INTRODUCTION

Anemotion is a particular feeling that characterizes a state of mind, such as joy, anger, love,
fear and so on. Automatic emotion detection from text has attracted growing attention due to
its potentially useful applications. For examples, psy chologists can better assist their
patients by analyzing their session transcripts for any subtle emotions; reliable emotion
detection can help develop powerful human-computer inter action devices; and deep
emotional analysis of public data such as tweets and blogs could reveal interesting insights
into human nature and behavior. Many current approaches to emotion detection are based
on supervised learning methods, in which a large set of an notated data (where text has
been labeled with emotions) is needed to train the model. Although the supervised learning
based methods can achieve good results, the availability of large annotated data sets is very
low and a model trained on one domain does not translate well to another. There are some
methods that do not use supervised learning. However, most of these methods use manually
designed dictionaries of emotion keywords. A problem with such an affect lexicon-based
method is that the number of emotion categories is fixed and limited in the dictionary.
Another problem is that if a sentence expresses emotion using words that do not appear in
the dictionary, then it would be considered to be unemotional. For example, the sentence
“Izzy got lots of new toys for her first birthday” conveys quite a happy feeling despite not
containing any Aijun An Department of Computer Science and Engineering York University,
Toronto, Canada aan@cse.yorku.ca obvious happy keywords such as joy, glad, etc. Affect
lexicon-dependent techniques may fail to detect emotions from such sentences. There are
also methods that rely on linguistic rules, but designing such rules is not a trivial task.
Moreover, most of these rules have not been made publicly available. In addition, most
current emotion detection methods look at individual words without considering the context a
word is in. However, a word can invoke different emotions in different contexts. We propose
a novel unsupervised context-based emotion detection method that does not rely on any
affect dictionaries or annotated training data. Therefore, the approach is not restricted to a
fixed number of emotion categories. We start with a small set of representative words which
are used to compute an emotion vector of an affect bearing word by calculating the semantic
relatedness score between this word and an emotion concept. To fine tune the emotion
vectors, the context of the word is considered using three types of syntactic dependencies.
Extensive evaluation of our framework shows promising results. The rest of the paper is
organized as follows. In the next section we present a literature survey of textual emotion
detection. Section 3 describes the details of our proposed algorithm. An extensive set of
experiments that evaluate the performance of our approach is presented in Section 4.
Finally, we conclude the paper and discuss some future avenues of research work

INTRODUCTION
Detecting emotional state of a person by analyzing a text document written by him/her
appear challenging but also essential many times due to the fact that most of the times
textual expressions are not only direct using emotion words but also result from the
interpretation of the meaning of concepts and interaction of concepts which are described in
the text document. Recognizing the emotion of the text plays a key role in the
human-computer interaction [1]. Emotions may be expressed by a person’s speech, face
expression and written text known as speech, facial and text based emotion respectively.
Sufficient amount of work has been done regarding to speech and facial emotion recognition
but text based emotion recognition system still needs attraction of researchers [14]. In
computational linguistics, the detection of human emotions in text is becoming increasingly
important from an applicative point of view. The paper is organized as following. Section 2
describes the methods used for text based emotion detection, which is classified into
keyword spotting technique, lexical affinity method, learning based method and hybrid
approach along with the limitations of these existing methods. A proposed architecture which
contains the emotion ontology and emotion detector algorithm is explained in Section 3.
Based on this, a system is designed for emotion detection from text documents. Finally
conclusion is given in Section 4.

Introduction
Emotion plays a critical role in our daily performance affect ing many aspects of our lives
including social interaction, behavior, attitude, and decision-making [1]. Understanding
human emotion patterns and how the people feel plays an essential role in various
applications such as public health and safety, emergency response, and urban planning.
Text is a particularly important source of data for detect ing emotion because the bulk of
textual data ranging from microblogs, emails, to SMS messages on a smart phone that has
become increasingly available. The rapid growth of emotion-rich textual data makes a
necessity to automate identification and analysis of people’s emotion expressed in text [1]

Introduction
The emotion recognition is the process of identifying human emotions, a task that is
automatically carried out by humans considering facial and verbal expressions, body
language, etc. However, this is a challenging task for an automatic system. In recent years,
the great amount of multimedia information avail able due to the extensive use of the
Internet and social media, along with new computational methodologies related to ma chine
learning, have led to the scientific community to put a great effort in this area [1, 2]. Emotion
recognition from speech signals relies on a num ber of short-term features such as pitch,
vocal tract features such as formants, prosodic features such as pitch loudness, speak ing
rate, etc. Surveys on databases, classifiers, features and classes to be defined in the
analysis of emotional speech can also be found in [3]. Regarding methodology, statistical
anal ysis of feature distributions has been traditionally carried out. Classical classifiers such
as the Bayesian or Super Vector Ma chines (SVM) have been proposed for emotion features
from speech. The modelofcontinuousaffective dimensions is also an emerging challenge
when dealing with continuous rating emo tion labelled during real interaction [1, 4]. In this
work, recur rent neural networks have been proposed to integrate contextual information and
then predict emotion in continuous time using a three-dimensional emotional model. Speech
transcripts have also been demonstrated to be a powerful tool to identify emotional states
[5]. Over the last decade, there has been considerable work in sentiment analy sis [6].
Moreover, the detection of emotions such as anger, joy, sadness, fear, surprise, and disgust
have also been addressed [7]. However, spoken language is informal and provides infor
mation in an unstructured way so that developing tools to select and analyse sentiments,
opinions, etc. is still a challenging topic [8]. In early systems dealing with emotion detection
in text, knowledge-based approaches were applied making use of emo tion lexicons, such as
Sentiwordnet [9]. Other methods, employ machine learning based approaches [10], where
statistical clas sifiers are trained using large annotated corpora and the emotion detection
can be seen as a multi-label classification problem. In this work we propose to use neural
networks to solve the regres sion problem given the 3-dimensional emotional model. The
main contribution of this work is the appropriate selec tion of a neural network architecture
for emotion detection con sidering both acoustic signals and their corresponding transcrip
tion. Additionally, the proposed architecture has been adapted to the 3-dimensional VAD
(Valence, Arousal and Dominance) emotional model [11]. Section 2 describes the two
proposed approaches for the au tomatic emotion recognition from used features to the
common network architectures basics. Experiments carried out are fully described in Section
3. Section 3.1 aims to explain the diffi culties to find out a Spanish corpus and it has led us to
create our own corpus. Section 3.2 mentions the used baselines meth ods and which
measure has been used for testing. Section 3.3 shows the experiments carried out under the
regression prob lem of emotional status with acoustic features whereas Section 3.4 deals
with the experiments achieved at the regression prob lem of emotional status with language
features. Finally some concluding remarks are reported in Section 4

1 Introduction
This survey describes recent works in the field of emotion or affect detection from text.
Emotion detection is part of the broader area of Affective Computing with aims to enable
computers recognize and express emotions [Picard 1997]. Current affect detection systems
are with respect to individual modalities or channels, such as face, voice and text [Calvo
2010]. In this survey, we have focused on reviewing works about emotion detection from
text. Emotion detection and analysis has been widely researched in neuroscience,
psychology and behavior science, as they are an important element of human nature. In
computer science, this task has also attracted the attention of many researchers, especially
in the field of human computer interactions [Strapparava 2008]. In computational linguistics,
the detection of emotion states of a person by analyzing a text document written by him/her
can have many applications in different fields, such as in e-learning environment [Rodriguez
2012] or suicide prevention [Desmet 2013, Vaassen 2014]. For this reason, we decided to
develop a survey about emotion detection systems from text and make it available to
researcher community. in accordance with the emotional model and the approach used. 1
Introduction This survey describes recent works in the field of emotion or affect detection
from text. Emotion detection is part of the broader area of Affective Computing with aims to
enable computers recognize and express emotions [Picard 1997]. Current affect detection
systems are with respect to individual modalities or channels, such as face, voice and text
[Calvo 2010]. In this survey, we have focused on reviewing works about emotion detection
from text. Emotion detection and analysis has been widely researched in neuroscience,
psychology and behavior science, as they In this survey, we classify the most relevant
emotion detection works in accordance with the emotional model and the approach used. A
numerical comparison is not possible since each work used different data sets to evaluate
their systems. Regarding the search strategy used in the survey, we have looked for all of
papers related to emotion detection from text in different research databases like Scopus1 or
IEEE Xplore2. Later on, we have reviewed the papers obtained of these databases and have
selected the best papers that use lexical approach or machine learning approach in their
emotion detection systems. The selection criterion used is based on the relevance of each
work in the field of Affective Computing. This paper is organized as follows. In section 2,
describes the emotional models. Section 3, the different computational approaches for
emotion detection is described. Finally, in section 4, we express our conclusions about this
survey.