidate to
find more representative word embeddings.
The current work follows one of our
previous works [17] on intensity prediction. However, our current research significantly differs from earlier work w.r.t.
the following points: a) Our previous
work [17] addressed only financial sentiment analysis task, whereas in the current
work we also focus on emotion intensity prediction for the four emotions, i.e.,
'anger', 'fear', 'joy' and 'sadness'. Please note
that intensity prediction of emotion is
completely a different task; b) We include
several features for training and testing of
the classifier; c) We incorporate various
normalization heuristics to address the
noisy text; d) We present a detailed analysis of the obtained results w.r.t. various
state-of-the-art and traditional techniques;

TABLE I Examples of Emotion and Sentiment analysis. Intensity values reflect the degree of emotion/sentiment in the respective
text. Examples are taken from the respective datasets [7], [8].
EMOTION ANALYSIS (0: NO EMOTION, 1: HIGH EMOTION)
TEXT

DOMAIN

EMOTION

INTENSITY

JUST DIED FROM LAUGHTER AFTER SEEING THAT.

TWITTER

JOY

0.92

MY UNCLE DIED FROM CANCER TODAY...

SADNESS

0.87

STILL SALTY ABOUT THAT FIRE ALARM AT 2AM THIS MORNING.

FEAR

0.50

HAPPINESS IS THE BEST REVENGE

ANGER

0.25

SENTIMENT ANALYSIS (−1: EXTREMELY NEGATIVE, +1: EXTREMELY POSITIVE)
TEXT

DOMAIN

TARGET

SENTIMENT

INTENSITY

BEST STOCK: $WTS +15%

MICROBLOGS

WTS

POSITIVE

0.857

UK GOVERNMENT CUTS STAKE IN LLOYDS TO BELOW 11 PCT

NEWS HEADLINE

LLOYDS

NEGATIVE

− 0.596

FEBRUARY 2020 | IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE

IEEE Computational Intelligence Magazine - February 2020

Table of Contents for the Digital Edition of IEEE Computational Intelligence Magazine - February 2020