A Review on Text-Based Emotion Detection - Techniques,

Applications, Datasets, and Future Directions

Sheetal Kusal1 · Shruti Patil2 · Jyoti Choudrie3 · Ketan Kotecha2 · Deepali Vora1 ·
Ilias Pappas4
© The Author(s) 2022

Abstract

AI has been used for processing data to make decisions, interact with humans, and understand
their feelings and emotions. With the advent of the internet, people share and express their
thoughts on day-to-day activities and global and local events through text messaging
applications. Hence, it is essential for machines to understand emotions in opinions, feedback,
and textual dialogues to provide emotionally aware responses to users in today's online world.
The field of text-based emotion detection (TBED) is advancing to provide automated
solutions to various applications, such as businesses, and finances, to name a few. TBED has
gained a lot of attention in recent times. The paper presents a systematic literature review of
the existing literature published between 2005 to 2021 in TBED. This review has
meticulously examined 63 research papers from IEEE, Science Direct, Scopus, and Web of
Science databases to address four primary research questions. It also reviews the different
applications of TBED across various research domains and highlights its use. An overview of
various emotion models, techniques, feature extraction methods, datasets, and research
challenges with future directions has also been represented.

1 Symbiosis Institute of Technology, Symbiosis International (Deemed University), Pune, Maharashtra, India.

2 Symbiosis Centre for Applied Artificial Intelligence (SCAAI), Symbiosis Institute of Technology, Symbiosis
International (Deemed University), Pune, Maharashtra, India

3 University of Hertfordshire, Hatfield, Hertfordshire, United Kingdom,

4 University of Agder, Norway.

1 Introduction

Artificial Intelligence (AI) is the part of computer science that focuses on designing
intelligent computer systems that show the traits we re-late with human intelligence like
comprehending languages, learning problem-solving, decision making, etc. One of the
significant contributions of AI has remained in Natural Language Processing (NLP), which
glued together linguistic and computational techniques to assist computers in understanding
human languages and facilitating human-computer interaction. Machine Translation,
Chatbots or Conversational Agents, Speech Recognition, Sentiment Analysis, Text
summarization, etc., fall under the active research areas in the domain of NLP. However, in
the past few years, Sentiment analysis has become a demanding realm. Nowadays, Artificial
Intelligence has spread its wings into Thinking Artificial Intelligence and Feeling Artificial
Intelligence (Huang and Rust 2021). Figure 1 shows the sub domain of artificial
intelligence. Thinking AI is de-signed to process information in order to arrive at new
conclusions or decisions. The data are usually unstructured. Text mining, speech
recognition, and face detection are all examples of how thinking AI can identify patterns
and regularities in data. Machine learning and deep learning are some of the recent
approaches to how thinking AI processes data. Existing decision-making applications
include IBM Watson - a question-answering computer system, recommender systems, and
expert systems. Feeling AI is developed for two-way human-machine interactions as well as
the analysis of human feelings and emotions. Sentiment analysis, text-to-speech technology,
chatbots that mimic human dialogue, and robots with special technology for detecting
emotional signals are some of the recent technologies in feeling AI. "Feeling AI" refers to
artificial intelligence (AI) that measures, understand, simulates, and reacts to human
emotions. Sentiment Analysis is the area that comes un-der the umbrella of feeling AI.
Emotion Detection, the adjunct of Sentiment Analysis, is intended to extract and analyze
fine-grained emotions such as happiness, anger, sadness, etc. In contrast, Sentiment analysis
is targeted to interpret opinions and give polarities like positive, negative, or neutral by
understanding the feelings (Francisca et al. 2020). Emotions have remained important in
Human-Computer Interaction and enable computer systems to understand human conduct or
be more emotionally aware systems. So, technology-driven by emotions plays a critical role
in decision-taking, which could be helpful to the broader range of application domains. This
more comprehensive range of application domains includes Management and Marketing,
User Interaction, Healthcare, Education, Finance, Public Monitoring, etc.
Owing to the evolution of technology and the tremendous growth of the Internet,
massive amounts of digitized data in text, images, videos, etc., are available on digital social
platforms. Digitized text is available on digital social media in blogs, news articles, customer
reviews on different products, services, and their experiences, discussion forums,
review/recommendation systems, conversational agents, etc. Moreover, these online social
platforms permit people to express their ideas, opinions. So, all these social platforms lead
the way to analyze the people’s emotions, feelings for investigating social trends, tracing the
customer feedback to understand and develop the business strategy, helping
customers/consumers in the decision-making process, etc. Therefore text-based emotion
detection has transformed into a significant research area.
Fig. 1 Subdomains of Artificial Intelligence (AI)

Since then, text-based emotion detection has been clasped in many real-world applications;
(Chung et al. 2018) describes a novel system for emotion analytics that uses social media
content to aid information collection, situational awareness, and administrative decision-
making. (Sameh et al. 2020) explored automatic sentiment analysis for assessing the polarity of
a product or a service review. (Winster et al. 2020) discussed emotion classification in News
Articles. (Chehal et al. 2020) implemented a system based on user reviews in e‑commerce
recommendations. (Graterol et al. 2021) propose a framework to allow social robots to detect
emotions. (Moisés R. et al. 2020) surveyed intelligent conversational agents and stated agents
potentially detect neuropsychiatric diseases. (Rodrigo Bavaresco et al 2021) focused on human-
like conversational agents for business. (Montenegro et al. 2019) explored conversational
agents in health care. (Topal et al. 2016) developed a movie review system to help users in
decision-making.
Detecting emotions from a massive amount of digitized text on social platforms to enhance
man-machine interaction and aid decision-making in the various application domains is the key
to encouraging research in the field. TBED has multidisciplinary facets thus applied to multiple
areas such as psychology, sociology, human-computer interaction (HCI), data mining. This
Systematic Literature Review (SLR) not only bring forth an outline of the current state of
research on text-based emotion detection but underline the various digital social platforms that
can be used for the purpose, what techniques have been applied, what different datasets have
been used in the research and what will be future directions in the field.

1.1 Significance

The importance of emotion detection from text is rising expeditiously hand in hand with the
Internet and online digital media. Therefore, technologists, business strategists, government
agencies, political analysts accentuated and want to take advantage of this field in all aspects of
the decision-making to improve businesses, reputations, etc. The Internet offers online social
relationship sites such as Facebook, media sharing networks like YouTube, Instagram,
microblogging sites like Twitter, Reddit. This increase in the advent and popularity of social
networks has motivated researchers to investigate online content and analyze users’ online
social behaviors (Bazzaz et al. 2020).
 Customer or product reviews play an essential role in determining consumer behaviors and
mindsets for businesses and simultaneously helping consumers decide to buy products/services
(Sunil et al. 2021). So, in addition, the adaption of chatbots or conversational agents in fields
like in healthcare to monitor and automate the delivery of healthcare services (Montenegro et
al. 2019), in business to improve the service quality, market competitiveness (Bavaresco et al.
2020), in education to promote meaningful interaction in distance learning or E-learning
(Donggil et al 2017). So, it becomes vital to study emotions from the text in different
application domains.

1.2 Motivation

There is no existing Systematic Literature Review (SLR) that emphasizes techniques, datasets,
application domains, comparative analysis, and future directions in text-based emotion
detection. Present reviews and surveys lack a comprehensive study of applications domains and
future trends in text-based emotion detection. The most common way for humans to interact
with computers is still text-based input; even in the advent of web 3.0, text-based emotion
detection should be addressed as a research area. There has been a considerable surge in the
field of sentiment analysis and emotion detection in the last few years. Although the present
paper focuses on the research articles on text-based emotion detection, many researchers have
focused on sentiment analysis. As per our extensive study, no systematic reviews have not been
performed on text-based emotion detection with significance on techniques, datasets,
applications domains, and future directions.
This systematic review intends to show insufficiency in the literature on the subject of
datasets. Even though the alike datasets are available, each dataset has a limited amount of
data, and the labels related to emotions are limited; as a result, emotion detection systems
cannot be reliably trained or generalized (Feng et al. 2020).
Another inspiration for this systematic review is that text-based emotion is highly domain-
dependent. Thus, approaches and datasets suitable in one domain like social media may not be
ideal for another application domain like review systems (Loitongbam et al. 2020). So, such
topics need to be studied to identify the research gaps in the field of TBED. The essence of this
SLR is to highlight existing approaches, available datasets, applications, challenges, and future
directions in text-based emotion detection.

1.3 Terms and terminologies

The terms (Tyng et al. 2017) that are typically used in text-based emotion detection research
are as follows:
1) Sentiment - Sentiment is defined in the Oxford dictionary as exaggerated and self-
indulgent feelings of tenderness, sadness, or nostalgia or a feeling or an opinion, especially
one based on emotions.
2) Emotion - Emotions represents a composite set of interactions between objectives and
subjective variable quantities that are facilitated by the brain and hormonal systems, which
can -
(a) escalate the affective experiences of emotional valence like displeasure and pleasure and
emotional arousals like low-high activation or arousing-calming.
(b) produce logical processes which are relevant to emotions such as appraisals, perceptual
affect, labeling.
(c) cause extensive physiological and psychological changes in response to the stimulating
circumstances.
 (d) encourage communicative, goal-directed, and adaptive conduct, albeit this isn't always
the case.
3) Mood - Moods remain retentive than emotions, which has both negative and positive
characteristics.
4) Feelings - Feelings are cognitive perceptions that are valanced, either positive or
negative, and are followed by the body's physiological changes in the body, particularly the
internal organs, in order to maintain or restore homeostatic balances.
5) Affect - Affect is a term that relates to consciously reachable basic processes such as
arousal and pleasure. It is difficult to explain individuals experienced emotional feelings but
has been connected to physical states such as homeostatic drives, e.g., thirst and hunger, and
external stimuli like touch, taste, smell, auditory, visual.

1.4 Evolution of text-based emotion detection

The term “Emotion” was inscribed in the work “The Expression of the Emotions in Man and
Animals” in 1872, by the famous Scientist, Charles Darwin. It was the first theory where
human emotion and expressions were identified. Then in 1980, Psychologist R. Plutchik
(Plutchik 1980) built the emotion wheel. He showed eight basic and principal emotions:
anticipation, anger, fear, disgust, joy, trust, surprise and sadness.
He further said that each fundamental emotion has a polar opposite. After that, the first
emotion recognition system was developed based on humans' facial expressions (Paul et al.
1980) in 1980. Paul Ekman and Wallace Friesen developed a system named EFACS (Emotion
Facial Action Coding System). Paul Ekman determined people of different cultures share six
fundamental emotions represented by certain facial expressions. Happiness, anger, disgust,
surprise, fear, and sadness are the six primary emotions. Then research in emotion detection
advanced with recognizing emotions in speech. The first seminal work on identifying emotions
in the speech was written by (Dellaert et al. 1996), which became the first research publication
in the field.
Figure 2 shows the brief history of emotion detection and how emotions and their research
evolved. Rosalind Picard's affective computing theory, published in 1997, was the next turning
point in the area. According to Picard (Picard et al. 1997), computers must have the ability to
identify, interpret, and express emotions if we want them to be truly intelligent and
communicate naturally with us. As a result, it shifted the field of emotion detection in a new
direction. Affecting computing theory enhanced human-computer interaction through text-
based inputs. In 1997, the first emotion dataset developed was ISEAR. The International
Survey on Emotion Antecedents and Reactions (ISEAR) database is created from multicultural
questionnaire studies from 37 countries. It consists of 7665 sentences labeled with emotions
and emotion labels as anger, disgust, fear, guilt, joy, sadness, and shame.
In 1998, the first English language lexical database was created for Natural Language
Processing (NLP) tasks as Wordnet (Fellaum et al. 1998). Then practical use of the term
sentiment and emotion analysis was coined in 2001 by Das and Chen (Das and Chen 2001) to
predict the stock market. After that, Wiebe (Wiebe et al. 2005) wrote the first article on
annotating emotion and opinion detection from the text in 2005. After that, in 2009, SenticNet,
is publicly available semantic resource for concept-level sentiment analysis, was developed.
Word embedding is a term used to represent words for text analysis in NLP. So, the first-word
embeddings Word2Vec were developed in 2013. Then in the same year as 2013, neural
networks were adopted for the first time in the NLP task. Then, in 2017, the concept of
transformers was explained in the publication (Vaswani et al. 2017), which gave a huge surge
in the field.
Fig. 2 Evolution of emotion detection – A brief history of emotion detection.

And then pretrained model brought an evolution in the field of automated text classification
and analysis in 2018 in the form of BERT. In 2019, Google developed the Transformer XL
model to help machines understand context beyond that fixed-length limitation. Microsoft
introduced a transformer-based generative language model - Turing Natural Language
Generation (T-NLG) in 2020. It is the largest model published with 17 billion parameters. In
2021, the research field directed towards integrating multimodal information using large
pretrained models. We look at new techniques and methods evolving in the text-based emotion
detection field in the coming sections.
1.5 Prior Research

On the topic of text-based emotion detection, inconsiderable literature reviews have been
written. As per our best knowledge, no systematic literature review has been written so far on
text-based emotion detection. Indeed, emotion detection evolved from sentiment analysis.
Hence, it becomes sensible to consider systematic reviews on sentiment analysis in this survey.
Table 2 gives an overview of survey papers. In a tertiary analysis, (Ligthart et al. 2021)
explored at sentiment analysis.
The authors gave a thorough overview of the various approaches to sentiment analysis. In
this sentiment analysis paper, the features, methodologies, and datasets utilized are explored.
Difficulties and unaddressed issues are also mentioned, which can aid in identifying areas
where sentiment analysis research is needed. (Alswaidan and Menai 2020) discussed explicit
and implicit emotion detection from the text in their survey paper. The authors presented the
state of art approaches, their main features with merits and demerits. Moreover, the authors
studied different corpora and different lexicons available for emotion detection from text. This
review also underlines the significance of NLP tasks, performances of different approaches;
and suggests some open issues. (Francisca et al. 2020) surveys the concept and main
approaches implemented in TBED systems. Authors added discussion on futuristic approaches
in the field with datasets, outcomes with positives, and negatives.
Also, annotated datasets are provided with suitable text for emotion detection. This article,
too, presents a few open problems and future research directions. (Akshi Kumar et al. 2019)
aims a systematic literature review to investigate and review current work on sentiment
analysis based on context and highlights research gaps and future directions. This SLR
identified and analyzed the use of context in Sentiment Analysis. (Seyeditabari et al. 2018)
reviews the work that has been done in TBED. It argues that many techniques, methodologies,
and models developed to detect emotion in text, are inadequate for various reasons. (Shervin et
al. 2021) presents a thorough overview of deep learning-based text classification models,
including their methodological contributions, similarities, and weaknesses. (Soujanya et al.
2017) reviews affective computing, multimodal affect analysis frameworks. (Bostan et al.
2018) focuses mainly on the use of audio, visual, and text information and studies the fusion
techniques of multimodal data. (Dang et al. 2020) reviews the methods of deep learning
employed to resolve sentiment analysis problems, like sentiment polarity and comparative
analysis of different deep learning methods. (Shilpi et al. 2017) reviews different approaches,
classifiers, applications domains, and issues related to text emotion detection. (Koswari et al.
2019) provides an overview of text classification algorithms, various feature extraction
approaches, existing algorithms, dimensionality reduction methods, methodologies, and
evaluation methods. It also covers the techniques' limitations and real-world implementations.
(Ali et al. 2017) presents a classification of sentiment analysis, a survey on emotion theories,
emotion mining-related polarity classification methods, and resources, useful resources, such as
datasets and lexicons. (Sailunoz et al. 2018) discusses advancements in emotion detection
research, including several emotion models, related datasets, detection algorithms,
characteristics, limits, and prospective future approaches in the field of text and speech-based
emotions.
Presented surveys ((Francisca et al. 2020, Ligthart et al. 2021, Alswaidan and Menai 2020,
Akshi et al. 2019, Seyeditabari et al. 2018, Shervin et al. 2021, Soujanya et al. 2017, Bostan et
al. 2018, Dang et al. 2020, Shilpi et al. 2017, Koswari et al. 2019, Ali et al. 2017, Sailunoz et
al. 2018) lack comprehensive study of datasets, technical approaches, a study of application
domains, and future directions. This review initiates prospects for future research work in this
area. It also aims at an in-depth survey on various datasets, techniques, approaches, application
domains or areas, and future trends in the field of TBED.

1.6 Research Goals

The research goal of this systematic literature review (SLR) is to gain a comprehensive
understanding of the current progress in TBED and recognize open challenges and barriers in
the area of TBED. Table 1 shows research goals.

Table 1 Research Questions with objectives

RQ
Research Question Objective
Number
The aim is to study the different Artificial
What are the different Artificial
Intelligence (AI) techniques for TBED, their
RQ1 Intelligence (AI) approaches used for
advantages, and limitations and to show a
TBED?
comparative analysis of different techniques.
Different application domains have different
requirements and, therefore, require different
What application domains have been methods, datasets. The aim is to study the
RQ2
adopted in text-based emotion detection? different application domains in text-based
emotion detection, their requirements, and
comparative analysis.
What are the different datasets available The goal is to review the available public
for research purposes in text-based datasets for text-based emotion detection by
RQ3 emotion detection, and which domains analyzing application domains, data sources,
have been acquainted in the available data size, emotion labels, and data
data sets? imbalance.
Text-based emotion detection is a domain-
dependent task. Therefore, it is one of the
What are the difficulties and open issues
RQ4 challenges. Therefore, the goal is to identify
concerning TBED?
the challenges and open issues in text-based
emotion detection.

1.7 Contribution of The Work

The Contributions of this systematic literature review are as highlighted:

• We provide a detailed overview of the existing literature on text-based emotion
detection with a particular focus on techniques, datasets, applications, associated
challenges, and future directions.
• We discuss and try to explore the techniques and methods used and how they are
refurbishing the research in text-based emotion detection.
• We also provide an overview of different publicly available datasets that can support
research in the domain.
• We also focus on analyzing the role of text-based emotion detection in various
application domains and explore each application domain.
• We present different types of challenges associated with text-based emotion detection,
including dataset, the accuracy of existing techniques, handling quality of data.
Figure 3 shows the outline of the SLR. It shows sections, subsections, and discussions
aimed at the specific sections.
Fig 3 Outline of the SLR
2 RESEARCH METHODOLOGY

An SLR is a type of ancillary research that practices a definite process to discover, analyze, and
understand existing information relevant to a certain research issue in a fair and (to some
extent) reproducible manner (Kitchenham et al. 2007). This systematic literature review (SLR)
follows the PRISMA guidelines. The Preferred Reporting Items for Systematic Reviews and
Meta-Analysis (PRISMA) method is a collection of guidelines for the composition and
structure of systematic reviews and other meta-analyses based on data. These guidelines were
published by (Moher et al. 2009) in 2009. Selection criteria, inclusion/exclusion criteria,
selection outcomes, and quality assessment are the four sections of the authors' research
methodology.

2.1 Selection Criteria

Authors predominantly employed IEEE, Science Direct, Scopus, and Web of Science databases
for searching the documents related to emotion detection. First, a specialized query was
constructed to get the relevant articles from multiple database searches. Then, the filtering
process is applied to the article selection to improve the results that meet our primary goals.
This process includes eliminating duplicates, applying inclusion and exclusion criteria, filtering
based on title and abstract, and finally, full-text screening. Figure 4 shows the process of
literature review in detail. Table 3 shows search keywords or queries used for selecting the data
from different databases using primary and secondary keywords using AND, OR Boolean
operators.

2.2 Inclusion/Exclusion Criteria

Authors prepared a set of inclusion criteria for research article selection and exclusion criteria
for research article rejection to choose relevant research studies for systematic review. After
identifying all relevant articles, the authors eliminated the articles that did not meet the criteria
by using exclusion criteria. Table 4 shows the inclusion and exclusion criteria.

2.3 Selection Results

With the initial query result, a total of 1782 articles were discovered. IEEE provided 584;
Science Direct supplied 886, SCOPUS found 145, Web of Science gave 167 papers, articles,
and research publications linked to Text-based emotion detection. Following that, 1133 articles
were obtained as a result of applying inclusion/exclusion criteria. Further search narrowed
down by studying keyword relevance, title, and abstract of articles; 266 articles were found.
Finally, full-text filtering and quality assessment criteria were applied to each item, and 63
articles were chosen for the final literature review. From 2005 through 2021, studies from
conferences, journals, and reviews were included.
Table 2 Summary of existing surveys related to text-based emotion detection

TBED Challenges
discussed in the survey
Feature paper
Techniques
Extractio
Emotion Multi- Pre- /Approach Applicat Future
Data n Performance

Technique

Quality of
Semantic
Ref. Models modal processing es ions Direction Overview

Dataset
set Techniqu Metrics
ity Techniques s

data
es

It surveys the
Francisca et al.

concepts, main
approaches, futuristic
2020

approaches, datasets
√ √ × × × √ × √ × √ √ × × used, performances,
strengths, and
weaknesses in the
area of TBED systems.
It presents a tertiary
study that gives
current research in
Ligthart et al.

sentiment analysis. It
covers only secondary
2021

studies. It covers key

× × √ √ √ √ √ √ × √ × √ × points and different
approaches,
algorithms, datasets,
the preferred deep
learning models for
sentiment analysis.
It focuses on implicit
Alswaidan and Menai

and explicit emotion

recognition from text.
It presents the state of
art approaches, their
2020

features with merits

√ √ √ × √ √ × √ √ √ √ × × and demerits,
different corpora, and
different lexicons
available for
text-based emotion
detection.
 It investigates and
Akshi et al.

reviews current work

on sentiment analysis
2019

× × √ × √ √ √ × × √ √ × × based on context and

discusses the
limitations and future
directions.
It reviews the work
that has been done in
Seyeditabari et

TBED. It argues that

al. 2018

many techniques,
√ √ √ × √ √ × × √ √ × √ × methodologies, and
models developed to
detect emotion in text,
are inadequate for
various reasons.
It provides a
Shervin et al.

comprehensive
review of deep
2021

learning models for

√ × × × × √ √ √ √ √ × √ × text classification with
contributions
technical, and
weaknesses.
It reviews affective
computing,
Soujanya et al.

multimodal affect
analysis frameworks.
2017

It focuses mainly on
√ √ √ √ √ √ × √ × × √ √ × the use of text, visual
and audio data and
studies the fusion
techniques of
multimodal data.
It surveys the
Bostan et al.

datasets, compares
emotion corpora, and
2018

√ √ √ × × √ √ × × √ × × × aggregates them in a
standard file format
with a common
annotation schema.
 It reviews the
Dang et al.

methods employed in
deep learning to
2020

√ × × √ √ √ √ × × × √ × × resolve sentiment
analysis problems,
like polarity and
comparative analysis.
It reviews different
Shilpi et al.

approaches,
classifiers,
2017

× × √ √ √ √ √ × × × × × √ applications domains,
and issues related to
text emotion
detection.
provides an overview
of text classification
Koswari et al.

algorithms, various
feature extraction
2019

approaches, existing
× × × √ √ √ √ × √ × × × × algorithms, dimensio
nality reduction
methods,
methodologies, and
evaluation methods
It presents a survey
on emotion theories
and emotion mining
Ali et al.
2017

related polarity
√ √ √ × √ √ × × × √ × × × classification methods
and resources
including lexicons
and datasets.
discusses advancements
in emotion detection
research, including
several emotion
Sailunoz et al.

models, related
datasets, detection
2018

algorithms,
√ √ √ × √ √ × × × × √ × √ characteristics, limits,
and prospective
future approaches in
the field of text and
speech-based
emotions.
 Table 3 Sources with search queries

Final Selection
Count of Count of articles after Selection-Based
after Quality
Database Search Query Articles inclusion and On Title and
Assessment &
obtained exclusion criteria Abstract
duplicates
("text emotion detection”) AND ("artificial intelligence or deep
learning or machine learning”) OR (“Chatbots or Conversational
IEEE agents or Social Media or Twitter or Facebook or Reddit or 584 482 164 31
Instagram or Reviews”) OR (“emotion analysis or emotion
recognition or sentiment analysis”)
(“text emotion detection”) AND (“artificial intelligence or deep
Science Direct 886 453 41 19
learning or machine learning”)
(TITLE-ABS-KEY (emotion AND detection) AND TITLE-ABS-KEY
(text) AND TITLE-ABS-KEY (artificial AND intelligence OR deep
AND learning OR machine AND learning) OR TITLE-ABS-KEY
Scopus (sentiment AND analysis OR emotion AND analysis OR emotion 145 140 38 7
AND recognition) OR TITLE-ABS-KEY (chatbots OR conversational
AND agents OR social AND media OR Twitter OR Facebook OR
Reddit OR Instagram OR reviews))
("text-based*") AND TOPIC: ("artificial intelligence" or "machine
learning*" or "Deep learning*" or "natural language processing*")
AND TOPIC: (“Emotion Detection*” OR “Sentiment analysis*” OR
Web of Science 167 58 23 6
“Emotion Analysis*” OR “Emotion Recognition*” OR “Chatbots,”
OR “Conversational agents*” OR “Social Media*” OR “Twitter” OR
“Facebook” OR” Reddit” OR “Instagram” OR “Reviews”)
Total Articles 1782 1133 266 63
2.4 Quality Assessment

The following steps were used to assess quality:

1) Emotion detection: Research must be focused on emotion recognition, techniques, or
datasets.
2) Text-based emotion detection: Research must focus on text-based emotion detection in
social media and different application domains.
3) Artificial intelligence techniques: Research must focus on different artificial intelligence
techniques used in text-based emotion detection.
4) Datasets: Research work also focuses on the publicly available datasets related to text-
based emotion detection.
5) Classification Techniques – The research work emphasizes classification techniques
used in text-based emotion detection.

Table 4 List of inclusion and exclusion criteria

Criteria Inclusion criteria Exclusion criteria

Publication Years of articles

1 Non-English Research articles
must be between 2005 to 2021
Article types considered
Duplicate research articles
2 Conference Paper, Article,
Review, Conference Review
Articles published in
computer science, engineering, Research articles with non-
3 psychology, decision sciences, availability of full text
and social sciences were
considered.
Research articles that are not
At least one of the search
relevant/ not focused on text-
4 terms should be met by
based emotion detection and
articles.
artificial intelligence.
Articles that do not provide
Articles should give answers
5 information related to the research
to the research questions.
question.
Fig. 4 shows a detailed process of literature review
2 LITERATURE SURVEY

A TBED system is a set of methods that process and classify text information to detect
emotions embedded in them. When we take a broad perspective, we may divide it into
many categories, as shown in Figure 5. Emotion models, datasets, approaches, application
domain, text pre-processing methods, feature extraction analysis, feature extraction
methods, and classifiers are shown in the figure. In text-based emotion detection, online
social media, product or customer reviews, conversational agents or chatbots, discussion
forums are some sources that are rich in textual information. So, these application domains
need to be focused on. Then there are some datasets or corpora available online which
already have annotated data or labeled data. It would be useful to gain a deeper
understanding of emotions to enhance the classification process. Modeling emotions has a
variety of methods and approaches, and it is still ongoing. The categorical, dimensional,
and componential models are preferred. The data needs to be preprocessed before features
extraction. These feature extraction steps are crucial for a classification process. It reduces
the original data to its most essential characteristics or makes it compatible with the
classification algorithms. Text analysis generally uses five types of analysis to extract the
features: lexical, syntactic, semantic, disclosure integration, and pragmatic analysis. There
are different methods used for feature extraction, such as BoW, TF-IDF, POS, etc. Then all
of these features are inputted into the classification system, which has a variety of
classifiers at its disposal. More recently, deep learning and machine learning classifiers are
commonly been used. We have given a detailed explanation of these approaches and
methods in the coming sections.

3.1 Text-Based Emotion Detection Model

Figure 6 shows the text-based emotion detection model. In the first step, the data is
acquired from different text sources. Then text preprocessing is performed to clean the
data. Mainly, the text from social media, product/customer reviews consist of slang words,
emojis, short text, incomplete words, etc., which requires preprocessing. So preprocessing
steps include tokenization, lowercasing, removing stop words and digits, removing
punctuation marks, removing white spaces, removing hashtags (#), removing URLs,
removing emoticons and emojis. Then applying stemming and lemmatization to get words
into their root form. The next step is feature analysis which includes feature extraction and
feature selection. In Feature extraction, essential features are extracted. If the feature set is
large, the feature selection can be made to choose the most informative. The final step is
classification and assigning data to one of the possible emotion classes. Classification
techniques are used to recognize emotions. Commonly known classification techniques are
machine learning, deep learning. In the literature, we can find, for example, Support
Vector Machine (SVM), K-Nearest Neighbor (K-NN), Naïve Bayes, many others. Deep
learning does not require specific features. It learns from data itself. However, we need a
relatively large data set to train generalization abilities. If the available data set is too
small, its size can be expanded artificially by data reinforcement techniques.
3.1.1 Emotion Models for Emotion Detection

Studies of emotions date back to Greek mythology. Ancient Greek literature exhibits a
wealth of emotions. According to Greek literature, emotions were organized into four basic
categories: metus (fear), aegritudo (pain), libido (lust), and Laetitia (pleasure). After that,
in the 19th century, Darwin’s emotion theory stated emotions and expressions of emotions
in humans and animals. In psychology, emotions are categorized into basic emotions and
complex emotions (that is, emotions that are hard to classify under a single term, such as
guilt, pride, shame, etc.). However, there is no universally accepted model of emotions.
There are two prevalent theories in this field: the discrete emotion model and the
dimensional model. And the third model has based on appraisal theory: the componential
model. Table 5 shows comparative analysis of different emotion models.

1) Discrete Emotion Model – It is also termed a categorical emotion model. As

stated by, discrete emotion theory, some emotions are distinguishable based on
neural, physiological, behavioral, and expressive features regardless of culture.
The discrete emotion model states that emotions can be placed into basic distinct
classes or categories. Different emotions correspond to different neurological
subsystems in the brain. Ekman's model (Ekman Paul et al. 1999) categorizes
emotions into six basic categories. These basic emotions are happiness, sadness,
anger, disgust, surprise, and fear. Plutchik's model (Plutchik 1980) states eight
emotions in opposite pairs as joy vs. sadness, trust vs. disgust, Anger vs. fear, and
surprise vs. anticipation.
2) Dimensional Emotion Model - This model assumes, emotions are dependent and
a relation exists between them. The dimensional model is based on the hypothesis
that all emotions result from a common and interconnected neurophysiological
system. Dimensional emotion models define a few dimensions with some
parameters and specify emotions according to those dimensions. Two or three
dimensions are used in most dimensional emotion models— ‘valence’ (indicates
the positivity or negativity of an emotion), ‘arousal’ (indicates the excitement
level of emotion), and ‘dominance’ (indicates the level of control over an
emotion). (Russell et al. 1980) presents a circular two-dimensional model
prominent in dimensional emotions representation called the circumplex of
effect. (Plutchik 1980) presents a 2-dimensional wheel of emotions showing
valence on the vertical axis and Arousal on the horizontal axis.
3) Componential Emotion Model - This approach is an expansion of the
dimensional approach. It is built on the appraisal concept (Scherer et al. 2005).
According to this paradigm, emotion is derived through an analysis of
circumstances if a person can be met with it. The outcome is determined by a
person's background, objectives, and possibilities for action.
Fig. 5 Overview of TBED systems
 Table 5 Comparative analysis of different emotion models

Discrete Model Dimensional Model Componential Model

Emotions can be
Emotions are not The componential model
placed into basic
independent, and that uses the appraisal theory
Description distinct classes or
there exists a relation based on a person’s
categories. Emotions
between them. experience.
are independent.
Complex/mixed emotions
Basic and Universal
are addressed well. Focus on the variableness
Model
Advantages Highly recommended for of different emotional
Widely adopted due
projects involving states.
to its simplicity.
emotional resemblances.
Reduction in 3-D space
Limited to fixed results in loss of
emotions. information.
No such standard appraisal
Disadvantages Difficult to address Not all fundamental
criteria. Different variants.
complex and mixed emotions are compatible
emotions with the dimensional
space.
Russell circular two-
Paul Ekman Model, dimensional model,
Robert Plutchik Plutchik 2-D wheel of
Models Model, Orthony, emotions, Russell and Scherer’s Appraisal theory.
Clore, and Collins Mehrabian 3-D emotion
(OCC) model model, The Hourglass of
Emotions
Ekman Paul et al. Plutchik 1980,
1999, Russell et al. 1980,
References Scherer et al. 2005
Plutchik 1980, Russell et al. 1977,
Ortony et al. 1990 Cambria et al. 2012
interaction modalities
acceptance, anger, (Pleasantness), interaction
anger, fear, disgust,
disgust, surprise, contents (Attention),
Emotions joy, surprise,
anticipation, fear, joy, interaction dynamics
sadness,
sadness (Sensitivity), interaction
benefits (Aptitude).
Fig. 6 Text-based emotion detection model
3.1.2 Text Preprocessing Methods

The initial step in text-based emotion detection prepares the data to the form appropriate
for any machine or deep learning system. Many deep learning and machine learning
algorithms require numerical data to perform any classification or regression task.
Therefore, text preprocessing and feature extraction are vital steps for text classification
problems. This segment introduces methods of text cleaning, tokenization, normalization,
and allowing for informative featurization. Mainly, the text from social media, customer/
product reviews consist of short text, emojis, slang words, incomplete words, etc., which
make preprocessing a necessity. Text preprocessing has the following steps:
1) Data cleaning – It includes lowercasing, punctuation, and digits removal, stop
words removal, hashtag & HTML tag removal, expanding contractions, etc.
2) Tokenization/Segmentation – Splitting text/strings into tokens representing
words.
3) Normalization - Stemming and lemmatization include truncating a word to its
root form.
Figure 6 shows all preprocessing steps in detail with examples. In addition, the figure
shows the output after applying each step.

Fig. 7 Elements of Feature Extraction Analysis

3.1.3 Feature Extraction Analysis

Commonly, documents and texts are unstructured data. Therefore, this section gives a brief
theoretical idea about feature extraction elements before diving into feature extraction
methods. Following are some elements that need to be considered while extracting features
in the text. Figure 7 shows different feature extraction analyses used in text-based emotion
detection.
1) Lexical Analysis – In linguistic analysis, the whole text is divided into
paragraphs, sentences, and words. It helps to identify and analyze the word
structures. Tokenization is one of the examples of linguistic analysis.
 2) Syntactic Analysis – It involves analyzing the grammar of words in a sentence
and showing the relationship among the words. Parts-of-Speech tagging (POS),
parsing are some examples of syntactic analysis.
3) Semantic Analysis – Semantic analysis determines the actual meaning of words
and evaluates the text's significance. Named Entity Recognition (NER) is an
example of semantic analysis.
4) Disclosure Integration – Disclosure integration gives a sense of the context. It
involves building the relationship and meaningfulness between two consecutive
sentences or phrases. Dependency parsing is one of the examples of disclosure
integration.
5) Pragmatic Analysis – The general interpretation of language is the focus of
pragmatic analysis. It is concerned with leveraging real-world knowledge to
derive meaningful interpretations of human language.

3.1.4 Feature Extraction Methods

The unstructured text must usually be converted to structured language using a

mathematical model or algorithm, as shown in figure 7. To begin, the data must be
cleansed to remove any irrelevant characters or words. Formal feature extraction
approaches can be used once the data has been cleansed. Feature extraction
algorithms are required to convert the text into a matrix (or vector) of features. This
transformation task is generally called feature extraction of data. Feature extraction
techniques are shown in figure 8. Text feature extraction techniques have two
approaches: the traditional approach based on syntactic word representations and the
modern approach based on feature learning techniques.

Fig. 8 Overview of feature extraction methods

1) Traditional Approach (Syntactic Word Representations) –

It includes the following methods -
a) Bag of Words (BoW) — This method is also known as the n-gram method. BOW
 counts the frequency of words in the text and generates a scarce vector that represents
present words as one and missing words as 0 in the text. Machine learning models use
these vectors as input. N-grams are groups of words that appear in close proximity to
one another and are merged into a single vector of features. Common representations
of N-grams are 1-gram, 2-gram, 3-gram.
b) Term Frequency-Inverse Document Frequency (TF-IDF) – The TF-IDF metric
indicates how significant or relevant a term is in a document. It identifies words or
sets that frequently appear in a single document yet are uncommon throughout the text
corpus. TF-IDF is defined as follows:
TF-IDF = TF * IDF
TF = Term Frequency = no. of term in a document / total terms in the document.
IDF = Inverse Document Frequency = Total No. of documents / no. of documents has
term.
So, TF gives the measure of how frequently the term occurs. And IDF gives the
measure of how important the term is.
c) Part of Speech Tagging (PoS) – The POS is the word category that showcases its
role and helps the NLP work. It includes categories like nouns, verbs, adjectives, and
adverbs.
d) Named Entity Recognition (NER) – In the disambiguation, the entities are
identified in a sentence, such as the famous person, brand, etc. In the named entity
recognition, the entity is identified and categorized as per the date, organization,
person, time, location, etc.
2) Modern Approach (Feature Learning Techniques) –
Embeddings - Word embeddings are techniques in which features are learned, and each
word from the dictionary is represented to N-dimensional integer. These feature vectors
are based on the context of the text dataset. To convert words into intelligible input, many
word embedding methods have been developed. Word2Vec, GloVe, FastText, and ELMo
are some common deep learning embedding approaches that have been successfully
applied.
• Word2Vec – (T. Mikolov et al. 2013) introduced the "word to vector"
interpretation. It is an upgraded architecture of word embedding. This approach
utilizes continuous bag-of-words, skip-gram, and lightweight neural networks
consisting of two hidden layers to generate a feature vector for every word. It
gives you immediate access to the vectors that represent words. A simple
continuous bag of words (CBOW) model looks for words that might appear near
one other, whereas skip-gram looks for words that might appear near each other.
This approach is a strong method for detecting correlations in a text dataset as
well as word resemblance.
• Glove – GloVe (Global Vectors for Word Representation) by (Pennington et al.
2014) is another technique for creating word embeddings. It captures a global
word-to-word conjunction matrix from a dataset to construct word embeddings.
The idea is to establish a context for words i and j in terms of whether or not they
occur near N-words apart. The ratio of the co-occurrence probability of two
words is encoded in the encoding vector, which is known as the count-based
technique. However, this method does not employ the entire corpus. It collects
 global statistics and improves word representation learning on tasks including
word comparison, word similarity, and named entity recognition.
• Fast Text - Facebook's AI research team created the FastText library. It is capable
of learning and calculating word representations as well as sentence
classification. It is a novel technique that solves the morphology issue by
allocating a different vector to each word. As a result, each word is represented
by an n-gram, which is a collection of characters.
• ELMo – Embeddings from Language Models (Peters et al. 2018) is a type of
contextual embedding that looks at the words around. It uses a bidirectional
LSTM algorithm, which means that the words before and after it are used in both
directions. This character-based language model is able to envision the previous
or next word, as well as handle words that aren't in the dictionary. It also handles
semantical and grammatical characteristics of words and how this differs around
language contexts like to develop lexical ambiguity. It advances modern
techniques in a broad range of demanding NLP tasks, involving textual context
question answering, and sentiment analysis.

3.2 Approaches for Emotion Detection–

The most common approaches for identifying text-based emotions are keyword-based,
rule-based, machine learning-based, deep learning-based. The approaches for TBED are
displayed in figure 3.
1) Keyword-based approach – This technique focused on locating keyword occurrences
in a given text and comparing them to the annotations registered in the dataset. Figure 9
outlines the method. The initial emotion keyword list is derived from conventional
lexical resources WordNet or WordNet-Affect in this method. After that, the dataset is
preprocessed. Then, keyword matching is performed between a predefined keyword list
and emotion words from the text. After that the power of the emotion keyword is
examined. After that, negation is examined to see whether there are any negation cues
and what is the scope of, and lastly, the emotion tag calculation is done. In this, we
studied (Tao J. 2004, Ma C et al. 2005, Perikos et al. 2013, Shivhare et al. 2015) which
are based on a keyword-based approach.

Fig. 9 Keyword-based approach lay-out

2) Rule-based approach – This approach utilizes linguistic rules to recognize the emotions
from the text. Figure 10 shows a general overview of the approach. The dataset is first
subjected to text preparation. It involves data cleaning, tokenization, POS tagging, etc.
Then rules for extracting emotions are built using a statistic, linguistic concepts. Then,
probabilistic affinity is stored with each word. Later, the best of the rules has been chosen
for the dataset to detect emotion labels. References (Lee et al. 2010, Udochukwu and He
2015, Liu and Cocea 2017) surveyed in rule-based approach.

Fig. 10 Rule-based layout

3) Machine learning-based Approach — Approach is depicted in Figure 11. Machine

learning-based methods allow systems to learn and improve on their own as a result of
their experiences. Text is classified into several emotion classes using machine learning

Fig. 11 Machine learning-based approach layout

algorithms. There are two types of machine learning algorithms: supervised and
unsupervised. Supervised machine learning algorithms are applied in the majority of the
studies reviewed. Text preprocessing, such as tokenization, POS tagging, and
lemmatization, is usually the first step in this technique. Then the text's useful features are
extracted, and only the features having the highest information yield are chosen. The
algorithm is then trained using the specified characteristics and emotion labels. finally, the
system is utilized to predict emotions from unobserved data using a trained system. The
authors examined (Aman et al. 2007, Ghazi et al. 2010, Bruyne et al. 2018, Suhasini et al.
2020, Singh et al. 2019, Allouch et al. 2018, Jonathan et al. 2017) in this approach.

4) Deep learning-based approach - Deep learning uses neural network layers to learn
unsupervised from unlabeled or unstructured data. It is one of the subclasses of machine
learning in Artificial Intelligence. Figure 12 depicts the deep learning approach. In the
deep learning, neural networks learn complex theories by building from simpler concepts.
The emotion dataset is first processed with preprocessing steps such as tokenization, stop
words removal, and lemmatization. Then, the embeddings are constructed. In this case,
tokens are represented by numbers. Then, using classification, these vectors are provided
to deep neural network layers with elements equivalent to emotion labels, where data
forms are discovered and used to estimate labels. References (Baziotis et al. 2017, Ezen
and Can 2018, Basile et al. 2019, Shrivastava et al. 2019, Xiao 2019, Rathnayaka et al.

Fig. 12 Deep learning-based approach layout

2019, Bian et al. 2019, Mohammad et 2021, Jose et al. 2020, Xia and Zhang 2018, Diogo
2021, Taiao et al. 2020, Polignano et al. 2019) surveyed in deep learning-based approach.

3.3 Datasets/Corpora

In this section, the authors discuss widely used datasets for TBED, which is the primary
focus of this article. In addition, this section informs researchers on recent advancements
in the domain of TBED. In TBED, researchers can either create their own datasets or use
publicly available datasets. In this section, the authors present some publicly available and
useful datasets with a reliable labeling or annotating process. Figure 13 gives a
representation of different datasets with application domains used. These datasets are
widely used by researchers in TBED and are described as follows:
1) ISEAR (The International Survey on Emotion Antecedents and Reactions):
This dataset was created in 1997 by (Scherer and Wallbott 1994). It's one of the first
curated datasets using emotion labels, and now it's free to download. The information was
gathered by a group of psychologists who worked on the ISEAR project in the 1990s.
There are seven major emotions included in this dataset: joy, anger, fear, disgust, guilt,
sadness, and shame. 3,000 students from 37 countries, including psychologists and non-
psychologists, were asked to explain situations in the form of emotions they had
encountered in this study.
2) Cecilia Ovesdotter Alms’ Affect data:
(Alm and Sproat 2008) developed this dataset in 2005. It includes 185 stories by Beatrix
Potter, Hans Christian Andersen, and the Brothers Grimm. It has 15000 sentences labeled
with the emotions fear, anger, disgust, sadness, happiness, and surprise. It also has
positively surprised, negatively surprised, or neutral if no emotion is displayed.
3) SemEval (Semantic Evaluations):
The database Semantic Evaluations (SemEval) (Rosenthal et al. 2019) contains Arabic and
English data. According to the data, this dataset has three versions. SemEval 2007 is a
collection of news headlines from various publications, including BBC News, The New
York Times, CNN, Google News, and others. Surprise, anger, fear, sadness, disgust, joy
are all annotated with one or more emotions in each headline. SemEval 2018 is made up
entirely of tweets. Anger, fear, disgust, love, joy, optimism, sadness, pessimism, trust, and
surprise are among the eleven emotions expressed in each tweet. Distinct test, trial, and
training datasets are provided for Spanish, Arabic, English, tweets. SemEval 2019 is made
up of textual conversations between 2 persons. Each interaction is categorized into one of
four groups: anger, sadness, joy, or others. The third turn of the discussion is used to
categorize emotion labels. Distinct training, trial, and test datasets are available.
4) Crowdflower:
This dataset (https://data.world/crowdflower/sentiment--in-analysistext) contains tweets
and their labels for the emotional content of texts. It was developed in 2016. It has a
substantial number of labels across 13 labels. It includes emotion labels: boredom, anger,
empty, fun, enthusiasm, happiness, love, hate, neutral, sadness, relief, worry, surprise. This
dataset has 40000 tweets.
5) Emobank:
Over 10,000 words have been dimensionally labeled using the Valence-Arousal-
Dominance (VAD) emotion representation paradigm. (Buechel et al. 2017) developed it in
2017. These sentences were collected from a wide range of sources, including essays, news
headlines, blogs, fiction, newspapers, travel guides, and letters written by writers and
readers. Ekman's basic emotion model has been used to categorize a subset of the dataset,
making it eligible for dual representational techniques.
6) Aman:
This corpus is made up of blogs that were obtained using Ekman's six fundamental
emotions as clue words, e.g., In the pleasure category, the words happy, delighted, and
enjoy were chosen. It was created by (Aman et al. 2007) in 2007. The sentence-level
annotation was done, and there are eight different emotion classes to choose from,
including two new ones: mixed emotion and no emotion. It has 1466 sentences labeled
with mixed emotion and no emotion, sadness, surprise, happiness, fear, disgust, and anger.
Sentences depicting neutral feelings are defined by the word "no emotion."
7) The Valence and Arousal Facebook Posts by Preotiuc-Pietro:
(Preotiuc and Pietro 2016) created this data set in 2016. It has been labeled on two scales
that are independent of one another: The polarity of the affective content of a post is
represented by valence (or sentiment), which is assessed on a nine-point scale from 1 (very
negative) to 5 (neutral/objective) to 9 (extremely positive) (very positive). Arousal (or
intensity) is a nine-point scale that ranks the strength of affective content from 1 (neutral
perspective) to 9 (very intense) (very high). Facebook status make up of this corpus. There
were 3120 posts in the original data set.
8) MELD (Multimodal Multi-Party Dataset for Emotion Recognition in Conversation):
It is a multimodal corpus that includes data from textual, audio, video data. It has around
1400 conversations and 13000 statements from the TV series Friends, which are labeled as
anger, grief, disgust, surprise, fear, joy, and neutral in dialogues. It was developed by
(Poria S et al. 2018) in 2018.
9) CBET (the Cleaned Balanced Emotional Tweets):
In 2015, (Gholipour Shahraki 2015) compiled this dataset from Twitter. He used hashtags
to detect tweets with one of nine emotions: joy, anger, sadness, love, disgust, surprise, fear,
guilt, and gratitude. The CBET is divided into two sections. One section has only one
labeled tweet, denoted as single labeled tweets. This section has 76,860 tweets, with 8,540
for each emotion, and is completely balanced over labels. Double-labeled tweets express
two emotions in the text, which are found in the smaller section. This part has 4,303 tweets
and is unbalanced since not all combinations of emotions occur equally and frequently.
10) GoEmotions:
This dataset comprises 58K Reddit comments from 2005 (the start of Reddit) to January
2019. In addition, this dataset has labeled for one or more of 27 emotion(s) or Neutral as
follows: admiration, amusement, approval, annoyance, anger, curiosity, confusion, caring,
desire, disgust, disapproval, disappointment, embarrassment, excitement, fear, joy, grief,
gratitude, love, sadness, nervousness, optimism, pride, remorse, relief, realization, surprise.
(Dorottya Demszky et al. 2020) developed this corpus in 2020.
11) Amazon:
The work of (Blitzer et al. 2007) in 2007 resulted in the name "Amazon." It's a product
review dataset compiled from the Amazon website. It features 2,000 reviews in four
distinct sectors, including DVDs, books, electronics, and kitchen products. This dataset is
balanced because each section has equal positive and negative reviews. Each sample in the
dataset contains specific information about a review, such as a title, the rating ranging from
0 to 5, and date of the review, and the substance of the review.
12) IMDB (Internet Movie Database):
It's the unlabeled version of the polarity dataset. The IMDB dataset
(https://datasets.imdbws.com/) was created to aid in the classification of binary sentiment
in movie reviews. There are equal positive and negative reviews on IMDB. It is split
uniformly in the training and test datasets, with each receiving 25,000 reviews. It was
created in the year 2011. Its data consists of 27,886 unprocessed and unlabeled HTML
files.
13) The Movie Review (MR) dataset:
It is a dataset of movie reviews curated to detect the sentiment related to detailed reviews
and to determine positive or negative reviews. It contains 10,662 review sentences with the
equivalent count of positive and negative instances. This dataset was developed in the
early 2000s by (Bo Pang and Lillian Lee 2002).

Fig. 13 Datasets with topics and domains

3.4 Application Domains

Text-based emotion detection has wide applicability to various application domains. This
wider range of application domains includes management and marketing, user interaction,
healthcare, education, finance, public monitoring, etc. An abundant amount of text data is
generated and analyzed to make decisions in many disciplines such as online social media,
customer or product review systems, recommendation systems, discussion forums, and
conversational agents. We are focusing on online social media, conversational agents, and
review systems.
3.4.1 Conversational Agents

Conversational agents are referred to by various terms, including chatbots, dialogue

systems, intelligent virtual assistants, intelligent virtual agents, smart bots, interactive
agents, digital assistants, and relational agents. A chatbot is a human-computer interaction
model with an artificial intelligence program (Eleni et al. 2020). A chatbot is described as
a "computer software meant to simulate interaction with human users, particularly over the
Internet, by using spoken or written natural language as a primary means of
communication.". Conversational agents, often known as chatbots, mimic human
interaction and are utilized in various areas such as business, education, healthcare,
entertainment, and many more. The primary goals of chatbots are entertainment, social
contact, and novelty interaction, with a strong emphasis on productivity. In chatbots, there
has been a tremendous requirement to simulate human-like characteristics and behavior.
Chatbots are usually depicted, built, and developed as a flow of communication between
multiple components as Natural Language Understanding (NLU), Dialogue Management
(DM), and Natural Language Generation (NLG). The NLU generates a semantic
representation of a user text, such as intent or purpose, to extract the "meaning" of the text.
The NLU's main role is parsing, which involves taking a collection of words, generating a
linguistic formation for the words, and storing in the knowledge repository. The second
component is dialogue management, which formalizes communication structure and
performs contextual understanding by assessing word disambiguation and identifying
phrases connected by references and supports domain knowledge supervision. It can match
expressions (emotion) using lexicon and handling synonyms, negation, and intensity based
on the presence of terms (emotional). The third component is NLG or response module.
Commonly, Chatbot responses are restricted to providing a direct response to the
participant's question. NLG gets a communication response from the dialogue manager and
creates a textual depiction that corresponds. Language generation interprets gist by
selecting the syntactical form and words required to represent it.
In business, conversational agents provide availability, scalability, and cost-
effectiveness by enhancing market competitiveness and service quality (Adikari et al.
2019). These chatbots also improve customer engagement by offering friendliness,
comfortness, flexibility, and efficient assistance. Chatbots should provide customers with
more engaging responses, directly addressing their issues. Users perceive chatbots as
companions rather than simple assistants. The majority of user requests are emotional
rather than informative. Chatbots gained the capacity to respond emotionally to customers
due to machine learning and sentiment analysis evolution. (Adikari et al. 2019) discussed
an artificial intelligence-based cognitive model to detect emotions in industrial chatbots
using word embedding, Markov chains, and NLP. Conversational agents are increasingly
being used in industrial applications, from simple conversation interfaces to sophisticated
human aides. Industries are infusing human-like qualities, features, and behaviors into
conversational agents to increase productivity. A chatbot implemented by (Thomas 2016)
responds to the users with AIML, and LSA uses a dataset of Frequently Asked Questions.
(Xue et al. 2018) designed a neural-based conversational agent that uses Bi-LSTM with
attention to reply to repetitive questions. (Sun and Zhang 2018) developed a personalized
recommendation system. It includes three key components: a deep belief network as a
tracker, an LSTM network, a deep policy network. In order to optimize user experience, it
also focuses on the personalization of the recommendation system. They used
Factorization Machines (FM) that glues the strengths of Support Vector Machines (SVM)
and factorization models.
In the Healthcare field, conversational assistants are used for diet management,
medication adherence, mental wellbeing, and physical activity (Fadhil et al. 2019).
Furthermore, due to the low learning curve, users, particularly the elderly, are comfortable
with chatbots. In (Fadhil et al. 2019) developed a CoachAI, a health coaching application
driven by a conversational agent with a supervised machine-learning model. In the
healthcare field, conversational agents could assist users by presenting appropriate
information about an illness or explaining the findings of clinical tests (Joao et al. 2019).
Clinicians use conversational agents to help them identify symptoms and improve their
assessment abilities, diagnosis, interview procedures, and interpersonal communication.
(Pacheco et al. 2021) reviewed smart conversational agent systems to detect
neuropsychiatric disorders. Affect and cognition are the human abilities most desirable by
chatbots to accomplish change behavior in humans (Pereira et al. 2019).
In the education field, educational conversational agents are designed to facilitate
and support online learning. These chatbots enable interactions between the learner and the
content; thus, the agent's role is to deliver the instructional content (Song et al. 2017).
Chatbots systems are employed for literacy tutoring chatbots, security training chatbots.
Chatbot assists students (Hien et al. 2018) in administrative concerns for example applying
for a course, finding out about examination schedules, results, and other academic
particulars. In (Jia and Jiyou 2004), a conversational agent was developed to converse with
English learners using a basic method of logical reasoning and inferences based on
semantic and syntactic analysis.
Chatbots can play different roles, such as recommenders, information providers,
tutors, advisors, conversational partners, entertainers, and can be used for multi-party
interaction, decision-making, conflict, and opinion resolution. Pattern matching and
machine learning are two approaches used to construct a chatbot, based on the algorithms
and methodologies
1) Pattern matching – It is also called a rule-based approach. Pattern matching algorithms
compare user input to a rule or pattern and choose from a set of replies a preset answer.
The context can also influence which rules are used and how the response is formatted.
The first chatbots to use the pattern matching approach were ALICE and ELIZA. It uses
logic programming clauses and if-else patterns for applying to reason. It traverses the tree
structure made up of rules to find the next action to perform.
2) Machine learning - For chatbot systems, various machine learning approaches exist,
each with its own set of learning scenarios. Each learning scenario produces a
conversational action to be taken next. Machine learning algorithms have grown in
popularity since they can choose which conversational act to use at any time, based on
prior knowledge removing the need for domain experts to introduce policies and rules to
the system regularly. Bayesian networks, neural networks, Markov models are different
mechanisms used in the machine learning approach. For example, the development of
chatbots based on Artificial Neural Networks (ANNs) uses score assignment and chooses
the most expected response from a collection of responses in retrieval-based models.
While in Generative models, on the contrary, the response is synthesized using deep
learning approaches.
Chatbots have several disadvantages and risks in addition to their tremendous
benefits. Data security is an important issue in conversational agents for both user and
agent service providers. Chatbots often fail to detect the intentions of their speaker. Failure
to detect the intent of the user may create frustration for the user. And its adverse effect
may prove ruinous to the service provider. Long responses covering the important
information that may discourage users and lead to a break in the conversation. Intention
classification can also be affected by user errors in spelling. Misused phrases, subtle
sarcasm, language impairments, usage of slang, and syntax faults are all examples of
failures caused by user input. The vast majority of chatbots are focused on business,
notably e-commerce; other areas, such as healthcare, finance, management, education,
have received less attention.

3.4.2 Online social media –

The majority of the world's population uses popular social networking sites or applications
such as Facebook, Instagram, Reddit, Twitter, YouTube, and WhatsApp. Furthermore,
microblogging services such as Instagram, Reddit, and Twitter are immensely popular.
Consequently, we may state that postings that include text, images, audio, and video,
online media has developed as a strong medium for expressing, communicating, and
sharing people's thoughts, opinions, perspectives, views, and including native to worldwide
issues and topics. So, it becomes important to analyze the emotions behind people's
opinions, thoughts, views, and perspectives. This study will be useful in finding the impact
of OSM on different online trends(positive/negative/influential/biased), social behavior,
and individual social conduct. Feelings, emotions, character traits, and their impact on
societal trends can all be found using TBED. There are various online social media
application areas such as business intelligence, behavioral analysis, epidemics, event
detection, crime detection, sentiment, opinion, and emotion analysis, recommendations.
The authors in (Sailunaz et al. 2019) developed a recommender system based
using sentiment analysis and emotion detection employing machine learning algorithms on
the Twitter posts. By using the Naïve Bayes algorithm, anger, disgust, joy, fear, surprise,
sadness, and neutral emotions are detected by their system. While positive, negative, and
neutral polarities are identified in sentiment analysis for each Twitter post. Their system
calculated users' influence scores using a variety of user- and tweet-based factors. And
used this data to create generalized and individualized suggestions for users based on their
Twitter behavior. (Pool et al. 2016) explored emotion classification using Facebook
reactions in a distant supervised context. The authors used Support Vector Machine (SVM)
with a simple BoW to classify the emotions like anger, joy, sadness, surprise. (Douiji et al.
2016) proposed a system coping with the complications of writing style in chats and
language progress to detect users' emotions from their textual interactions from YouTube
comments. Based on a dataset produced from YouTube comments, their approach used an
unsupervised machine learning technique to do emotion classification. The authors utilized
Latent Semantic Analysis, and its Probabilistic version, Non-negative Matrix Factorization
to detect sadness, joy, fear, surprise, disgust, Anger from YouTube comments.
In (Kumari et al. 2021) authors discovered a text-based cyberbullying detection system in
social media using semantic autoencoders. They proposed a method named semantic-
enhanced marginalized stacked denoising auto-encoder (smSDA) via semantic dropout
noise based on word embeddings of domain knowledge and sparsity constraints. They
evaluated the performance of the system using Twitter and myspace corpora of
cyberbullying. In (Wang et al. 2017) authors developed a model to discover a group of
closely connected people and highly involved in their sentiments about a product/service.
They introduced the concept of sentiment community using the optimization models of
semi-definite programming (SDP). The authors developed an influential activity
maximization model to identify the influential node. Independent cascade and linear
threshold, these diffusion models are used in their activity maximization model. Authors in
(Song et al. 2017) studied the spam messages on social media sites. They proposed a novel
methodology for social spam detection using labeled latent Dirichlet allocation (L-LDA)
on the word-based, topic-based, user-based features, and accuracy improved using
incremental learning. In (Gupta et al. 2021) the authors present a research approach that
effectively measures population opinions and identifies misunderstandings, informing
public health communications in epidemics. They studied the ability of weather to affect
the spread of COVID-19 pandemics. Their work was based on Twitter data and used
Gradient Descent Support Vector Machine (GDSVM), random forest, naïve Bayes, and
TF-IDF machine learning algorithms. Topic modeling is used to find out major topics of
discussion in the pandemic. Authors in (Sun et al. 2015) designed a social regulation
method that incorporates social network information to benefit recommendation systems.
The authors considered the social relationship among the users for improving the accuracy
of the recommendation system. They used a bi-clustering algorithm to calculate the trusted
friends in the network.
Many difficulties must be overcome while identifying emotions via social media.
First, the casual writing style of users in social media is a big challenge. Grammatical and
spelling mistakes, as well as slang language, sarcasm, and irony, may be found in social
media posts. Simultaneously, the use of informal language and short messages has been
studied, but only to a limited extent. Second, the expression of human emotions and the
texts that convey them are subjective and ambiguous. As a result, correctly inferring and
interpreting the author's emotional states is challenging. Third, emotions are multifaceted
constructs with unclear confines and varying expressions. As a result, human mapping is
important. Finally, affective expression is difficult for automated systems to understand.
Because of the wide range of topics discussed on social media, manually creating a
complete dataset of labeled data that includes all conceivable emotional circumstances is
difficult.

3.4.3 Review systems

BI (Business Intelligence) (Sandra et al. 2021) is a set of methodologies, frameworks, and

approaches for transforming data into actionable information that helps businesses run
more efficiently. It is, in fact, a set of apps and services that convert data into usable
knowledge and experience. The use of the BI paradigm to social media data benefits
businesses of various domains. Social media platform data can benefit various business
domains. Business-related information gathered from product/customer reviews, tweets,
likes, shares, comments of specific and related business data and disseminated through
online social media platforms can assist businesses in producing high-quality products and
services. Especially in product design and development, early feedback, detection, and
conceptual concepts from customer study are crucial. A good user experience with the
product, system, and service is gradually recognized as a significant product/service design
advantage. In general, user experience/feedback (UX) refers to how a user feels about
utilizing a product, system, service, or object (Bai et al. 2019). Customer holding and
trustworthiness are critical components of business success. Because there is widespread
agreement that customer satisfaction is strongly linked to slow customer destruction, many
business strategies focus on analyzing and improving customer satisfaction. According to
the application area, emotion recognition from product and customer reviews has been
researched more than service reviews.
The authors of (Bai et. al 2019) examined how user input data may be retrieved
and restored from online customer reviews to help product design. A multi-aspect
proposed framework is developed as an operational technique for highlighting the key
components of user feedback in product design. The authors also suggested a technique for
constructing a knowledge base of user comments from online consumer reviews. User
feedback discovery, which extracts user feedback data from a single review, integrates
with similar group data, and user feedback network formalization, which establishes causal
linkages among user feedback data groups, are the three stages of their approach. By
analyzing a large number of online customer data, this study analyses the feasibility of
automatically uncovering relevant user feedback data and their correlations for product
creation and strategic business planning. The overall goal of (Monireh et al. 2018) research
is to create an automatic tool that analyses customer reviews of a product and determines
how they feel about various characteristics of that product. Authors offer the Pros/Cons
Sentiment Analyzer (PCSA) framework for collecting sentiment knowledge from
pros/cons reviews using dependency relations. Feature extraction using syntactic rules and
determining the opinion’s polarity according to their emotional power from pros/cons
reviews highlight their work. Authors in (Cambria et al. 2020) used proposed domain
adaptation using genetic programming in product reviews. The authors proposed a
framework, Genetic Opinion Adaptation Learning (GOAL) which combines convolutional
deep belief network (CBDN) with Genetic Programming (GP). Authors use a variable-
length tree called a genetic program to model each text's features. At internal nodes in the
tree, mathematical operators such as '+' or '-' can denote the polarity of phrases. The
proposed model is tested using Amazon product reviews for a variety of products and
languages.
Even while this field's research isn't inextricably connected to the crucial user
feedback idea "emotion" (e.g., happiness, joy, fear, surprise, and disgust), user sentiment
like a neutral, negative, or positive does refer to feelings, albeit not to the same extent.
Therefore, it is needed to present a possible map from sentiments to feelings by further
leveraging the language elements in the feedback/ evaluations, such as sorting them based
on emotion categories. Emotion detection from review text has mainly been done in movie
review systems, hotel review and recommendation systems, E-commerce review systems,
agriculture product review systems.
3.5 Classifiers

After generating feature vectors and creating word embeddings next significant step is to
classify the emotions. Deep learning and machine learning are the two classification
techniques used by most researchers. Overview of different classifiers used in text-based
emotion detection is shown in figure 14. This section describes different deep learning and
machine learning classifiers used to classify emotion purposes in detail.

3.5.1 Machine Learning Classifiers

Machine learning classifiers are used extensively and comprehensively in the field of text-
based classification. These classifiers work on annotated/labeled datasets, so known as
data-driven (Nantheera 2020). Computing methods in machine learning are used to directly
learn knowledge from large amounts of training examples/samples. These algorithms
adaptively arrive at optimal solutions and generally enhance their effectiveness as the
number of samples supplied for learning increases. There are three different categories of
learning classifiers: supervised, unsupervised, and reinforcement learning. A supervised
learning classifier contains inputs and desired output labels, whereas an unsupervised
learning classifier deals with unlabeled data. Reinforcement learning classifiers learn from
trial and error and are self-supervised means a limited set of annotated data provided for
labeling and usually a large amount of unlabeled data. Decision Tree (DT), K Nearest
Neighbor (KNN), Support Vector Machine (SVM), Random Forest (RF), Naive Bayes
(NB), Linear Regression (LR), Multinomial Naive Bayes (MNB) are some most preferred
machine learning classifiers.
Decision Trees (DT):
A decision tree is a rule-based supervised classifier (Allouch et al. 2018). It is a tree
structure made up of nodes and branches/directed edges. Each node signifies an
attribute/feature, and each branch signifies the decision rules or tests on attributes (Hasan
et al. 2019), and each leaf node demonstrates a class label/output. There are three steps for
decision tree learning (Adivi et al. 2019): features choosing, decision tree generation, and
pruning. Simple decision rules from the features are used for the learning process (Shayaa
2018). The decision tree is a frequently used classifier for regression and classification.
This method represents decision-making through a tree-like structure (Pradhan et al. 2016).
When using a DT, figuring out which attributes to be utilized, what criteria to be utilized
for splitting, and when to terminate are crucial factors. Throughout this procedure, all
attributes are examined, and several division points are explored and tried. The split that
costs the least is picked to determine the optimum path. Pruning a tree can boost its
performance even further. Pruning is the process of deleting less important elements from
a tree to reduce its complexity and, as a result, increase the predictive power by reducing
overfitting. (Kaur 2012) utilized the decision tree to investigate human emotion variance
(Kaur 2012). The decision tree can be used to examine and classify many types of
emotions. To detect the emotions of children with diverse disabilities, Kaur employed
outlier analysis.
Random Forest (RF):
Random Forest is a supervised classification algorithm. It is built using the decision tree
approach. The term "forest" refers to a collection of decision trees that have been
combined to produce a more accurate, robust prediction and enhance the overall result.
Because of its simplicity, this approach is flexible and straightforward to use, so it is used
for classification and regression problems. The trees continue to grow, and random forest
adds randomness by observing the finest attributes within a subset of random features
rather than only investigating the most significant ones. It helps increase variation in this
way, leading to improved models and results (Fang and Zhan 2015). Boosting and bagging
are the two most used classification approaches (Segnini et al. ). Bootstrap aggregation is
referred to as "bagging." It combines many learners/predictors to minimize the decision
tree's variance. Random forest trains N decision trees on diverse random subclasses of the
data and calculates the average of all the predictions for final prediction (Abu et al. 2018).
Boosting method is used to create a strong classifier from a set of weak classifiers.
Boosting try to deal with errors from the prior decision trees by forming a group of
sequential predictors. Boosting techniques can be used to track down the model that failed
to provide an accurate prediction. Overfitting is less of an issue with boosting approaches.
Ensemble learning generates predictions by averaging the decisions of numerous
predictors. It produces better predictions compared to a single decision tree.
k-NN (k - Nearest Neighbor) –
k-NN is one of the simplest categories. The nearest neighbor of K is the meaning of k
nearest neighbor. The central impression of the k-NN algorithm is that a sample is
classified based on the majority votes of its neighbor, with the sample classified to the
class that is most common between its k nearest neighbors. The method determines the
type of sample classified based on one or more recent samples (Kaur and Duhan 2015). It
used Euclidean distance to determine the distance of an attribute from its neighbors (Wu
et al. 2008). The KNN algorithm is a slow, non-parametric algorithm that makes no
assumptions about the underlying data distribution and does not require any explicit
training before classification (Shayaa 2018). It is used for classification and regression.
Support Vector Machine (SVM) –
SVM is a supervised learning model. It is used to solve problems involving classification
and regression. It's mostly used to organize linearly separable data, although it can also
handle non-linear data in a high-dimensional feature space. It constructs a set of
hyperplanes (decision boundaries) in high-dimensional space with the largest margin to
classify the data (Kaur and Duhan 2016). It aims to improve the classification's robustness
by using the largest margin. In most classification problems, more complex structures are
required to achieve optimal separations. A nonlinear separation is required to better
identify the clusters in most circumstances than a linear line. SVM overcomes this
challenge by employing kernel functions, which are mathematical functions, to reorganize
the data set. To do the linear separation, the technique entails plotting the input data in a
new space and translating it into a higher dimensional feature space. SVM has the
advantages of being effective in high-dimensional space, memory efficiency, and versatile
as various kernel functions available (Zahid et al. 2020). However, overfitting may arise if
the count of features exceeds the count of samples. To categorize reviews, (Cheng and
Tseng 2011) employed two multiclass SVM-based techniques: To assess the reviews'
quality, researchers used Single-Machine Multiclass SVM and One-versus-All SVM
(Cheng and Tseng 2011).
Naïve Bayes (NB)-
The Naïve Bayes classifier is one of the simplest classifiers used for text classification. It is
derived from the Bayesian theorem of probability, and the Naïve Bayes assumption
assumes that the features of an object are conditionally independent, given the class label
of the object. There are many variations, such as the Multinomial Naïve Bayes, the
Gaussian Naïve Bayes, and the Bernoulli Naïve Bayes, all of which operate on the
probability of the occurrence of words in the text. This classification is straightforward and
effective, and it's especially well-suited to inputs with a lot of dimensions. Naive Bayes
outperforms more advanced categorization systems despite their simplicity (Sharupaet
2020). The Bayes rule, for example, determines the likelihood of features appearing in
each class in Bayesian analysis and then delivers the most probable class. This approach
was previously described in (Tang et al. 2009). They also introduced a Multinomial Naive
Bayes (MNB) method for resolving existing uncountable and meaningless features, which
takes into account the amount of positive and negative words while computing weights and
eliminates irrelevant words while selecting features.

3.5.2 Deep Learning Classifiers

Deep learning classifiers come under unsupervised learning. However, it can be supervised
or semi-supervised too. Deep learning classifiers learn and extract features automatically,
resulting in enhanced accuracy and performance. Convolutional Neural Network,
Recurrent Neural Network, BERT, Bi-LSTM, GRU, and pre-trained models are some most
preferred deep learning classifiers.
Convolutional Neural Network (CNN) –
CNN is widely seen in Recommender Systems, Computer Vision, and NLP applications. It
is a convolutional and pooling or subsampling layer of a deep feed-forward neural network
that feeds data into a fully connected neural network layer (Zhang et al. 2018).
Convolution layers obtain features by filtering input data, and several filters are merged to
get outputs. With pooling or subsampling, layers feature resolution is lowered to improve
the CNN's distortion and noise resistance. Minimizing the dimension of the result from
one layer to the next subsequent layer is called pooling. Various pooling algorithms are
utilized to lower the number of results or outputs while maintaining essential attributes.
The commonly used pooling approach is max pooling, which selects the largest factor in
the pooling window. Fully connected layers carry out classification tasks. When it comes
to identifying local and position-invariant patterns, CNNs perform well. CNN's have been
employed efficiently in text classification (Shervin et al. 2021).
Recurrent Neural Network (RNN) –
Researchers used a recurrent neural network for text mining and classification (Sutskever
et al. 2011). It is a kind of neural network architecture in which previous data points in a
sequence are given higher weight by RNN. As a result, this strategy effectively classifies
text, string, and sequence data. Furthermore, in a very sophisticated approach, an RNN
examines the information of prior nodes, allowing for a greater semantic understanding
structure of a dataset. RNN (Britz 2015) consists of feedback loops in which the
connections between neurons create a directed cycle. The basic role of RNN is to process
Fig. 14 Overview of Classifiers

sequential input using internal memory obtained during directed cycles. RNN can recollect
and apply prior knowledge to the next element in the input sequence, unlike traditional
neural networks. When gradient descent error is backpropagated across the network, RNN
is prone to vanishing gradient and exploding gradient difficulties. As a result, RNN
primarily use LSTM or GRU for text categorization.
LSTM -
S. Hochreiter and J. Schmidhuber (Hochreiter et al. 1996) invented LSTM, which has
since been expanded by a number of research experts (Graves et al. 2005). The LSTM is a
form of RNN that solves issues more successfully than the basic RNN by retaining long-
term dependency. When it comes to addressing the vanishing gradient problem, LSTM
proves to be useful (Pascanu et al. 2013). LSTM also had chain-like architecture like RNN,
but it uses many gates to closely limit the extent of information permitted into each node
state. The input is preprocessed to rearrange it for the embedding matrix. The next layer is
the LSTM, which has 200 cells. The final layer comprises 128 text classification cells and
is fully connected. The final layer uses the sigmoid activation function to reduce the 128-
height vector to a one-height output vector if there are two classes (negative, positive) to
forecast. Bidirectional LSTM networks are one sort of LSTM network. Bi-LSTMs are
made up of two hidden layers. The input sequence is processed forward using the first
hidden layer and backward using the second hidden layer. The output layer combines the
hidden layers, allowing it to access the past and future context of each point in the
sequence. The LSTM and its bidirectional variations have proven to be extremely useful.
They may learn when and how to forget particular pieces of knowledge, as well as when
and why not to use specific gateways in their architecture. A Bi LSTM network has the
advantages of a faster learning rate and greater performance (Schuster et al. 1997).
GRU –
(J. Chung et al. 2014) and (Farzi et al. 2016) proposed GRUs as a gating mechanism for
RNN. The GRUs architecture is a simpler version of the LSTM. A GRU, on the other
hand, differs from an LSTM in that it has two gates and no internal memory. In addition, a
second non-linearity is not used. (Cho et al. 2014) presented the Gated Recurrent Unit, a
recurrent neural network that attempted to solve the vanishing gradient problem (GRU).
Because both are built similarly, GRU is a variant of the LSTM. To overcome the
vanishing gradient problem, GRU employs an update and reset strategy. The model uses
the update gate to decide the amount of previous data from prior time steps that should be
utilized in the next future steps. The reset gate decides the amount of data is discarded. A
Bi-GRU network that is bi-directional is a variant of GRU (Bi-GRU).
Transformer Models –
Transformer-based deep learning models use the self-attention process, weighing the
significance of each portion of the input data differently. Comparison analysis of different
pre-trained models is shown in table 6.
BERT – The BERT stands for Bidirectional Encoder Representations from Transformers. It
is a language transformation methodology launched by Google (Devlin et al. 2019). It is
"deeply bidirectional". It acquires the extensive interpretations of texts by taking into
consideration right and left contexts equally. This method is utilized to train general-
purpose language models on big datasets and solve NLP problems. The pre-training and
fine-tuning steps are involved in employing BERT. BERT model is trained on unlabeled
data in the pre-training phase. The model is then fine-tuned for specific NLP tasks after
being initialized with the pre-trained parameters. The cost of fine-tuning the BERT model
is substantially lower in terms of system resources. BERT employs the same architecture
for a variety of activities. The Transformers (Vaswani et al. 2017) are used to construct
BERT. The model is available in two sizes: BERT-base and BERT-large. BERT-base has
12 Transformer blocks with a hidden size of 768 and 12 self-attention heads, whereas
BERT-large has 24 Transformer blocks with a hidden size of 1024 and 16 self-attention
heads. The English Wikipedia and BooksCorpus datasets were used to train BERT.
DistilBERT focuses on shrinking the dimensions of the BERT while maintaining
performance (Sanh et al. 2020), to enhance performance, RoBERTa (Liu et al. 2019)
exploits fully, hyper-parameters of BERT and XLNet acquires contexts in both directions
using all permutations of the factorization order (Yang et al. 2020), are three BERT
variants that have recently been introduced.
ULMFiT – ULMFiT (Howard et al. 2018) was developed at the beginning of 2018.
ULMFiT was the first "pure" transfer learning application. It employs AWD-LSTMs, an
LSTM version that employs Drop Connect for improved regularization and improves using
averaged stochastic gradient descent (ASGD). This model has 400 dimensions in the
embedding layer and has three LSTM layers, with 1150 hidden units in each of them. On
top of this approach, the authors stacked a SoftMax classifier of size 50 to pre-train the
model. During fine-tuning, this last layer is supplemented with a task-specific layer. The
vocabulary is constrained to 30k words. The model was evaluated on several data sets
IMDb, TREC-6, Yelp-bi, AG’s news, Yelp-full, DBpedia.
GPT – The Open AI GPT (Radford et al. 2018) design is purely attention-based, with no
recurrent layers. By combining learned position embeddings with byte-pair encoded token
embeddings and incorporating these embeddings into a cross-layer transformer decoder
framework with the purpose of standard language modeling, pre-training is achieved. The
model performs at each step by utilizing a decoder architecture. It has access to the tokens
that came before it in the sequence. As a result, the GPT model is a one-way attention
model. GPT uses transformer decoders to model language using a semi-supervised
learning technique. The GPT, primarily used for text representation, is formed by the
transformer decoder with 12 attention heads and 12 transformer layers. Pre-processing
enormous unlabeled datasets, such as the BooksCorpus dataset on the restricted supervised
datasets, trains and fine-tunes them. In 2019, the Open AI team released GPT2 (Radford
et.al 2019), a scaled-up version of GPT.
XLNet - In XLNet (Yang et al. 2019), Transformer-XL and Permutation Language (PLM)
models are employed. XLNet is a language model that is both auto-regressive and auto-
encoding. BERT masks the data and uses a bi-directional context to try to predict the
masked data, whereas XLNet uses the permutation objective. In a sentence, complete
distinct combinations of words are trained, as the model may learn bi-directional context
with the help of PLM. The techniques of positional encoding and recurrence are then
applied. For maintaining the location information of the token that is to be predicted, it
employs two-stream self-attention. XLNet is pretrained and fine-tuned on English
Wikipedia, Common Crawl, Clueweb 2012-13, Giga 5, and BooksCorpus.
RoBERTa – RoBERTa (short for Robustly optimized BERT technique) is a precise
architectural clone of BERT with a larger dataset and tuned hyperparameters for
pretraining that was introduced in (Liu et al. 2019). The pre-training masking approach has
been changed from static to dynamic. Static masking is performed just once during pre-
processing, whereas dynamic masking is performed many times during pre-processing,
i.e., every sequence is masked just before feeding it to the model. Furthermore, it makes
use of larger pre-training datasets. The model was trained on five English language
datasets: English Wikipedia, BooksCorpus, the Stories dataset, the Open Web data, the
CC-News data.
DistilBERT – (Tang et al. 2109) proposes distillation in neural networks, which tries to
increase the speed of models. It is accomplished by using a simplified version of the BERT
architecture with few parameters. It takes the original BERT's design, decreases the layers
to half of the original BERT, and eliminates poolers and token embeddings to create a
faster and smaller BERT for common applications. For better inference, it uses dynamic
masking and overlooks the next sentence estimates. The model was trained on the same
datasets as BERT, English Wikipedia, and BooksCorpus datasets.

4 OUTCOME OF SURVEY

This section will discuss the results for each of our research questions. It discusses the
findings of primary studies that were used to generate research questions. For every study
question, tables with combined outcomes as well as detailed explanations and analyses are
provided. It's important to remember that different types of emotion detection tasks
necessitate distinct features and methodologies. As a result, each approach is given a brief
description.
4.1 RQ1 - What are the different Artificial Intelligence (AI) approaches
used for text-based emotion detection?

There are different ways to detect emotions in text. These different ways constitute the
different approaches in text-based emotion detection. Different approaches follow the
different tasks, distinct processes. As a result, it is important to consider which tasks
comprise which approach. The approaches utilized in various emotion detection tasks are
discussed in section 3. Keyword-based, rule-based, machine learning-based, and deep
learning-based techniques are commonly employed for text-based emotion detection tasks.
Table 9 presents an overview of emotion models, feature extraction methods, approaches
used in text-based emotion detection. Approaches that are expressly mentioned in the cited
article are marked with a checkmark. The findings are organized into seven categories,
each with its subcategories. The following is a list of each category and its subcategories:
1) Emotion models – In the text-based emotion detection literature, the discrete or
categorical emotion model is widely used, which states that some emotions are more basic,
more universal, or more important.
Emotion models have been extensively used in the keyword-based, rule-based, and
machine learning approaches, but are less preferred in the deep learning approach.
2) Feature extraction methods – Various feature extraction methods are based on
lexical analysis, semantic analysis, syntactic analysis. Parts of Speech (PoS) Tagging, Bag
of Words (BoW), Term Frequency Inverse Document Frequency (TF-IDF), all are the
most used feature extraction methods in keyword-based, rule-based, and machine learning
approaches. At the same time, word embeddings are the most preferred feature extraction
method in the deep learning approach. In most deep learning-based emotion detection
studies, pre-trained word embeddings Glove, Word2Vec, and FastText are employed. The
table shows the most preferred word embeddings in TBED. Pre-trained word embeddings
 Table 6 Comparison analysis of different pre-trained models
Ref. Model Architecture Description Tuning Advantages Challenges Trained on Datasets
Devlin et al.

Bidirectional 1) limited to monolingual classification,

Pre-Training 1) ability to handle contextual BooksCorpus datasets
2019

Transformer, LSTM, Uses and the Next Sentence 2) Fixed length of input sentences
BERT Fine information and the English
Bidirectional Prediction (NSP), Masked Language 3) Suffers from logical inference
Tuning 2) Faster Training Wikipedia.
Modelling (MLM) mechanisms. 4) computationally expensive.

1) Transfer Learning in Language Wikitext-103,

Howard et al.

Fully Connected Domain IMDb, TREC-6,

Pre-Training
2018

Neural Network, 2) very well on small and medium 1)Model not trained on GLUE benchmark. Yelp-full,
ULMFiT Uses AWD-LSTMs Fine
Unidirectional corpora 2) Unidirectional Yelp-bi,
Tuning
3) Prevent over-catastrophic forgetting DBpedia AG’s news,
when fine-tuning
Radford et al.

1) improved lexical robustness.

to model language, a semi-supervised 1) The model's resource-intensive nature
2) Outperforms various models trained
2018

Transformer, learning approach employed utilizing a Pre-Training makes the pre-training step costly. 2) 2)
GPT on domain-specific datasets BooksCorpus, WebText
Unidirectional multi-head attention layer transformer Fine Tuning inability to deal with dependencies that are
3) no-fine tuning required for GPT-2 and
decoders. larger than the given fixed lengths.
GPT-3.

1) The ability to extract bi-directional

Yang et al.

English Wikipedia,
Auto-Regressive Use two-stream self-attention and contextual
2019

Pre-Training 1) XLNet computationally Common crawl, Giga 5,

XLNet Transformer, Permutation Language Model Information
Fine Tuning intensive. Books Corpus,
Bidirectional. principles are used (PLM) 2) It removes BERT's fixed-length
Clueweb 2012-13
limitation.

1) The use of more extensive pre-training English Wikipedia,

Liu et al.

Replication of BERT with larger 1) resource-intensive nature

2019

Transformer, Pre-Training data results in improved performance. Books Corpus,

RoBERTa corpus and tuned hyper-parameters. 2) It is computationally intensive and takes
Bidirectional Fine Tuning 2) In downstream NLP tasks, CC-News, Stories,
Uses Dynamic Masking. longer to complete.
outperforms XLNet and BERT. OpenWebText.
Tang et al. 2019

1) Language modeling capability with

Uses Early Form of BERT by reducing
Distil Transformer, Pre-Training the pre-training Book Corpus and English
the of layers by a factor of 2, Uses 1) Fixed length limitation
BERT Bidirectional Fine Tuning 2) The new model is faster and lighter Wikipedia
dynamic masking.
compared with BERT.
are embeddings that have been learned in one task and can be utilized to solve an identical
problem. These embeddings are created by training them on big datasets, saving them, and
then using them to solve other problems. Pretrained word embeddings are a type of
transfer learning.
3) TBED approaches – TBED algorithms learn a model from text, and emotion labels, are
often referred to as training data. The Model then guesses the emotion label for new,
previously unknown text. Finally, the accuracy of the model is tested on a test set for
which emotion labels are also available is used to determine its performance. The table 7
shows a summary of approaches used in text-based emotion detection.

1) Deep Learning - Deep learning models are convoluted structures that use many layers
of neural networks to obtain complex information from input. In text-based emotion
detection, CNN, RNN, Pre-trained deep learning models based on transformers are
used. CNN uses convolutional filters to identify the pattern of data. CNN's are
typically utilized in image processing applications. Still, they have recently begun to
be applied in the field of natural language processing (NLP) in conjunction with other
deep learning approaches. In NLP tasks, RNN is the most popular deep learning
algorithm. RNNs are useful for storing sequential information. LSTM and GRU are
special types of modern RNN. These modern RNN are good at learning long-term
dependencies in the text. In NLP, where long-term dependencies are prevalent, Bi-
LSTM and Bi-GRU, variants of LSTM and GRU, are extremely useful. Attention
networks are used to concentrate on data parts that are of particular interest. Then
transformer-based pre-trained models used attention with encoders and decoders to
improve the extraction of relational context, eliminate the problem of long-term
sequence dependencies, and allow input sequences to be delivered in parallel.
Bidirectional Encoder Representations from Transformers (BERT), Generative Pre-
Training (GPT), Transformer-XL, Robustly Optimized BERT pre-training Approach
(RoBERTa), XLNet, and DistilBERT are popular pre-trained models. Due to its
bidirectional capability, BERT has the potential to manage contextual information
extraction and the fact that it trains faster and has been employed in various language
modeling applications. Following the popularity of BERT, various modifications on
top of BERT were developed, including RoBERTa, DistilBERT, and SpanBERT.
RoBERTa is a variation of BERT that has the advantage of using a larger amount of
pre-training data, which results in superior performance on a range of tasks. The
concept of distillation in neural networks was suggested by DistilBERT as a technique
to speed up models. DistilBERT is a language modeler that may be pre-trained to
perform different linguistic modeling tasks. It significantly achieves the original
BERT in terms of speed and lightness. XLNet is a language model that is self-
regressive that can extract contextual information by using a permutation language
model. It also removes BERT's fixed-length restriction. For predicting the next word,
GPT is a good language model. The strength of GPT is improved lexical robustness.
GPT-2 and GPT-3 variations do not require finetuning. The discussed deep learning
algorithms are thought to be viable strategies for improving text-based emotion
detection task performance.
2) Machine learning - The machine learning technique tackles the emotion detection
 problem by implementing machine learning algorithms to classify texts into distinct
emotion categories. A supervised or unsupervised ML approach is frequently used for
detection. In text-based emotion detection, commonly supervised machine learning
algorithms are implemented. Classical machine learning techniques are still frequently
utilized in all types of emotion detection tasks similar to text classification. Even
though deep learning is a rising field, for several tasks, classical machine learning
algorithms perform as well as or better than deep learning approaches, especially on
smaller datasets. Algorithms for supervised machine learning that have been around
for a long time are Naive Bayes (NB), Support Vector Machines (SVM), Random
Forest (RF), Neural Networks (NN), k-Nearest Neighbor (k-NN), Logistic Regression
(LR), and Decision Trees (DT). SVM, NB, and RF are more popular machine
learning algorithms as compared to others.
3) Rule-Based Approach – In rule-based learning, grammatical and logical concepts are
used to construct the rules for detecting the emotions from the text. Rule-based
learning is built on exploiting information to understand data in a meaningful way.
4) Keyword Based Approach - The traditional approach to TBED is the keyword-based
approach. It looks for words in documents that indicate different emotions in humans.
This method does not require any learning data because words are specified in a
dictionary/dataset. The goal is to locate occurrences of search words at the sentence
level in a given document. Once the keyword has been located inside the sentence, the
sentence is labeled according to the dataset.

4.2 RQ2 - What application domains have been adopted in text-based

emotion detection?

Public thoughts, opinions, views provide us with valuable insights. Detecting the emotions
from these opinions on the online digital platform will help to learn users, and it has a wide
range of applications domains. This task of analyzing emotions can help in a wider range
of application domains, including management and marketing, healthcare, education,
finance, public monitoring, dialog systems/chatbots, social media mining.
We used the methodological choices on these application domains in the realm of text-
based emotion recognition to survey articles on online social media, review systems, and
conversational agents in section 3. Table 8,10,11 shows the different methods and
techniques used in different text-based emotion detection applications.
Online social media (OSM) applications have become the universal applications on the
Internet. These are rapid data generation applications making it subtle to study these data.
Text-based emotion detection has been widely used in online social media for a variety of
tasks. These tasks include detecting sarcasm, irony, tracing the flow of emotions in social
media, determining personality traits, depression detection, etc. Table 10 shows
applications in social media on which research has been done in TBED includes Twitter,
Facebook, YouTube. Other topics like emotion detection from news articles, behavioral
analysis, crime detection, business intelligence, recommendation systems, stock market
prediction, politics have also been studied. The domain of online social media has
preferred machine learning classifiers: SVM, k-NN, random forest, decision tree, Naïve
Bayes, and logistic regression, whereas Bi-LSTM, Bi-GRU techniques from deep learning
are the most used.
Customer reviews on products or services can provide sentiment or emotions towards
various features or aspects. Therefore, the analysis of these reviews can extract meaningful
insights, which will help businesses maintain customer relationships, gain feedback from
users to help in early product designs, and thus improve the businesses.
In the domain of review systems, some work has been done in product review, movie
review, hotel recommendations systems, etc. Techniques used in review systems are naive
Bayes, SVM, maximum entropy, k-means clustering, convolutional deep belief networks
(CDBN), expectation-maximization semi-supervised algorithm, and syntactic rules with
polarity determination.
Conversational agents, often known as chatbots, are beneficial in a variety of industries,
including education, commerce and eCommerce, health, and entertainment. The most
crucial goal for chatbot users is productivity, while other factors such as enjoyment, social
factors, and novelty engagement are also important. Table 11 shows different applications
areas where conversational agents can be used.

4.3 RQ3 - What are the different datasets available for research purposes
in text-based emotion detection, and which domains have been
addressed in the available data sets?

Labeled or annotated tailored datasets are available for text-based emotion detection
research. Existing datasets were favored by researchers, or they built their own datasets
based on the needs of studies in certain application domains. Here the authors have
identified some understanding of datasets with different perspectives.
The authors have inferred from table 12, online social media, customer/product
review systems, conversational/dialogue systems, and news articles are the application
areas addressed by currently available datasets in text-based emotion detection. Most of
the datasets have been created to study the emotions/sentiments on online social media.
Many datasets are imbalanced in terms of emotion/sentiment labels. All online available
datasets are annotated/labeled. The majority of the datasets are openly accessible and can
be downloaded. The presence of a checkmark in the table indicates that specifically
mentions datasets are balanced or not, annotated or not, access free or not, and multi-
modal or not. In contrast, most datasets are not multi-modal. The important inference is
that the majority of datasets are multi-labeled datasets that can be used to detect text
emotion.
Emotion detection from text is considered a multi-class classification problem. In
text-based emotion detection, researchers use machine learning or deep learning methods,
which require huge amounts of data. On the other hand, supervised learning algorithms
provide good results with annotated datasets. But the main problem with annotated
datasets is, very few annotated datasets are available. Although in Table 12, all datasets
studied are annotated, but these datasets are very less in number.
Table 7 Summary of text-based emotion detection approaches

Ref Approach Description Merits Demerits

- The
emotion
keyword
Based on finding the - Simple does not
Tao J. 2004,
occurrence of a word method always
Ma C. et al. 2005 Keyword
and comparing it - Most used match.
Perikos et al. 2013 Based
with annotated approach. - Emotion can
Shivhare et al. 2015
labels in the dataset. present
without
emotional
words.
- Rules creation
for large data
- Simple
Based on sets complex
Lee et al. 2010, approach.
grammatical and task.
Liu et al. 2017 - Rules are
Rule-Based logical rules from - Lack of
Udochukwu et al. 2015 easily created
text to interpret the linguistic
for small sets.
information. information
affect the
performance
Aman et al. 2007,
- Widely
Ghazi et al. 2010 - Not robust.
Based on the ability implemented
Bruyne et al. 2018, Machine - Not explicitly
to learn from - Better
Suhasini et al. 2020 Learning- extract
experience and detection
Singh et al. 2019, Based semantic
develop. results.
Allouch et al. 2018 information.
Jonathan et al. 2017
Baziotis et al 2017,
Ezen and Can 2018
Basile et al. 2019,
Xiao 2019
- More robust
Shrivastava et al. 2019,
Based on to learn - Extract
Polignano et al. 2019 Deep Need large
without human deeply
Rathnayaka et al. 2019 Learning data to train
intervention from hidden
Bian et al. 2019, Based the system.
data. details.
Jose et al. 2020
Mohammad et al. 2021,
Xia and Zhang 2018,
Diogo 2021,
Taiao et al. 2020
 Table 8 Summary of Review systems in different application areas

Ref. Application Area in Description Techniques/ Methods used Dataset/data Challenges

Review Systems utilized
The proposed model does not directly link
PoS tagging, Bootstrapping
to emotions or user sentiment at the
Utilizing user experience/feedback technique, feature ranking
Bai et al. detailed level.
Product Design from online customer reviews for approach, Expectation- Amazon
2019 Limited domain knowledge and other
product design. maximization semi-supervised
online product review websites are not
algorithm
considered.
Proposed an automatic tool to Pros/Cons
Monireh Syntactic rules with polarity Cannot analyze the service/movie
analyses customer reviews of a Product review
et al. Product Review determination, Lexicons used reviews.
product and determine customer for digital
2018 – SentiStrength, Bing Language dependent.
sentiments about that product. camera
Examines the information quality of
three major online review sites in Latent Dirichlet Allocation, TripAdvisor,
Xiang et Hospitality and Tourism Data used are destination-specific.
relation to online reviews about the Elbow Method, Naïve Bayes, Expedia, and
al. 2017 online review system A very specific case discussed
entire hotel population in New York TF-IDF Yelp
City.
Proposed domain adaptation using Domain Adaptation using
Cambria Multi-domain
genetic programming for Amazon genetic programming, Semi-
et al. Product Review Amazon Variance is high during the search.
product reviews for a variety of supervised convolutional deep
2020 dataset
products and languages. belief network (CDBN)
Presented the notion of emotion
Topal et k-means clustering, emotion heat IMDb movie Very few movies were considered for the
Movie Review maps centered on machine learning
al. 2016 maps. dataset study.
techniques
Customer care center Predict customer satisfaction Human-in-the-loop / active learning
Teja et GRU, CNN, GNN
Product Review following interactions with a Amazon approaches can improve system
al. 2021
corporate call center.
 Table 9 Overview of emotion models, feature extraction methods, approaches used in TBED with is subcategories
Emotion models
Feature Extraction Methods Approaches Used for text-based emotion detection
Used
Word Keyword- Rule-
Machine Learning-Based Deep Learning-Based
Embeddings based Based

Text Pre-processing
Componential
Dimensional RNN Transformer Based

Negation

Not Mentioned
Ref.
Discrete

TF-IDF

GoogleEmb
BoW

NER

Word2Vec
PoS

FastText
Glove

k-NN

ANN

CNN
SVM

DistilBERT
NB

DT

LR
RF

RoBERTa
Bi-LSTM

ULMFiT
XLNet
BERT
GRU

GPT
Tao 2004

Ma et al. 2005

Perikos et al. 2013

Shivhare et al.
2015
Lee at al. 2010
Udochukwu et al.
2015
Liu et al. 2017

Aman et al. 2007

Ghazi et al. 2019

Bruyne et al. 2018

Suhasini et al.
2020
Singh et al. 2019

Allouch et al. 2018

Jonathan et al.
2017
Baziotis et al. 2017

Ezen et al. 2018

Basile et al. 2019

Shrivastava et al.
2019
Xiao 2019
Rathnayaka et
al.2019
Bian et al. 2019
Mohammad et al.
2021
Jose et al. 2020

Xia et al. 2018

Diogo 2021

Taiao et al. 2020

Polignano et al.
2019
Table 10 Summary of online social media in different application aeras

Application Area in Dataset/data

Ref. Description Techniques/ Methods used Challenges
Online social media utilized
Not considered word order in messages which
Kumari et A proposed text-based cyberbullying Stacked denoising auto-encoder, Word Twitter and
Behavioral Analysis can further improve the robustness of the learned
al. 2021 detection system in social media embeddings myspace
representation.
Accuracy needs to be improved of community
Developed a methodology to identify
sentiment detection methods.
a group of people who are strongly Flixster.com
Wang et al. Linear threshold and independent Need to consider the demographic characteristics
Business intelligence linked and emotionally engaged in movie review
2017 cascade diffusion models. of the most valuable customers.
their feelings about a product or dataset
Not observed dynamic customer sentiments from
service.
social media.
Aside from topic-based elements, other high-level
features must be learned.
Song et al. Presented Social Spam detection Labeled latent Dirichlet allocation (L- Need to explore other application domains
Crime detection You-tube
2017 system LDA) besides YouTube.
Need to improve the quality of topic-based
features.
Mislabeled data.
The weather, which has been Machine learning algorithms SVM, NB, Not considered the magnitude of tweets while
Gupta
Epidemics discussed, impacts the virus's RF, and K-means clustering with Topic Twitter annotating the data.
2021
transmission over the world. modeling. Need to improve the performance of the classifier
for sarcasm.
Need to improve the performance.
Sun et al. Recognizing the appropriate class of Time-series information, place information, and
Recommendation systems Bi-clustering Algorithm, SVM Del.icio.us
2015 friends to offer recommendations. the dynamic variations among users are not
considered for recommendations.
Latent Semantic Analysis (LSA),
Created a framework to handle the
Probabilistic Latent Semantic Analysis
complexities of chat writing styles and You-tube NA
Douiji et (PLSA), Non-negative Matrix
determine user sentiment.
al. 2016 Factorization (NMF)
Pool et al. Proposed system to classify emotions There is a lot of variation in ratings for the same
Sentiment, opinion, and SVM Facebook
2016 in Facebook posts emotion among datasets.
emotion analysis
Sailunaz et Not considered abbreviations, emoticons, short-
al. 2019 Calculate user’s influence scores using hand text.
user and tweet information to classify Naïve Bayes Twitter The system is static, with a narrow range of data
the emotion and sentiment. acquired in less span of time and a major amount
of missing information on individuals.
Table 11 Summary of Conversational Agent systems in different application areas

Application Area in Dataset/data

Ref. Description Techniques/ Methods used Challenges
Conversational Agents utilized
English user-chatbot Bi-LSTM with deep attention very unbalanced distribution of data.
Basile et To recognize emotional responses
interaction in the Indian mechanism, USE, BERT, Ensemble SemEval-2019 Accuracy is not up to the state of art systems
al. 2019 between a human and a Chatbot.
chat room. Model (RF, SVM, LR)
Adikari
Proposed cognitive model to detect Markov chains, word embedding, and service dialog
et al. Industrial chatbots System performance can be improved.
emotions in industrial chatbots. NLP using emotion vocabulary dataset
2019
The frequently
answered
Proposed a chatbot that responds to Artificial Intelligence Markup
Thomas questions from Only designed for FAQ’s.
E-business chatbots people using a frequently asked Language (AIML) and Latent
2016 the internet for General questions are not considered.
questions (FAQs) database. Semantic Analysis (LSA).
the e-business
domain
Developed a neural-based
Xue et al. Call center customer support conversational solution to boosts the Raw Slack Less labeled data used for training the intent
Bi-LSTM with an attention mechanism
2018 agent. effectiveness of customer support channel data classification model
agents.
CoachAI is a conversational agent-
assisted health coaching system that An ambulatory
Fadhil et Support Vector Machine (SVM), Finite- Activities suggested are not clinically validated and
Healthcare supports the delivery of health health care clinic
al. 2019 state Machine architecture. guided by a health care provider.
interventions to individuals or data was utilized.
groups.
Discussed smart learning contexts
Student Data was System created for a specific application domain.
and demonstrated a chatbot that, on
utilized from the There is no knowledge structure between the user
Hien et behalf of the academic staff, Named Entity Recognition (NER), Text
Education Ho Chi Minh City inputs and the responses generated.
al. 2018 responds to a question from students Classification
University of Only dealt with the user's context to a limited
regarding the services provided by
Science. extent.
the education system.
The second problem is that creating these annotated datasets is based on a particular
emotion model, so the dataset displays only those emotions present in that specific emotion
model. Again, the third problem is creating an annotated dataset with manual labeling is a
time-consuming and error-prone task. Datasets for emotion detection are nothing but user-
generated text, and it completely depends on the domain from which data is generated and
the source it has generated. TBED has spread the wing into application domains like online
social media and product/customer reviews. If we study, the writing style of online social
media is very informal, with lots of emojis, hashtags, etc. In contrast, product and
customer reviews are formal and explicate a particular product/service. Different online
datasets available to researchers have been discussed in section 3.3. These datasets have
some application domains adopted in them. When it comes to evaluating any learning
model on datasets, it performs well on specific application domains, but performance gets
degraded when tested on different application domains. The domains of the adopted
datasets with other details are depicted in Table 12.

4.4 RQ4 - What are the challenges and open problems with respect to
text-based emotion detection?

The authors have presented different datasets and broad categories of text-based emotion
detection applications from the literature survey. Though our study highlights the different
revelations of text-based emotion detection, there are still some challenges and research
directions that need to be addressed to align emotion detection with the requirements of the
real world. Table 13 shows the challenges identified. The following are some of the
research challenges and drawbacks in this field:
A major challenge in text-based emotion detection is related to datasets. Only a
few datasets are existing for research, and the majority of them are imbalanced datasets.
These datasets were created for specific analysis; therefore, application areas are
dependent. So, domain dependency of datasets is one of the challenges in text-based
emotion detection. Language dependency is also one of the major challenges. Most of the
research was done in English text corpora. Other languages such as Chinese and Arabic are
preferred after English. So, building multilinguistic systems is a challenging task.
The second challenge is the performance of machine learning and deep learning
techniques. These techniques have been used extensively in text-based emotion
recognition. Though, many machine learning algorithms need labeled datasets, that is a
tedious task and is reliant on the efficacy of humans, lowering the performance of machine
learning techniques. Deep learning approaches, on the other hand, are emerging, but they
are complex procedures that can keep word order and syntactic patterns. Deep learning
approaches still pose certain issues, like a vast amount of data required for training. Text-
based emotion detection is a multi-class problem where the efficiency and performance of
deep learning and machine learning techniques can be degraded. The third challenge in
TBED is the quality of text used in online media. Incomplete information, typing mistakes,
slang words, short texts, emojis, sarcasm, irony, harmony, etc., all these inconsistencies are
used in languages on online media by users. These inconsistencies make emotion detection
a challenging task. Other further challenges related to text semantics are the inability to
 Table 13 Overview of Datasets

Access
Application Balanced/ Annotated/ Type Multi
Dataset Granularity Size Emotion Labeled
Domain Unbalanced Not annotated Free/not Modality
free
ISEAR 7666 Anger, disgust, fear, sadness,
Social Media Blogs ×
Scherer et al. 1994 Sentences shame, joy, and guilt
Disgust, anger, fear, sadness, joy,
Alms’ Affect data 15302
-- Story Tales × positive surprise, and
Alm et al. 2008 sentences
negative surprise
SemEval 2016
Social Media, Blogs, Tweets 22761 × × Positive, Negative, Neutral
Rosenthal et al. 2019
SemEval 2007
Discussion Forums News articles 1,250 × × Anger, disgust, fear, sadness, joy, and surprise
Rosenthal et al. 2019
SemEval 2019
Conversations Dialogues 38424 × × Happy, sad, angry, other
Rosenthal et al. 2019
Crowdflower
https://data.world/crowdflower/ 40001 Boredom, anger, empty, fun, enthusiasm, happiness, love,
Social media Tweets × ×
sentiment--in-analysistext tweets hate, neutral, sadness, relief, worry, surprise.

Emobank
Social media Blogs 10062 × Valence, Arousal, Dominance
Buechel et al. 2017
Aman 4266 Sadness, happiness, anger, disgust, surprise, fear, no
Social media Blogs × × ×
Aman et al. 2007 Sentences emotion, and mixed emotion
The Valence and Arousal
3120
Facebook Posts Preotiuc et al. Social media Facebook × Valence, Arousal
posts
2016
MELD Dialogues/ anger, disgust, fear, joy, neutral, sadness, and surprise,
Conversations 15141 ×
Poria et. al. 2018 utterances neutral
CBET 76,860 Joy, anger, love, surprise, fear, sadness, thankfulness,
Social media Tweets (Twitter)
Ameneh et al. 2015 tweets guilt, and disgust
admiration, amusement, approval, anger, annoyance,
curiosity, caring, confusion, desire, disgust,
GoEmotions
Social media Blogs (Reddit) 58009 × disappointment, disapproval, excitement, embarrassment,
Dorottya et al. 2020
fear, grief, gratitude, joy, love, sadness, nervousness,
optimism, pride, remorse, realization, relief, surprise.
Amazon Blitzer et al. 2007 Review systems Product review 8000
Positive, Negative
reviews
IMDB Review Systems Movie Review 50000 ×
https://datasets.imdbws.com/ reviews Positive, Negative

The Movie Review Pang et al. Review Systems Movie Review 10,662 ×
positive, negative, or neutral
2002 sentences
recognize implicit emotions, unable to extract the semantic data, classifying emotions according to their
intensities, mislabeling emotions. These, too, are important challenges that need to be focused on by
researchers.
Table 12 Challenges identified through literature survey

Problem
Challenges in text-based emotion detection Future Directions
area
Lacking labeled/annotated datasets.
Domain Adaptation
Imbalanced datasets Transfer Learning
Datasets Domain dependent datasets GAN
Language dependent datasets Multi-modality in text and data sources

Mislabeled emotions
Several techniques are not robust
Auto-encoders
Accuracy
Accurateness of current systems Ensemble methods

Incomplete information, typing mistakes Multitask Learning System

Quality of Transfer learning
Slang words, short texts, emojis
text Pattern-based approach.
Detection of sarcasm, irony, harmony
Inability to recognize implicit emotions Pre-trained word embeddings
Best-Worst Scaling (BWS) Annotation
Semantics Unable to extract the semantic data
Scheme.
Categorizing emotions considering their intensities Deep Learning algorithms

5 Future Directions in Text-Based Emotion Detection

Techniques or methodologies in TBED have witnessed huge evolutions over the past several years.
Combining keyword-based approaches with n-hidden layer-based deep learning techniques, for example.
Artificial Intelligence advances in recent years have bolstered text-based emotion detection algorithms.
Some of the outstanding difficulties in text-based emotion detection have been addressed by AI-based
technologies such as data augmentation (generative adversarial networks), transfer learning, explainable
AI, autoencoders, adversarial machine learning, and different learning styles. Figure and table showcase
several challenges and their solutions using these methods with references.

5.1 Data Augmentation

In Natural language processing (NLP), many tasks like text classification to question answering require
vast training data to improve the performance of the model. However, acquiring and annotating additional
data can be an expensive and time-consuming process in order to generate a large amount of training data.
In this scenario, data augmentation techniques can be applied. Data augmentation is a machine learning
technique that artificially generates additional synthetic training data through label preserving
transformations (Bayer et al. 2021). Though data augmentation techniques have achieved great success in
computer vision applications, they are powerful in NLP applications. Data augmentation is achieved by
thesaurus, word embeddings, back translation, replacing synonyms, text generation, etc. In (Bayer et al.
2021), the authors proposed a novel text generation method to increase the performance of short and long
text classifiers using data augmentation techniques. The authors used GPT-2 355 million parameter
model for text generation in the implementation. The authors in (Shi et al. 2020) developed an efficient
data augmentation algorithm for labeled sentences called AUG-BERT for text classification. A deep
neural network-based text classifier based on data augmentation and semi-supervised learning built using
free-text obtained from a survey on sleep-related issues (Shim et al. 2020). One of the techniques trending
for data augmentation is GAN, explained in detail as follows –
Generative Adversarial Networks (GAN) – The lack of balanced data over all labels in the data is one of
the most critical issues in text-based emotion detection. Class imbalance causes over-representation of
majority classes, poor model performance, and classification models may disregard or treat small classes
as noise. GAN is the most recent technique for dealing with imbalance. Figure 15 shows the architecture
of the generative adversarial network. A generative adversarial network (GAN) is a new framework that
sets two neural networks against one other (thus the term "adversarial") to create new, synthetic data
instances that look like actual data. The discriminator evaluates whether each instance of data it examines
corresponds to the actual training dataset or not, while the generator generates new data instances. (Gray
et al. 2019) proposed spamGAN, a generative adversarial network for detecting opinion spam that uses
both labeled and unlabeled data. SpamGAN outperforms state-of-the-art GAN-based text classification
algorithms, according to the authors. TripAdvisor data was used in the experiments to show that
spamGAN outperforms existing techniques when labeled data is limited. GANs have recently been
suggested by a number of experts for unlabeled or imbalanced data (Li et al. 2020, Croce et al. 2020).

Fig. 15 Architecture of Generative Adversarial Networks

5.2 Transfer Learning

Data collection and labeling are time-consuming and frequently demand expert knowledge. Transfer
learning is an excellent technique to handle difficulties like data scarcity and a lack of human labeling.
Transfer learning allows you to reduce the amount of labeled data you need and take advantage of a
previously trained model to increase performance. The knowledge of the source domain can be
transferred to separate but related target domains using transfer learning. A set of data containing
adequate data samples, a large number of labels, and relatively high quality is referred to as a source

Fig. 16 A closer look at transfer learning

domain. A target domain's data, on the other hand, may contain a small number of samples, limited
amounts of data, or no labels at all, as well as being noisy. Transfer learning strategies try to increase the
learning of the target task using information from the source domain when given a source and a target
(Feng and Chaspuri 2019). Figure 16 gives a closer look at transfer learning. Transfer learning for
emotion recognition via cross-lingual embeddings was proposed by the authors in (Zishan et al. 2020).
Their developed a deep transfer learning framework using CNN and Bi-LSTM efficiently transfers the
captured relevant information through the shared space of English and Hindi. (Omara et al. 2019,
Devamanya et al. 2020, Yasaswini et al. 2021) proposed using a transfer learning approach for the
scarcity of data.

5.3 Explainable Artificial Intelligence (XAI)

Deep learning and machine learning models have drawback of not being human interpretable. These
models suffer the major problem of lack of transparency. The interpretability of AI systems is the topic
that is currently generating interest in the scientific community. The non-linear and nested structure of
deep neural networks, which enhances prediction model accuracy, makes it difficult to understand how
information flows from a human-understandable input such as an image or a piece of text to a predicted
output. Deep neural networks are regarded as “black boxes” in this sense. Explainable AI proposes that
models built-in machine learning or deep learning should explain their decision-making process and
actions. Very little work has been done in this area. Figure 17 presents the architecture of XAI. The
Authors in (Zucco et al. 2018)] highlighted the need for explainable sentiment analysis for medical
applications in order to extract human-reliable knowledge related to opinions and emotions from the user-
generated text information. In (Turcan et al. 2021) authors discussed detecting psychological stress in
online posts using a more explainable black-box model. (Nawshad et al. 2021) discussed Depression
Symptoms Detection (DSD) from text using Explainable Zero-shot modeling.

Fig. 17 Explainable Artificial Intelligence (XAI)

5.4 Auto-Encoders

These are a kind of neural network where feature vectors are transferred into an arbitrary higher or lower
dimensional space during the encoding phase, allowing the original feature vector to be rebuilt with
minimal reconstruction error in a subsequent decoding phase. Figure 18 shows the representation of auto-
encoders. Autoencoders can deal with missing data, improve system performance in noisy environments,
and be trained unsupervised. Denoising autoencoders, a form of autoencoder, have been employed in
domain adaptation, which involves training a model on one labeled corpus and then applying it to a
separate unlabeled corpus with a different context. In (Hesam et al. 2017) the authors used stacked
denoising autoencoders (SDAs) to perform sentiment recognition on a multitude of domains and
languages (Kumari et al. 2021) have used autoencoders to construct a deep learning model for detecting
distinct stages of violence (indirect, direct, and no aggression) from posts on social media in a multi-
lingual setting.
Fig. 18 Autoencoders

5.5 Adversarial Machine Learning (AML)

In numerous Natural Language Processing (NLP) tasks, many machine learning and deep learning models
have had remarkable success. On the other hand, machine learning and deep learning models, are
increasingly vulnerable to adversarial examples, according to a growing number of research publications.
Adversarial machine learning (AML) is an emerging research area that brings together the best machine
learning practices, robust statistics, and computer security (Deldjoo et al. 2021). The act of deploying
attacks towards machine learning-based systems and making systems robust to attacks is known as
Adversarial Machine Learning (AML). In adversarial machine learning, inputs are modified by creating
small perturbations to cause a malfunction in the target model, which outputs incorrect results. Figure 19
shows the adversarial machine learning model. The aim is to exploit the weaknesses or vulnerabilities of
the machine learning model. As a result, the model’s usefulness can significantly be reduced. The authors
in (Wei et al. 2019, Basemah et al. 2020) surveyed adversarial attacks on deep learning models in natural
language processing. While in (Deldjoo et al. 2021), discussed attack/defense strategies for adversarial
recommender systems, generative adversarial networks in detail. (Alsmadi 2021) examines features of
adversarial machine learning, specifically in text analysis and generation, and outlines key research
advances in the field, including GAN algorithms, models, attack types, and countermeasures.

5.6 Learning Styles

Machine learning is the learning from experience, that is, acquiring skills or knowledge from experience.
It means creating useful information from past data. There are many different learning styles depending
on how algorithms use multiple layers to extract increasingly higher-level features from the raw input.
Figure 20 shows different learning styles. These learning styles are becoming the advanced trends in text
classification or NLP applications. Table 14 gives the summary of learning styles.
1) Co-learning – Co-learning is based on a co-training concept. The co-training approach was first
introduced by (Blum and Mitchell 1998). Their approach proposed that multiple classifiers can learn from
each other using different features of data. First, all classifiers are trained on a small dataset (labeled) in a
completely supervised method in co-learning. Then each trained classifier is tested on unlabeled data.
This automatically
Fig. 19 working of Adversarial Machine Learning (AML)

labeled data, by each classifier, is then used to re-train the other classifiers in a semi-supervised method.
Co-learning is used when lack of training data or low/unavailability of labeled data. This learning strategy
can improve the generalization capacity of classifiers with learned features, resulting in better
classification models. The authors developed a unique technique for multi-domain text categorization
called dual adversarial co-learning (Wu et al. 2020). Their method uses shared-private networks to extract
both domain-invariant and domain-specific features, then uses the extracted features to train two
classifiers. Finally, these classifiers are co-learned on unlabeled data in an adversarial manner. With the
learned characteristics, the authors intended to improve the generalization capacity of the classifiers. The
credibility of Arabic web blogs is assessed in (Helwe et al. 2019) by utilizing a deep co-learning
approach. The authors showed a lack of training data in the Arabic web blogs co-learning approach
performed well with other baseline methods.
2) Multi-task learning - Multi-task learning is a method in which the model is trained simultaneously on
many tasks rather than separately on each task. The goal is to create a generalized model that can tackle a
variety of tasks. Specifically, information from the training signals of related tasks is utilized in multi-task
learning. To learn several linked tasks, use share representation. By sharing representations between
related tasks, Multi-Task Learning (MTL) facilitates model generalization. There are two mechanisms in
multi-task learning: 1) (a) the soft-sharing mechanism that applies a task-specific layer to different tasks.
(b) the hard-sharing mechanism that utilizes a shared feature space to extract features for different tasks.
Multi-task learning has the potential to make a model more robust in the limited dataset. The authors of
(Zhang et al. 2021) created a multi-task learning framework for sentiment analysis in several application
domains based on a hard-sharing approach. (Akhtar et al. 2020) proposes a multi-task learning
framework for finding and classifying aspect elements in a unified model. The authors used a Bi-
directional Long Short-Term Memory (Bi-LSTM) and a Convolutional Neural Network (CNN) network
in a cascade for collaborative learning to extract aspect terms and classify aspect sentiment in a multi-task
framework.
3) Zero-shot learning/Few shot learning – In most machine learning classification models, all data labels
are evaluated in the course of training, but, in some cases, it is necessary to categorize instances whose
classes are not observed at all or only a few times during training. Issues related to target classes are such
as a small number or a large number of classes, expensive class labeling, or changing classes over time.
At the time of the testing when there is no labeled example, it is known as zero-shot learning; at the time
of the testing when there is only one labeled example, it is termed one-shot learning; and at the time of the
testing when there are a few labeled instances, it is referred as few-shot learning. In (Thakkar et al. 2021),
the authors use zero-shot and few-shot learning to do a cross-lingual sentiment analysis of news articles.
The authors evaluated their system on Croatian and Slovene datasets using single-task and multi-task
environments. A multi-label few-shot learning system was proposed in (Hu et al. 2020) for Aspect
category detection (ACD) in sentiment analysis.
4) Meta-learning – To solve a novel problem, the meta-learning framework collects a large number of
tasks, each of which is treated as a training example, and then trains a model to adapt to all of those
training tasks. Finally, this model should perform well in the new assignment. Meta-Learning is an
algorithm that trains itself by learning to perform many tasks on training data before being tested on new
ones. Because it enhances data efficacy, allows learned information transfer, and facilitates unsupervised
learning. Meta-learning is crucial for both task-specific scenarios and task-agnostic. It is a method of
addressing challenges that makes advantage of the experience gained while completing related tasks. The
major objectives are to expedite the learning process and improve quantitative model performance.
(Madrid et al. 2019) highlighted the problem of impulsive recognizing the category of text classification
task using meta-features. The authors provided a collection of 73 meta-features that were tested in 81 data
sets related to five different types of tasks. (Yin 2020) surveyed meta-learning for a few shot NLP
applications.
5) Reinforcement learning – It is a machine learning technique where an agent is trained in an interactive
environment through reward and punishment mechanisms using feedback from its own actions and
experiences. The agent is rewarded for correct actions and punished for the wrong ones. By learning from
own actions and experiences, the agent tries to minimize wrong actions and maximize the right
ones. Reinforcement learning is used in which a sequence of decisions is needed. Most reinforcement
learning environments use Markov decision processes as a mathematical framework. In natural language
processing (NLP) applications, reinforcement learning is used for text summarization, Q&A, dialogue
systems, machine translation, etc. In (Duo et al. 2020), the authors proposed a framework to solve the
problem of meaningless labels. The authors presented a descriptive-based text classification that generates
class-specific descriptions using the reinforcement learning approach. Authors in (Chen 2019) developed
a framework to analyze the word-level and sentence-level sentiment analysis using reinforcement
learning.
 Multi-task Learning
Co-learning

Few-shot Learning
Meta Learning

Reinforcement Learning Self-supervised Learning

Fig. 20 Learning Styles
 Table 14 Summary of different learning style

Application
Ref. Learning Style Description Purpose
areas
Improve the
generalization
Multiple classifiers can
Wu et al. 2020 effectiveness of classifiers
learn from each other • Text classification
Helwe et al. Co-learning using learned features, as
using different features • Sentiment classification
2019 well as classification
of data
models based on
unknown data.
Model is trained on
• Stock Prediction
multiple tasks
Zhang et al. To make a model more • Multi-Domain Sentiment
simultaneously by
2021 Multi-task robust in a limited Classification,
sharing representations
Akhtar et al. Learning dataset. • Machine Translation
information instead of
2020 Faster performance. • Syntactic Parsing
training each task
• Microblog Analysis
separately.
• Sentiment Analysis
It can detect classes that • Multi-Label Text
Thakkar et al. To provide good
Zero/Few Shot the model has never Classification
2021 performance in unlabeled
Learning seen or seen very few • Machine Translation
Hu et al. 2020 datasets.
classes during training. • User Intent Classification
For Dialog Systems
To increase the • Text Classification
It learns to handle new quantitative performance• Machine Translation,
Madrid et al.
tasks when tested by of models by speeding up• Dialogue Generation
2019 Meta-Learning
performing numerous the learning process. • Intent classification
Yin et al. 2020
tasks on training data. Allows knowledge • Dialog state tracking
transfer. • Question answering
An agent is trained in an
interactive environment Used in automated goal-• Text summarization
Duo et al. 2020 through a reward and directed learning and • Question and answering
Reinforcement
Chen et al. punishment mechanism sequential decision- • Conversational System
Learning
2019 using feedback from its making. • Machine Translation
own actions and
experiences.
It trains the system itself To improve the
by precisely predicting performance of
Su et al. 2021 Self- • Classification task
the other part of the data classification from
Salim et al. supervised • Sentiment analysis
and generating labels unlabeled data and also
2021 learning • Text generation
using one part of the when a dataset is large to
data. label it.
6) Self-supervised learning – Self-supervision is another term for self-supervised learning. In this type of
learning, the model learns themselves by utilizing one part to anticipate the other of the data and
accurately create annotations; this is called a pretext task and then using representations learned from the
pretext task for other supervised tasks. Finally, the self-supervision learning method transforms an
unsupervised learning problem into a supervised learning problem. Self-supervised learning is used where
the dataset is too large, and labeling/annotating it will be time-consuming. This method makes data
labeling an automated process, and human intervention is eliminated. (Su et al. 2021) the authors
proposed an approach to automatically train attention supervision information for neural Aspect based
sentiment analysis. In (Salim et al. 2021) The authors created SSentiA (Self-supervised Sentiment
Analyzer), a self-supervised hybrid methodology for sentiment classification from unlabeled data that
combines a machine learning classifier with a lexicon-based strategy.

6 CONCLUSION

Digital online media has increased demand in recent years. This article provides a comprehensive review
of the literature on TBED using digital online platforms. In many applications, artificial intelligence
techniques have been applied to digital online media analysis. This review provides the outcomes of a
systematic literature review on TBED methods whereby the authors intended to highlight the employed
emotion models, adopted features extraction methods, approaches, the approved data sets, adopted
application domains, and the challenges with respect to TBED. In this study, different SLR phases were
planned, conducted, and executed on TBED. The research questions were based on criteria like datasets,
approaches, application domains, and investigating gaps in the literature. Existing literature shows
different artificial intelligent approaches used like deep learning-based, machine learning-based, rule-
based, and keyword-based approaches for TBED. Deep learning and machine learning-based approaches
are trending with the help of available datasets and automated feature extraction methods. The authors
surveyed publicly available datasets for TBED. In RQ 3, different applications domains are explored for
TBED research. Research challenges in the field of TBED are identified, such as domain dependency,
imbalanced datasets, and future directions for TBED based on artificial intelligence are presented.

References
Huang, MH., Rust, R.T. (2021). A strategic framework for artificial intelligence in marketing. J. of the Acad. Mark.
Sci. 49, 30–50 https://doi.org/10.1007/s11747-020-00749-9
Francisca Adoma Acheampong, Chen Wenyu, Henry Nunoo‐Mensah (2020), Text‐based emotion detection: Advances,
challenges, and opportunities. Wiley Online Library.
Chung W, Zeng D, (2018), Dissecting Emotion and User Influence in Social Media Communities: An Interaction Modeling
Approach, Information, and amp; Management,
https://doi.org/10.1016/j.im.2018.09.008
Sameh Al-Natour, Ozgur Turetken (2020), A comparative assessment of sentiment analysis and star ratings for consumer
reviews. Elsevier International Journal of Information Management 54,
https://doi.org/10.1016/j.ijinfomgt.2020.102132
S. Godfrey Winster, M. Naveen Kumar (2020), Automatic classification of emotions in news articles through Ensemble
decision tree classification techniques. Journal of Ambient Intelligence and Humanized Computing,
https://doi.org/10.1007/s12652-020-02373-5
Dimple Chehal, Parul Gupta, Payal Gulati, (2020), Implementation and comparison of topic modeling techniques based on
user reviews in e‑commerce recommendations, Journal of Ambient Intelligence and Humanized Computing, Part of
Springer Nature, https://doi.org/10.1007/s12652-020-01956-6
Graterol. W, Díaz-Amado J, Cardinale Y, Dongo I, Lopes-Silva E, Santos-Libarino C (2021) Emotion Detection for Social
Robots Based on NLP Transformers and an Emotion Ontology. Sensors, 21,1322. https://doi.org/10.3390/s21041322
M.R. Pacheco-Lorenzo, S.M. Valladares-Rodríguez, L.E. Anido-Rifón et al. (2020) Smart conversational agents for the
detection of neuropsychiatric disorders: A systematic review, Journal of Biomedical Informatics, doi:
https://doi.org/10.1016/j.jbi.2020.103632.
R. Bavaresco, D. Silveira, E. Reis, J. Barbosa, R. Righi, C. Costa, R. Antunes, M. Gomes, C. Gatti, M. Vanzin, S. Clair Jr., E
Silva, C. Moreira, (2020), Conversational agents in business: A systematic literature review and future research
directions, Computer Science Review 36, 100239, https://doi.org/10.1016/j.cosrev.2020.100239, 1574-0137, 2020
Elsevier
Joao Luis Zeni Montenegro, Cristiano Andréda Costa ∗, Rodrigo da Rosa Righi (2019) Survey of conversational agents in
health, Expert Systems with Applications 129, 56–67, 0957-4174, Elsevier https://doi.org/10.1016/j.eswa.2019.03.054
Kamil Topal, Gultekin Ozsoyoglu, (2016) Movie Review Analysis: Emotion Analysis of IMDb Movie Reviews,
IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining, 978-1-5090-2846-7/16
Bazzaz Abkenar S, Haghi Kashani M, Mahdipour E, Mahdi Jameii S (2020) Big data analytics meets social media: A
systematic review of techniques, open issues, and future directions, Telematics and Informatics, doi:
https://doi.org/10.1016/j.tele.2020.101517
Sunil Kumar, Arpan Kumar Kar, P. Vigneswara Ilavarasan (2021) Applications of text mining in services management: A
systematic literature review, International Journal of Information Management Data Insights, 2667-0968/2021, Elsevier,
https://doi.org/10.1016/j.jjimei.2021.100008.
Donggil Song, Eun Young Oh, Marilyn Rice (2017) Interacting with a Conversational Agent System for Educational
Purposes in Online Courses, 978-1-5090-4688-1/17, IEEE
Feng K and Chaspari T (2020) A Review of Generalizable Transfer Learning in Automatic Emotion Recognition. Front.
Comput. Sci. 2:9. doi: 10.3389/fcomp.2020.00009
Loitongbam Gyanendra Singh · Sanasam Ranbir Singh (2020) Empirical study of sentiment analysis tools and techniques on
societal topics, Journal of Intelligent Information Systems, https://doi.org/10.1007/s10844-020-00616-7
Tyng CM, Amin HU, Saad MNM and Malik AS (2017) The Influences of Emotion on Learning and Memory. Front.
Psychol. 8:1454. doi: 10.3389/fpsyg.2017.01454
Plutchik R. (1980) A general psycho evolutionary theory of emotion. Amsterdam, Netherlands: Elsevier; (pp. 3–33).
Paul Ekman, Wallace Friesen (1980) Facial Signs of Emotional Expressions Journal of personality and social psychology,
Vol 39, No. 6, 1125-1134.
Dellaert, F., Polzin, T. and Waibel, A (1996) Recognizing emotion in speech. In Proceedings of ICSLP 3, (Philadelphia, PA,
1996). IEEE, 19701973.
R. W. Picard, (1997) Affective Computing, M.I.T Media Laboratory Perceptual Computing Section Technical Report No.
321, MIT Media Laboratory; Perceptual Computing; 20 Ames St., Cambridge, MA 02139
Fellbaum C (1998) WordNet: an electronic lexical database. Language, speech, and communication. MIT Press, Cambridge.
Das, Sanjiv Ranjan, and Chen, Mike Y (2001) Yahoo! For Amazon: Sentiment Parsing from Small Talk on the Web
(August 5, 2001). EFA 2001 Barcelona Meetings, http://dx.doi.org/10.2139/ssrn.276189
Wiebe, J., Wilson, T., and Cardie, C (2005) Annotating expressions of opinions and emotions in language, Language
Resources and Evaluation (formerly Computers and the Humanities), pp. 1-54.
Vaswani, Ashish, and Shazeer, Noam and Parmar, Niki and Uszkoreit, Jakob and Jones, Llion and Gomez, Aidan N. and
Kaiser, undefined ukasz and Polosukhin, Illia (2017) Attention is All You Need, Proceedings of the 31st International
Conference on Neural Information Processing Systems, NIPS'17.
Alexander Ligthart · Cagatay Catal · Bedir Tekinerdogan (2021) Systematic reviews in sentiment analysis: a tertiary study,
Artificial Intelligence Review, https://doi.org/10.1007/s10462-021-09973-3
Nourah Alswaidan, Mohamed El Bachir Menai (2020) A survey of state-of-the-art approaches for emotion recognition in
text, Knowledge and Information Systems, https://doi.org/10.1007/s10115-020-01449-0
Akshi Kumar, Geetanjali Garg (2019) Systematic literature review on context-based sentiment analysis in social multimedia,
Multimedia Tools and Applications, https://doi.org/10.1007/s11042-019-7346-5
Seyeditabari, A., Tabari, N., & Zadrozny, W (2018) Emotion Detection in Text: a Review. ArXiv, abs/1806.00674.
Shervin Minaee, Nal Kalchbrenner, Erik Cambria, Narjes Nikzad, Meysam Chenaghlu, and Jianfeng Gao. (2021) Deep
Learning-based Text Classification: A Comprehensive Review. ACM Comput. Surv., 54, 3, Article 62 (June), 40
pages. DOI: https://doi.org/10.1145/3439726
Soujanya Poria, Erik Cambria, Rajiv Bajpai, Amir Hussain (2017) A review of affective computing: From unimodal analysis
to multimodal fusion, Information Fusion, Volume 37, Pages 98-125, ISSN 1566-2535,
https://doi.org/10.1016/j.inffus.2017.02.003.
Bostan, Laura-Ana-Maria and Klinger, Roman (2018) An Analysis of Annotated Corpora for Emotion Classification in Text,
Proceedings of the 27th International Conference on Computational Linguistics, Aug, 2104—2119,
https://aclanthology.org/C18-1179.
Dang, N.C.; Moreno-García, M.N.; De la Prieta, F (2020) Sentiment Analysis Based on Deep Learning: A Comparative
Study. Electronics, 9, 483. https://doi.org/10.3390/electronics9030483
Shilpi Goyal, Nirupama Tiwari (2017) EMOTION RECOGNITION: A LITERATURE SURVEY, International Journal
For Technological Research In Engineering Volume 4, Issue 9, May, ISSN (Online): 2347 – 4718
Kowsari, K.; Jafari Meimandi, K.; Heidarysafa, M.; Mendu, S.; Barnes, L.; Brown, D (2019) Text Classification Algorithms:
A Survey. Information, 10, 150. https://doi.org/10.3390/info10040150
Ali Yadollahi, Ameneh Gholipour Shahraki, and Osmar R. Zaiane (2017) Current State of Text Sentiment Analysis from
Opinion to Emotion Mining. ACM Comput. Surv. 50, 2, Article 25 (June 2017), 33 pages. DOI:
https://doi.org/10.1145/3057270
Sailunaz, K., Dhaliwal, M., Rokne, J. et al. (2018) Emotion detection from text and speech: a survey. Soc. Netw. Anal.
Min. 8, 28. https://doi.org/10.1007/s13278-018-0505-2
Kitchenham et al (2007) Guidelines for performing Systematic Literature Reviews in SE.
Moher D, Liberati A, Tetzlaff J, Altman DG, The PRISMA Group (2009) Preferred Reporting Items for Systematic Reviews
and Meta-Analyses: The PRISMA Statement. PLoS Med 6(7): e1000097. doi:10.1371/ journal.pmed.1000097
Published July 21, 2009
Ekman P. (1999) Basic emotions. Handbook Cognit Emot.; 98(45-60):16.
Plutchik R. (1980) A general Psycho Evolutionary theory of emotion. Amsterdam, Netherlands: Elsevier; (pp. 3–33).
Russell J. A (1980) A circumplex model of affect. J Pers Soc Psychol.;39(6):1161.
Scherer KR (2005) Appraisal theory. In: Dalgleish T, Power MJ (eds) Handbook of cognition and emotion. Wiley, New
York, pp 637–663
Ortony A, Clore GL, Collins A (1990) The Cognitive Structure of Emotions. Cambridge, MA: Cambridge University Press.
Russell JA, Mehrabian A (1977) Evidence for a three-factor theory of emotions. J Res Pers.;11(3):273-294.
Cambria E., Livingstone A., Hussain A. (2012) The Hourglass of Emotions. In: Esposito A., Esposito A.M., Vinciarelli A.,
Hoffmann R., Müller V.C. (eds) Cognitive Behavioural Systems. Lecture Notes in Computer Science, vol 7403.
Springer, Berlin, Heidelberg. https://doi.org/10.1007/978-3-642-34584-5_11
Tomás Mikolov, Kai Chen, Greg Corrado, Jeffrey Dean, (2013) Efficient Estimation of Word Representations in Vector
Space. ICLR (Workshop Poster)
Jeffrey Pennington and Richard Socher and Christopher D. Manning (2014) Glove: Global vectors for word representation,
In EMNLP.
Peters, Matthew & Neumann, Mark & Iyyer, Mohit & Gardner, Matt & Clark, Christopher & Lee, Kenton & Zettlemoyer,
Luke. (2018). Deep contextualized word representations.

Tao J (2004) Context-based emotion detection from text input, In Proceedings of the 8th international conference on spoken
language processing (ICSLP), , pp 1337–1340
Ma C, Prendinger H, Ishizuka M (2005) Emotion estimation and reasoning based on affective textual interaction. In: Tao J,
Tieniu T, Picard RW(eds) Affective computing and intelligent interaction. Springer, Berlin, pp 622–628
Perikos I, Hatzilygeroudis I (2013) Recognizing emotion presence in natural language sentences. In: Iliadis L, Papadopoulos
H, Jayne C (eds) Engineering applications of neural networks. Springer, Berlin, pp 30–39
Shivhare SN, Garg S, Mishra A (2015) Emotionfinder: detecting emotion from blogs and textual documents. In: International
conference on computing, communication & automation (ICCCA), pp 52–57
Lee SYM, Chen Y, Huang CR (2010) A text-driven rule-based system for emotion cause detection. In: Proceedings of the
theNAACLHLT 2010workshop on computational approaches to analysis and generation of emotion in text,
CAAGET’10. Association for Computational Linguistics, Stroudsburg, pp 45–53
Udochukwu O, He Y (2015) A rule-based approach to implicit emotion detection in text. In: Biemann C, Handschuh S,
Freitas A, Meziane F, Métais E (eds) Natural language processing and information systems. Lecture notes in computer
science. Springer, Cham, pp 197–203
H. Liu and M. Cocea (2017) Fuzzy rule-based systems for interpretable sentiment analysis, Ninth International Conference
on Advanced Computational Intelligence (ICACI), 2017, pp. 129-136, doi: 10.1109/ICACI.2017.7974497.
Aman S, Szpakowicz S (2007) Identifying expressions of emotion in text. In: Proceedings of the 10th international
conference on text, speech, and dialogue, TSD’07. Springer, Berlin, pp 196–205.
Ghazi D, Inkpen D, Szpakowicz S (2010) Hierarchical versus flat classification of emotions in text. In: Proceedings of the
NAACLHLT 2010workshop on computational approaches to analysis and generation of emotion in text,
CAAGET’10. Association for Computational Linguistics, Stroudsburg, pp 140–146
De Bruyne L, De Clercq O, Hoste V (2018) LT3 at SemEval-2018 task 1: a classifier chain to detect emotions in tweets. In:
Proceedings of the 12th international workshop on semantic evaluation. Association for Computational Linguistics, pp
123–127
Suhasini M, Srinivasu B. (2020 ) Emotion Detection Framework for Twitter Data Using Supervised Classifiers. New York,
NY: Springer (pp. 565–576).
Singh L, Singh S, Aggarwal N (2019) Two-stage text feature selection method for human emotion recognition. In: Krishna
CR, Dutta M, Kumar R (eds) Proceedings of 2nd international conference on communication, computing and
networking, lecture notes in networks and systems, vol 46. Springer, Singapore, pp 531–538
Allouch M, Azaria A, Azoulay R, Ben-Izchak E, Zwilling M, Zachor DA (2018) Automatic detection of insulting sentences
in conversation. Paper presented at Proceedings of the 2018 IEEE International Conference on the Science of Electrical
Engineering in Israel (ICSEE),1-4; IEEE.
Jonathan Herzig, Michal Shmueli-Scheuer, and David Konopnicki. (2017) Emotion Detection from Text via Ensemble
Classification Using Word Embeddings. In Proceedings of the ACM SIGIR International Conference on Theory of
Information Retrieval (ICTIR '17). Association for Computing Machinery, New York, NY, USA, 269–272.
DOI:https://doi.org/10.1145/3121050.3121093
Baziotis C, Pelekis N, Doulkeridis C (2017) Datastories at SemEval-2017 task 4: deep LSTM with attention for message-
level and topic-based sentiment analysis. In: Proceedings of the 11th international workshop on semantic evaluation
(SemEval-2017). Association for Computational Linguistics, pp 747-754
Ezen-Can A, Can EF (2018) RNN for affects at SemEval-2018 task 1: formulating affect identification as a binary
classification problem. In: Proceedings of the 12th international workshop on semantic evaluation. Association for
Computational Linguistics, pp 162–166
Basile A, Franco-Salvador M, Pawar N, Štajner S, Chinea Rios M, Benajiba Y (2019) “combined neural models for emotion
classification in human chatbot conversations(eds) Proceedings of the 13th international workshop on semantic
evaluation. Association for Computational Linguistics, Minneapolis, pp 330–334
Shrivastava K, Kumar S, Jain DK (2019) An effective approach for emotion detection in multimedia text data using
sequence-based convolutional neural network. Multimedia Tools Appl 78:29607–29639
Xiao J (2019) SemEval-2019 task 3: an ensemble of transfer learning methods for contextual emotion detection.
Proceedings of the 13th international workshop on semantic evaluation. Association for Computational Linguistics,
Minneapolis, pp 220–224
Rathnayaka P, Abeysinghe S, Samarajeewa C, Manchanayake I, WalpolaMJ, NawaratneR, Bandaragoda T, Alahakoon D
(2019) Gated recurrent neural network approach for multi-label emotion detection in microblogs. arXiv preprint
arXiv:1907.07653
W. Bian, C. Wang, Z. Ye, and L. Yan 2019 Emotional Text Analysis Based on Ensemble Learning of Three Different
Classification Algorithms 10th IEEE International Conference on Intelligent Data Acquisition and Advanced
Computing Systems: Technology and Applications (IDAACS), pp. 938-941, doi: 10.1109/IDAACS.2019.8924413.
Mohammad Ehsan Basiri, Shahla Nemati, Moloud Abdar, Erik Cambria, U. Rajendra Acharya (2021) ABCDM: An
Attention-based Bidirectional CNN-RNN Deep Model for sentiment analysis, Future Generation Computer Systems,
Volume 115, Pages 279-294, ISSN 0167-739X, https://doi.org/10.1016/j.future.2020.08.005.
José Ángel González, Lluís-F. Hurtado, Ferran Pla (2020) Transformer based contextualization of pre-trained word
embeddings for irony detection in Twitter, Information Processing & Management, Volume 57, Issue 4, , 102262,
ISSN 0306-4573, https://doi.org/10.1016/j.ipm.2020.102262.
F. Xia and Z. Zhang (2018 ) Study of text emotion analysis based on deep learning," 13th IEEE Conference on Industrial
Electronics and Applications (ICIEA), pp. 2716-2720, doi: 10.1109/ICIEA.2018.8398170.
Diogo Cortiz (2021) Exploring Transformers in Emotion Recognition: a comparison of BERT, DistillBERT, RoBERTa,
XLNet and ELECTRA, Computation and Language (cs. CL), arXiv:2104.02041
Taiao Liu, Yajun Du, and Qiaoyu Zhou (2020) Text Emotion Recognition Using GRU Neural Network with Attention
Mechanism and Emoticon Emotions. In Proceedings of the 2020 2nd International Conference on Robotics, Intelligent
Control and Artificial Intelligence (RICAI 2020). Association for Computing Machinery, New York, NY, USA, 278–
282. DOI:https://doi.org/10.1145/3438872.3439094
Polignano, Marco & Basile, Pierpaolo & de Gemmis, Marco & Semeraro, Giovanni (2019) A Comparison of Word-
Embeddings in Emotion Detection from Text using BiLSTM, CNN, and Self-Attention. 63-68.
10.1145/3314183.3324983.
Scherer KR, Wallbott HG (1994) Evidence for universality and cultural variation of differential emotion response patterning.,
J Personal Soc Psychol, 66(2):310–328.
Alm ECO (2008) Affect in Text and Speech. Citeseer. Urbana: the University of Illinois at Urbana-Champaign.
Rosenthal S, Farra N, Nakov P (2019) SemEval-2017 task 4: sentiment analysis in Twitter. arXiv preprint arXiv:1912.00741.
Buechel S, Hahn U (2017) Readers vs. writers vs. texts: coping with different perspectives of text understanding in emotion
annotation. Paper presented at Proceedings of the 11th Linguistic Annotation Workshop;:1-12.
Aman S, Szpakowicz S (2007) Identifying expressions of emotion in text. In: Proceedings of the 10th international
conference on text, speech, and dialogue, TSD’07. Springer, Berlin, pp 196–205.
Preotiuc-Pietro D, Schwartz HA, Park G, et al. (2016) Modelling valence and arousal in Facebook posts. Paper presented at
Proceedings of the 7th Workshop on Computational Approaches to Subjectivity, Sentiment, and Social Media Analysis
9-15.
Poria S, Hazarika D, Majumder N, Naik G, Cambria E, Mihalcea R. (2018) MELD: a multimodal multi-party dataset for
emotion recognition in conversations;. arXiv preprint arXiv:1810.02508.
Ameneh Gholipour Shahraki (2015) Emotion Mining from Text, Thesis, Master of Science
https://doi.org/10.7939/R3C53F63N
Dorottya Demszky, Dana Movshovitz-Attias, Jeong-woo Ko, Alan Cowen, Gaurav Nemade, Sujith Ravi, (2020)
GoEmotions: A Dataset of Fine-Grained Emotions, proceedings of the 58th Annual Meeting of the Association for
Computational Linguistics, pages 4040–4054, July 5 - 10
John Blitzer, Mark Dredze, Fernando Pereira (2007) Biographies, Bollywood, Boom-boxes, and Blenders: Domain
Adaptation for Sentiment Classification. Association of Computational Linguistics (ACL),.
Pang, Bo, and Lee, Lillian and Vaithyanathan, Shivakumar (2002) Thumbs up? Sentiment Classification using Machine
Learning Techniques, Proceedings of the 2002 Conference on Empirical Methods in Natural Language Processing,
EMNLP, Jul 2002, Association for Computational Linguistics, doi = "10.3115/1118693.1118704", pages 79-86
Eleni Adamopoulou, Lefteris Moussiades (2020) Chatbots: History, technology, and applications, Machine Learning with
Applications, Volume 2, 100006, ISSN 2666-8270, https://doi.org/10.1016/j.mlwa.2020.100006.
A. Adikari, D. De Silva, D. Alahakoon and X. Yu, (2019 ) A Cognitive Model for Emotion Awareness in Industrial Chatbots
IEEE 17th International Conference on Industrial Informatics (INDIN), 2019, pp. 183-186, doi:
10.1109/INDIN41052.2019.8972196.
N. T. Thomas (2016 ) An e-business chatbot using AIML and LSA, International Conference on Advances in Computing,
Communications and Informatics (ICACCI), 2016, pp. 2740-2742, doi: 10.1109/ICACCI.2016.7732476.
Z. Xue, T. Ko, N. Yuchen, M. D. Wu and C. Hsieh (2018) Isa: Intuit Smart Agent, A Neural-Based Agent-Assist Chatbot,
IEEE International Conference on Data Mining Workshops (ICDMW), 2018, pp. 1423-1428, doi:
10.1109/ICDMW.2018.00202.
Yueming Sun and Yi Zhang (2018) Conversational Recommender System. In The 41st International ACM SIGIR
Conference on Research & Development in Information Retrieval (SIGIR '18). Association for Computing Machinery,
New York, NY, USA, 235–244. DOI: https://doi.org/10.1145/3209978.3210002
FADHIL, Ahmed, Yunlong WANG, Harald REITERER (2019)Assistive Conversational Agent for Health Coaching: a
Validation Study. In: Methods of information in medicine. 58(1), pp. 009-023. ISSN 0026-1270. eISSN 2511-705X.
Available under: doi: 10.1055/s-0039-1688757
Joao Luis Zeni Montenegro, Cristiano André da Costa, Rodrigo da Rosa Righi (2019) Survey of conversational agents in
health, Expert Systems with Applications, Volume 129, Pages 56-67, ISSN 0957-4174,
https://doi.org/10.1016/j.eswa.2019.03.054.
Pacheco-Lorenzo MR, Valladares-Rodríguez SM, Anido-Rifón LE, Fernández-Iglesias MJ (2021) Smart conversational
agents for the detection of neuropsychiatric disorders: A systematic review. J Biomed Inform. Jan,113:103632. doi:
10.1016/j.jbi.2020.103632. Epub 2020 Dec 2. PMID: 33276112.
Pereira, J., Díaz, Ó. (2019) Using Health Chatbots for Behavior Change: A Mapping Study. J Med Syst 43, 135
https://doi.org/10.1007/s10916-019-1237-1
D. Song, E. Y. Oh and M. Rice (2017) Interacting with a conversational agent system for educational purposes in online
courses," 10th International Conference on Human System Interactions (HSI), 2017, pp. 78-82, doi:
10.1109/HSI.2017.8005002.
Hien, Ho & Pham-Nguyen, Cuong & Nam, Le & Nhung, Ho & Le Dinh, Thang. (2018). Intelligent Assistants in Higher-
Education Environments: The FIT-EBot, a Chatbot for Administrative and Learning Support. 69-76.
10.1145/3287921.3287937.
Jia, Jiyou. (2004). NLML--a Markup Language to Describe the Unlimited English Grammar.
Kashfia Sailunaz, Reda Alhajj (2019) Emotion and sentiment analysis from Twitter text, Journal of Computational Science,
Volume 36, 101003, ISSN 1877-7503, https://doi.org/10.1016/j.jocs.2019.05.009.
Pool, C., & Nissim, M (2016) Distant supervision for emotion detection using Facebook reactions. ArXiv, abs/1611.02988.
Douiji Yasmina, Mousannif Hajar, Al Moatassime Hassan (2016) Using YouTube Comments for Text-based Emotion
Recognition, Procedia Computer Science, Volume 83, Pages 292-299, ISSN 1877-0509,
https://doi.org/10.1016/j.procs.2016.04.128.
Kumari, K., Singh, J.P., Dwivedi, Y.K., et al. (2021) Bilingual Cyber-aggression detection on social media using LSTM
autoencoder. Soft Comput 25, 8999–9012. https://doi.org/10.1007/s00500-021-05817-y
Wang, D., Li, J., Xu, K. et al. (2017) Sentiment community detection: exploring sentiments and relationships in social
networks. Electron Commer Res 17, 103–132 https://doi.org/10.1007/s10660-016-9233-8
Song, L., Lau, R.Y.K., Kwok, R.CW., et al. (2017) Who are the spoilers in social media marketing? Incremental learning of
latent semantics for social spam detection. Electron Commer Res 17, 51–81. https://doi.org/10.1007/s10660-016-9244-
5
Gupta M, Bansal A, Jain B, Rochelle J, Oak A, Jalali MS. (2021) Whether the weather will help us weather the COVID-19
pandemic: Using machine learning to measure Twitter users' perceptions. Int J Med Inform.;145:104340.
doi:10.1016/j.ijmedinf.2020.104340
Zhoubao Sun, Lixin Han, Wenliang Huang, Xueting Wang, Xiaoqin Zeng, Min Wang, Hong Yan (2015) Recommender
systems based on social networks, Journal of Systems and Software, Volume 99, Pages 109-119, ISSN 0164-1212,
https://doi.org/10.1016/j.jss.2014.09.019.
Sandra Maria, Correia Loureiro, João Guerreiro, Iis Tussyadiah (2021) Artificial intelligence in business: State of the art and
 future research agenda, Journal of Business Research, Volume 129, Pages 911-926, ISSN 0148-
2963,https://doi.org/10.1016/j.jbusres.2020.11.001.
Bai Yang, Ying Liu, Yan Liang, Min Tang, (2019) Exploiting user experience from online customer reviews for product
design, International Journal of Information Management, Volume 46, Pages 173-186, ISSN 0268-
4012,https://doi.org/10.1016/j.ijinfomgt.2018.12.006.
Monireh Alsadat Mirtalaie, Omar Khadeer Hussain, Elizabeth Chang, Farookh Khadeer Hussain (2018) Extracting
Sentiment Knowledge from Pros/Cons Product Reviews Discovering features along with the polarity strength of their
associated opinions, Expert Systems With Applications, doi: 10.1016/j.eswa.2018.07.046
I. Chaturvedi, E. Cambria, S. Cavallari, and R. E. Welsch (2020 ) Genetic Programming for Domain Adaptation in Product
Reviews IEEE Congress on Evolutionary Computation (CEC), 2020, pp. 1-8, doi: 10.1109/CEC48606.2020.9185713.
Nantheera Anantrasirichai (2020) Artificial Intelligence in the Creative Industries: A Review, July.
M. Allouch, A. Azaria, R. Azoulay, E. Ben-Izchak, M. Zwilling and D. A. Zachor (2018) Automatic Detection of Insulting
Sentences in Conversation," 2018 IEEE International Conference on the Science of Electrical Engineering in Israel
(ICSEE), pp. 1-4, doi: 10.1109/ICSEE.2018.8646165.
Hasan, M., Rundensteiner, E. & Agu, E. (2019) Automatic emotion detection in text streams by analyzing Twitter data. Int J
Data Sci Anal 7, 35–51. https://doi.org/10.1007/s41060-018-0096-z
Ghanbari-Adivi, F., & Mosleh, M. (2019) Text emotion detection in social networks using a novel ensemble classifier based
on Parzen Tree Estimator (TPE). Neural Computing and Applications, 31(12), 8971-8983.
S. Shayaa et al. (2018) Sentiment Analysis of Big Data: Methods, Applications, and Open Challenges, IEEE Access, vol. PP,
no. c, p. 1.
Pradhan, J. Vala, and P. Balani (2016) A Survey on Sentiment Analysis Algorithms for Opinion Mining, Int. J. Comput.
Appl., vol. 133, vol. 133 no. 9, pp. 7-11, 2016
J. Kaur, (2012) Analysis and Identifying Variation in Human Emotion Through Data Mining, Int. J. Comput. Technol. Appl.,
vol. 3, no. 6, pp. 1963–1966.
X. Fang and J. Zhan (2015) Sentiment analysis using product review data, J. Big Data 2:5, pp. 1-14.
A. Segnini and J. J. T. Motchoffo (2019) Random Forests and Text Mining.
B. Abu-Salih, P. Wongthongtham, and K. Y. Chan (2018) Twitter mining for ontology-based domain discovery
incorporating machine learning,” J. Knowl. Management, vol. 22, no. 5, pp. 949-981.
A. Kaur and N. Duhan (2015) A Survey on Sentiment Analysis and Opinion Mining,” Int. J. Innov. Adv. Comput. Sci., vol.
4, no. May, pp. 107–116.
Wu, X., Kumar, V., Ross Quinlan, J. et al. (2008) Top 10 algorithms in data mining. Knowl Inf Syst 14, 1–37.
https://doi.org/10.1007/s10115-007-0114-2
Zahid Halim, Mehwish Waqar, Madiha Tahir (2020) A machine learning-based investigation utilizing the in-text features for
the identification of dominant emotion in an email, Knowledge-Based Systems, Volume 208, 106443, ISSN 0950-
7051, https://doi.org/10.1016/j.knosys.2020.106443.
C. C. Chen and Y.-D. Tseng (2011) Quality evaluation of product reviews using an information quality framework,” Decis.
Support Syst., vol. 50, no. 4.
N. A. Sharupa, M. Rahman, N. Alvi, M. Raihan, A. Islam and T. Raihan (2020) Emotion Detection of Twitter Post using
Multinomial Naive Bayes, 11th International Conference on Computing, Communication and Networking
Technologies (ICCCNT), 2020, pp. 1-6, doi: 10.1109/ICCCNT49239.2020.9225432.
H. Tang, S. Tan, and X. Cheng (2009) Expert Systems with Applications A survey on sentiment detection of reviews, Expert
Syst. Appl., vol. 36, no. 7, pp. 10760–10773.
Zhang, L.; Wang, S.; Liu, B. (2018) Deep learning for sentiment analysis: A survey. WIREs Data Min. Knowl. Discov. 8,
e1253.
Sutskever, I.; Martens, J.; Hinton, G.E. (2011) Generating text with recurrent neural networks. In Proceedings of the 28th
International Conference on Machine Learning (ICML-11), Bellevue, WA, USA, 28 June–2 July 2011; pp. 1017–
1024.
Britz, D. Recurrent Neural Networks Tutorial, Part 1–Introduction to RNN. http://www.wildml.com/2015/09/recurrent-
neural-networks-tutorial-part-1-introduction-to-rnns/
Hochreiter, S.; Schmidhuber, J. (1996) LSTM can solve hard long time lag problems. In Proceedings of the Advances in
Neural Information Processing Systems, Denver, CO, USA, 2–5 December 1996; pp. 473–479.
Graves, A.; Schmidhuber, J. Framewise phoneme classification with bidirectional LSTM and other neural network
architectures. Neural Netw. 18, 602–610.
Pascanu, R.; Mikolov, T.; Bengio, Y. (2013) On the difficulty of training recurrent neural networks. ICML 2013, 28, 1310–
1318.
Schuster, M.; Paliwal, K.K. (1997) Bidirectional recurrent neural networks. IEEE Trans. Signal Process. 1997, 45, 2673–
2681.
Chung, J.; Gulcehre, C.; Cho, K.; Bengio, Y. (2014) Empirical evaluation of gated recurrent neural networks on sequence
modeling. arXiv 2014, arXiv:1412.3555.
Farzi, R.; Bolandi, V. (2016) Estimation of organic facies using ensemble methods in comparison with conventional
intelligent approaches: A case study of the South Pars Gas Field, Persian Gulf, Iran. Model. Earth Syst. Environ. 2016,
2, 105.
Cho, K.; Merrienboer, B.; Gülçehre Ç.; Bougares, F.; Schwenk, H.; Bengio, Y. (2014) Learning Phrase Representations
using RNN. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP),
Doha, Qatar, 25–29 October 2014.
Devlin, J.; Chang, M.W.; Lee, K.; Toutanova, K. (2019) BERT: Pre-training deep bidirectional transformers for language
understanding. In Proceedings of the NAACL HLT 2019—2019 Conference of the North American Chapter of the
Association for Computational Linguistics: Human Language Technologies, Minneapolis, MI, USA, 2–7 June 2019.
Vaswani, A.; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Jones, L.; Gomez, A.N.; Kaiser, Ł.; Polosukhin, I. (2017) Attention is all
you need. Adv. Neural Inf. Process. Syst. 2017, 30, 5998–6008.
Sanh, V.; Debut, L.; Chaumond, J.; Wolf, T. (2020) DistilBERT, a distilled version of BERT: Smaller, faster, cheaper and
lighter. arXiv 2020, arXiv:1910.01108.
Liu, Y.; Ott, M.; Goyal, N.; Du, J.; Joshi, M.; Chen, D.; Levy, O.; Lewis, M.; Zettlemoyer, L.; Stoyanov, V. (2019)
RoBERTa: A Robustly Optimized BERT Pretraining Approach. arXiv 2019, arXiv:1907.11692.
Yang, Z.; Dai, Z.; Yang, Y.; Carbonell, J.; Salakhutdinov, R.; Le, Q.V. (2020) XLNet: Generalized Autoregressive
Pretraining for Language Understanding. arXiv 2020, arXiv:1906.08237.
Howard, Jeremy, and Sebastian Ruder (2018) Universal Language Model Fine-tuning for Text Classification.” ACL.
Radford, A., Narasimhan, K., Salimans, T., and Sutskever, I. (2018). Improving language understanding by generative pre-
training. URL https://s3-us-west-2. amazonaws. com/openai assets/researchcovers/languageunsupervised/language
understanding paper. pdf.
Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., and Sutskever, I. (2019). Language models are unsupervised multitask
learners. OpenAI Blog, 1(8).
Z. Yang, Z. Dai, Y. Yang, J. Carbonell, R. R. Salakhutdinov, and Q. V. Le, (2019) Xlnet: Generalized autoregressive
pretraining for language understanding," in Advances in neural information processing systems, pp. 5753-5763.
Liu, Y., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D., Levy, O., Lewis, M., Zettlemoyer, L., and Stoyanov, V. (2019).
Roberta: A robustly optimized Bert pretraining approach. arXiv preprint arXiv:1907.11692.
R. Tang, Y. Lu, L. Liu, L. Mou, O. Vechtomova, and J. Lin, (2019) Distilling task-specific knowledge from BERT into
simple neural networks, arXiv, p. 8,.
Xiang, Zheng & Du, Qinzhou & Ma, Yufeng & Fan, Weiguo, (2017) A comparative analysis of major online review
platforms: Implications for social media analytics in hospitality and tourism," Tourism Management, Elsevier, vol.
58(C), pages 51-65.
K. Topal and G. Ozsoyoglu (2016) Movie review analysis: Emotion analysis of IMDb movie reviews, IEEE/ACM
International Conference on Advances in Social Networks Analysis and Mining (ASONAM), 2016, pp. 1170-1176,
doi: 10.1109/ASONAM.2016.7752387.
Teja Kanchinadam, Zihang Meng, Joseph Bockhorst, Vikas Singh Kim, Glenn Fung, (2021) Graph Neural Networks to
Predict Customer Satisfaction Following Interactions with a Corporate Call Center
arXiv - CS - Machine Learning (IF), Pub Date: 2021-01-31, DOI: arxiv-2102.00420
Markus Bayer and Marc-André Kaufhold and Björn Buchhold and Marcel Keller and Jörg Dallmeyer and Christian Reuter,
(2021) Data Augmentation in Natural Language Processing: A Novel Text Generation Approach for Long and Short
Text Classifiers,2021, arXiv:2103.14453v1 [cs.CL].
Shi, Linqing & Liu, Danyang & Liu, Gongshen & Meng, Kui. (2020). AUG-BERT: An Efficient Data Augmentation
Algorithm for Text Classification. 10.1007/978-981-13-9409-6_266.
Heereen Shim, Stijn Luca, Dietwig Lowet, and Bart Vanrumste (2020) Data augmentation and semi-supervised learning for
deep neural networks-based text classifier. In Proceedings of the 35th Annual ACM Symposium on Applied
Computing, Association for Computing Machinery, New York, NY, USA, 1119–1126.
DOI:https://doi.org/10.1145/3341105.3373992
Gray Stanton and Athirai A. Irissappane, (2019) GANs for Semi-Supervised Opinion Spam Detection, arXiv:1903.08289v2
Li, Q., Yu, G., Wang, J. et al. (2020) A deep multimodal generative and fusion framework for class-imbalanced multimodal
data. Multimed Tools Appl 79, 25023–25050, https://doi.org/10.1007/s11042-020-09227-4
Croce, Danilo, Castellucci, Giuseppe, basili (2020) Roberto, GAN-BERT: Generative Adversarial Learning for Robust Text
Classification with a Bunch of Labeled Examples, Proceedings of the 58th Annual Meeting of the Association for
Computational Linguistics, 2020, Association for Computational Linguistics",doi = 10.18653/v1/2020.acl-
main.191,2114--2119
Feng K and Chaspari T(2019) Transfer learning and generalizability in automatic emotion recognition. Front. Comput. Sci.
2:9. doi:10.3389/fcomp.2020.00009
Zishan Ahmad, Raghav Jindal, Asif Ekbal, Pushpak Bhattacharyya, (2020) Borrow from a rich cousin: transfer learning for
emotion detection using cross-lingual embedding, Expert Systems with Applications, Volume 139, 112851, ISSN
0957-4174, https://doi.org/10.1016/j.eswa.2019.112851.
E. Omara, M. Mosa, and N. Ismail (2019) Emotion Analysis in Arabic Language Applying Transfer Learning," 2019 15th
International Computer Engineering Conference (ICENCO), pp. 204-209, doi: 10.1109/ICENCO48310.2019.9027295.
Devamanyu Hazarika, Soujanya Poria, Roger Zimmermann, Rada Mihalcea (2020) Conversational Transfer Learning for
 Emotion Recognition, Information Fusion, doi: https://doi.org/10.1016/j.inffus.2020.06.005
Yasaswini, Konthala, and Puranik, Karthik and Hande, Adeep and Priyadharshini, Ruba and Thavareesan, Sajeetha and
Chakravarthi, Bharathi Raja, (2021) Transfer Learning for Offensive Language Detection in dravidian Languages,
Proceedings of the First Workshop on Speech and Language Technologies for Dravidian Languages,2021, Association
for Computational Linguistics, https://aclanthology.org/2021.dravidianlangtech-1.25,187--194.
C. Zucco, H. Liang, G. D. Fatta and M. Cannataro (2018) Explainable Sentiment Analysis with Applications in Medicine,
IEEE International Conference on Bioinformatics and Biomedicine (BIBM), 2018, pp. 1740-1747, doi:
10.1109/BIBM.2018.8621359.
Turcan, Elsbeth and Muresan, Smaranda and McKeown, Kathleen (2021) Emotion-Infused Models for Explainable
Psychological Stress Detection, "Proceedings of the 2021 Conference of the North American Chapter of the
Association for Computational Linguistics: Human Language Technologies", year = "2021", publisher =
"Association for Computational Linguistics", url = "https://aclanthology.org/2021.naacl-main.230", doi =
"10.18653/v1/2021.naacl-main.230",2895--2909",
Farruque, Nawshad & Goebel, Randy & Zaïane, Osmar & Sivapalan, Sudhakar. (2021). STEP-EZ: Syntax Tree guided
semantic ExPlanation for Explainable Zero-shot modeling of clinical depression symptoms from the text.
Hesam Sagha, Nicholas Cummins, Björn Schuller (2017) WIREs Data Mining Knowl Discov, 7:e1212. doi:
10.1002/widm.1212
Deldjoo, Yashar and Noia, Tommaso Di and Merra, Felice Antonio, (2021) A Survey on Adversarial Recommender
Systems: From Attack/Defense Strategies to Generative Adversarial Networks, Association for Computing Machinery,
doi = {10.1145/3439729},
Wei Emma Zhang and Quan Z. Sheng and Ahoud Alhazmi and Chenliang Li, (2019) Adversarial Attacks on Deep Learning
Models in Natural Language Processing: A Survey, arXiv:1901.06796v3
Alshemali, Basemah and J. Kalita (2020) Improving the Reliability of Deep Neural Networks in NLP: A Review. Knowl.
Based Syst. 191: 105210.
Alsmadi, Izzat. (2021). Adversarial Machine Learning in Text Analysis and Generation.
Avrim Blum and Tom Mitchell. (1998) Combining labeled and unlabeled data with co-training. In proceedings of the
eleventh annual conference on Computational learning theory. Association for Computing Machinery, New York, NY,
USA, 92–100. DOI: https://doi.org/10.1145/279943.279962
Wu, Y., & Guo, Y. (2020). Dual Adversarial Co-Learning for Multi-Domain Text Classification. Proceedings of the AAAI
Conference on Artificial Intelligence, 34(04), 6438-6445. https://doi.org/10.1609/aaai.v34i04.6115
Helwe, Chadi & Elbassuoni, Shady & Zaatari, Ayman & El-Hajj, Wassim. (2019). Assessing Arabic Weblog Credibility via
Deep Co-learning. 10.18653/v1/W19-4614.
Zhang, J.; Yan, K.; Mo, Y. (2021) Multi-Task Learning for SentimentAnalysis with Hard-Sharing and TaskRecognition
Mechanisms. Information, 12, 207. https://doi.org/10.3390/info12050207
Md Shad Akhtar, Tarun Garg, Asif Ekbal, (2020) Multi-task learning for aspect term extraction and aspect sentiment
classification, Neurocomputing, Volume 398, Pages 247-256, ISSN 0925-2312,
https://doi.org/10.1016/j.neucom.2020.02.093.
Thakkar, Gaurish (2021) Multi-task Learning for Cross-Lingual Sentiment Analysis, CLEOPATRA Workshop 2021
https://doi.org/10.5446/52943
Hu, Mengting, Shiwan Zhao, Honglei Guo, Chao Xue, H. Gao, Tiegang Gao, Renhong Cheng and Z. Su. (2020) Multi-
Label Few-Shot Learning for Aspect Category Detection. ACL/IJCNLP (2020).
Madrid J.G., Escalante H.J. (2019) Meta-learning of Text Classification Tasks. In: Nyström I., Hernández
Yin, Wenpeng. (2020) Meta-learning for Few-shot Natural Language Processing: A Survey.” ArXiv abs/2007.09604
Duo Chai and Wei Wu and Qinghong Han and Fei Wu and Jiwei Li, (2020) Description Based Text Classification with
Reinforcement Learning, arXiv:2002.03067v3
Ruiqi Chen et al. 2019 IOP Conf. Ser.: Mater. Sci. Eng. 490 062063
Su, Jinsong & Tang, Jialong & Jiang, Hui & Lu, Ziyao & Ge, Yubin & Song, Linfeng & Xiong, Deyi & Sun, Le & Luo,
Jiebo. (2021). Enhanced Aspect-Based Sentiment Analysis Models with Progressive Self-supervised Attention
Learning.
Salim Sazzed, Sampath Jayarathna, (2021) SSentiA: A self-supervised Sentiment Analyzer for classification from unlabeled
data, Machine Learning with Applications, Volume 4, 00026, ISSN 2666-8270,
https://doi.org/10.1016/j.mlwa.2021.100026
https://data.world/crowdflower/sentiment--in-analysistext
https://datasets.imdbws.com/