Emotion recognition based mental heathcare chat bots a survey

International Conference on IoT based Control Networks and Intelligent Systems (ICICNIS 2020)

Emotion Recognition-Based Mental Healthcare Chat-bots: A Survey
*Carol Antony, Bestina Pariyath, Seema Safar, Aswin Sahil, Akash R Nair
Department of Computer Science and Engineering, Rajagiri School of Engineering and Technology, Kochi 682039, Kerala, India

ABSTRACT
The average human attention span has cascaded down fro m 12 to a mere 8 seconds. This is a direct consequence of man’s tech saturated lifestyle. Hu mans have shown a changing trend of preferring quick and informat ive dig ital conversations over a time consuming hu man-to-human interaction. Concurrently, a humongous increase in research being done on chat-bot technology is
seen. A well-trained chat-bot capable of having a fu lfilling and productive conversation has seen keen interest in users. Ever since
the outbreak of the Covid-19 pandemic our lives have been forced to change drastically. Difficu lty managing to adapt to the postCovid lifestyle has raised concerns about psychological resilience to adversity. The situation calls for immediate attention to
mental healthcare. In this survey, a study of the latest papers on how emotion recognition, in addit ion to sentiment analysis can be
integrated into a chat-bot to help identify and resolve a user's mental anguish is done. Th is survey is aimed at finding and
analysing the existing methods used to develop a self-sufficient emotion recognition-based chat-bot system that can take up the
role of a therapist.
Keywords: Chat-bot - Mental health - Natural language Understanding - Natural Language Generation - Deep Learning.

1. Introduction
1.1. Background
According to the World Health Organization (WHO), “Mental health is a state of well -being in wh ich an indiv idual realizes his
or her own ab ilit ies, can cope with the normal stresses of life, can work productively, and is ab le to make a contribution to h is or
her commun ity.” Despite it having a direct correlation with one’s physical well-being resulting in ill-health tendencies and high
mental mo rbidity rates [1], mental health is one of the most neglected areas of health care. The society still stigmatizes mental
health-related issues to a great extent even in this modern era. This discrimination certain ly discourages people facing such issues
fro m bringing them out in the open and they keep suffering on the inside. We need to recognize trau ma as a normal hu man
being’s response to an abnormal scenario and embrace people dealing with it. To pave the way for a healthier society, we need to
stop questioning the existence of people who suffer fro m some kind of mental dysfunction and start supporting them by
addressing their mental health issues [2].
The novel coronavirus outbreak has had a significant impact on the functioning of society as a whole and has evenly wreaked
havoc in both the rural and urban lives. A natural disaster of this intensity holds the capacity to affect the psychological
vulnerability of the mass population. Firstly, those who have contacted the virus and liv ing in care facilities or those plac ed at
high risk of carry ing the virus in them or the frontline health care wo rkers who have direct c ontact with patients are exposed to

developing mental disorders like an xiety disorder, post-traumat ic stress disorder or clinical depression. Also, this unsolicited
situation has led to touch starvation in many indiv iduals [3]. Humans crave physical touch as much as they need verbal attention.
The social distancing norm introduced during the pandemic has heightened concerns of people feeling deprived of affection,
because of limited or no physical contact. Besides the direct psychological d istress caused by the ailment, people dread the

69

Electronic copy available at: />

International Conference on IoT based Control Networks and Intelligent Systems (ICICNIS 2020)

thought of being left alone in isolation at homes or hospitals, which ultimately leads to unhealthy sleep and diet patterns [4]. The
fear of not being able to see their loved ones or confirm their well-being can really shake people up. The economic downfall is yet
another factor that has contributed greatly to emot ional instability among people. This has had adverse effects on the rural folk
who are economically less -privileged, due to uncertainty about their livelihood and scarcity of new job opportunities. This season
has witnessed a massive increase in the nu mber of intimate partner violence (IPV) [5] and suicide victims [6] as well. Therefo re,
mental health has become one of the most widely discussed subjects these days. In most issues related to mental health, the
affected merely need a space to vent out. The listener is only required to validate the person’s emotions by understanding th eir
situation and helping them come up with constructive solutions to tackle it.
Chatbots are being widely used and experimented on in recent years. The direction of research has slightly inclined towards
their potential in being emotionally intelligent [7]. The current scenario has increased the demand for such chatbots. Being isolated
fro m their loved ones and the sudden disruptive change in lifestyle has taken its toll on people globally and left a good number o f
them feeling mentally vulnerab le and unstable. But o wing to social stig ma and limited knowledge in the field of mental
healthcare, people are often scared to reach out to others. This is where a chatbot can come into play. The knowledge of
anonymity wh ile conversing with a chatbot enables the user to drop their facade and feel more co mfortable about opening up.
Most existing chatbots generally enable faster and easier access to knowledge for its users and are incapable of ascertaining or
reciprocating the emotions of the user. But, once we equip the chatbot with this quality, it can provide emotional support an d help
reduce the load on therapists at the same time. Therefore, the issues that rose or went out of hand owing to the pandemic or lack
of access to mental resources can be controlled to an extent with chatbots that can understand the user’s emotions and respon d the
way a human can.

1.2. Objective
The objective of this paper is to prov ide a co mprehensive overview of the existing research literature on the use of empathet ic
conversational agents. This paper discusses the most efficient methodologies available to enable a chatbot syste m to understand a
user's emotions from the input user utterance with emot ion recognition, by employing deep learn ing methods and Natural
Language Understanding (NLU). The paper aims to give further insight on how to generate a response that is emotionally
compatible to the user’s input, without losing contextual information, thus enabling the development of therapeutic chat -bots.
Finally, in line with the observed gaps in the literature, this paper seeks to provide reco mmendations for future research on design
and application of this technology.

2. Scope
Chatbots can be explicit ly used for diagnosis, prevention, in -time interventions and follow-up in the mental healthcare sector.
Research being conducted lately in the field suggests how suicidal tendencies and psychological distress can be perceived from
social media activity using Artificial Intelligence. In such cases where immediate act ion is required, the conversational age nts can
prompt the user to reach out to emergency helpline facilities or alert nearby therapists in order to prevent any mishaps. These
agents can also keep the user’s mental health stabilized by providing regular emotional support to avoid any chances of relap se
and even facilitate early screening and quicker diagnosis of psychological dis orders [9].

70

Electronic copy available at: />

International Conference on IoT based Control Networks and Intelligent Systems (ICICNIS 2020)

Fig. 1- Difference between conventional and empathetic chatbots
Adolescence is a crucial period for developing and maintaining social and emotional habits important for mental well-being
and youth these days rely extensively on s martphones. At this age, where the child is prone to all kinds of emotional vulnerability
and is in the process of transforming into an adult , the intervention of an emot ionally intelligent chatbot would be v ery
resourceful. The use of these agents is not just limited to youth but it can also be used by adults to relieve them of the stress and
anxiety that comes with the responsibility of tending to family and society, and help them cult ivate healthy physical habits so their

day becomes more productive. It can be utilized by NGOs and other social or rehabilitation workers to reach out to people in
need. In post-Covid times, where many have been fo rced to stay home, we have realized how such a technology holds the
potential to co mbat the ill-effects as well as make up for the lack of access to in-person healthcare facilities during natural
calamities like floods, virus outbreaks, earthquakes etc
Low and mediu m-inco me countries till date face shortage of qualified mental health professionals. The doctor to patient ratio
is too low. Also, there is no provision for insurance coverage for mental health services. Reaching out and seeking help is q uite an
expensive affair. These conversational agents can revolutionize mental health care in such areas and relax the burden on clinical
psychologists, by making needed facilities accessible to anyone, anywhere at their fingertips.

3. Phases of an empathetic chatbot
The interaction of a user with an empathetic chatbot [8] consists of precisely the following four phases:
1. Emotion expression by the user
2. Pre-processing of the input utterance
3. Sentiment analysis to detect the underlying emotion
4. Emotionally appropriate response generation

71

Electronic copy available at: />

International Conference on IoT based Control Networks and Intelligent Systems (ICICNIS 2020)

Fig. 2 - Phases of an empathetic chatbot

3.1 Emotion expression
In this init ial phase, the input utterance fro m the user is fed into the system. Emot ions are context -sensitive in nature. More
often, a sentence may express mu ltiple emotions simu ltaneously. Somet imes the emot ion might not be apparent and sometimes
the overall tone of the statement might be neutral. Words holding a strong emotional charge will allow the system to better
interpret the emot ional state of the user. Such contextual word associations and emotional information can be identified by
sentiment analysis, after pre-processing the statement. A human’s feelings and response toward stimu li are guided by emotional

mental models.
Setting the number of emotion categories to be used for classification of tokens at this stage itself will lessen the confusio n for
emotional model selection on the future phases. Most popular model availab le is the Ekman’s six basic emotions model [10] —
happiness, sadness, fear, anger, disgust, and surprise. Other less frequently used models for emot ion detection are Plutchik’ s
wheel of emotions [11] — 8 primary bipolar emot ions: joy versus sadness; anger versus fear; trust versus disgust; and surprise
versus anticipation. and advanced emotions based on the differences in intensity of the primary emotions and Parrott’s Emot iona l
Layers [12], consisting of thirty-one different emotions. Ekman’s model is preferred owing to higher accuracy in mental health
predictions and lesser categories, leading to a faster and more efficient system.

3.2 Pre-processing

Fig. 3 - Steps involved in pre-processing

72

Electronic copy available at: />

International Conference on IoT based Control Networks and Intelligent Systems (ICICNIS 2020)

At this stage, the statement must be broken down for further analysis and this is the fundamental step in any NLP method.
Provided a character sequence, tokenizat ion divides it into smaller, indivisible units, called to kens, also scraping off all unwanted
characters, repetitions and numbers. In case acronyms or messaging language are present in the input, we perform normalizat io n
to bring it down to its standard form. All negations are kept intact to preserve contextual information. After pre-processing,
machine learning models are utilized to optimally tokenize each word in the testing set.
Tokenization improves the machine’s learn ing rate and in ferences to a great extent. A good tokenization mechanis m is taking
a domain specific corpus and giving them higher scores and further normalizing the remaining dataset with a generic corpus.
Finally, the model assigns parts of each to each of the tokens and correlations are established.

3.3 Sentiment Recognition
The tokens identified during the pre-processing stage are embedded into binary vector representations, which contains the

mean ing corresponding to each word. These word vectors are then combined and into matrix representation of a sentence.
Attention mechanis ms are applied to filter out the crucial informatio n, thereby co mpacting the matrix while retaining the
contextual data. Finally, we predict the target sentiment of the statement . This can be done in following ways:

Fig. 4 - Steps involved in sentiment recognition in automatic and hybrid models

•

•

•

Rule Based Models: These models make use of a fin ite set of rules crafted by the system developer to pred ict the
sentiment/emotion within a statement. Primarily, each word in a statement is identified as one that shows or aids a sentiment
or as neutral words. Then using the rules that deal with all the sentiment deciders like sentiment lexicons, adjectives,
negators, intensifiers and more, the sentiment behind the statement is decided. These models wor k well if we know the exact
context of the user’s statement.
Automatic Models: Such models make use of machine learn ing techniques that learn fro m an input dataset. Sentiment
analysis is usually modified as a classifier algorith m. Here, we need to build a machine learning algorith m and feed it with
large amounts of input data to train it. The most widely proposed classifier models for this are Naive Bayes, Linear
regression, Support Vector Machines (SVMs) and deep learning models.
Hybrid Models: These models comb ine both automatic and rule -based models. They are built fo r improved accuracy. In this
type, the developer includes the beneficial co mponents from the other two models to build a better one. These types are
generally used when the developer cannot be very sure of the context of the user’s input but needs accurate outputs.

73

Electronic copy available at: />

International Conference on IoT based Control Networks and Intelligent Systems (ICICNIS 2020)

3.4 Response Generation
Once the sentiment behind the user’s input is identified, the chatbot needs to provide a matching human -like response to the
user. For this, the identified sentiment is fed to the response generator model along with the user’s input. The model then
generates a response that corresponds to the user’s emotion.
• Retrieval-based Models: This is an easier model wherein a predefined set of queries and responses are fed into the system and
it chooses a response that best matches the context as well as takes into account the emotion conveyed in the input utterance .
These models are best suited for do main-specific applicat ions like customer care for a certain business. The possible response set
is known in advance, hence there’s no ambiguity regarding the generated response.
• Generative Models: There are no predefined set of responses in these models as they generate comp letely new responses from
scratch. This approach is quite advanced and is seldom used in practice. It is difficult to develop bots using this model as it
requires a massive emotionally variant hu man-human interaction dataset and huge amounts of training time to achieve
contextually accurate responses. But recent years have seen a lot of advancement in this field.

4. Result Analysis

Fig. 5 – Basic architecture of the chatbot

4.1. Development Approaches
Table 1 shows a set of sentiment recognition methods used by researchers in the past few years for classifying various social
med ia co mments and tweets/posts while Tab le 2 shows recently developed response-generation approaches for generat ing
context-relevant and empathetic responses corresponding to user input utterances. The tables were created as a result of the
literature review conducted regarding the same.

74

Electronic copy available at: />

International Conference on IoT based Control Networks and Intelligent Systems (ICICNIS 2020)

4.2. Datasets available
Table 3 and 4 list out some of the public datasets available on Kaggle platform to facilitate sentiment analysis and more hu manlike response generation.

Table 1 – Sentiment Detection and Analysis Approaches
Authors

Year Focus

Suggested Approach

Chao Song et. Al.

2020 Short text sentiment analysis [13]

A framework that is based on probabilistic linguistic
terms called SAPC is created. It makes use of
probabilistic linguistic terms sets and support vector
machines to create the framework

Kashfia Sailunaz et. 2019 To conduct emotion and sentiment Naïve Bayes classifier is used to recognize emotion
Al.
analysis on twitter data [14]
as well as sentiment.
Shailendra Kumar
Singh et. Al.

2019 To classify social med ia texts using SentiVerb system – It uses the dictionary approach
sentiment analysis [15]
(opinion verb dictionary) and a binary classifier.

Mohsin Manshad
Abbasi et. Al.

2019 To summarize emotions from text [16]

Uses Plutchik’s wheel of emotions to classify the
emotion conveyed in the text.

Md. Rakibul Hasan 2019 To conduct sentiment analysis with NLP Uses Bag of Words and TF-IDF model concept to
et. Al.
on twitter data [17]
analyse the sentiment in the tweet.
Xian Zhong et. Al.

2019 To develop an emotion classification Sentiment categorization done by modificat ions to
algorithm based on SPT-CapsNet [18]
SPT-CapsNet (capsule network)

Ankush Chatterjee
et. Al.

2018 To understand emotions in text using SSBED – Sentiment and Semantic Based Emotion
deep learning and big data [19]
Detector model is proposed for sentiment
recognition.

Guixian Xu et. Al.

2017 To conduct sentiment analysis on social TF-IDF is used to process the comment and the
media comment texts [20]

resulting vectors are given as input to a Bi-LSTM
model to analyse the sentiment.

Anna Jurek et. Al.

2015 To develop an imp roved lexicon-based Rule based approach which considers almost all
sentiment analysis algorithm for social emotion determining factors.
media analytics [21]

Monisha Kanakaraj 2015 To conduct sentiment analysis using NLP based approach to strengthen the sentiment
et. Al.
ensemble classifiers on twitter data [22]
categorization by incorporating semantics in feature
vectors and consequently using ensemble methods
for the same.

75

Electronic copy available at: />

International Conference on IoT based Control Networks and Intelligent Systems (ICICNIS 2020)

Table 2 – Response Generation Approaches
Authors

Year Focus

Suggested Approach

Zhang et al.

2020 Retrieval-Polished Response Generation A retrieved response is polished by considering a
for Chatbot [23]
contextually similar p rototype and the better of the
retrieved and polished responses is chosen as the
final response. The method uses the background
provided by the context and the sentence style
provided by the retrieved response

Kim et al.

2020 Knowledge-Grounded Chatbot Based on
Dual
Wasserstein
Generative
Adversarial Net works with Effective
Attention Mechanisms [24]

Zhang et al.

2019 Dialo GPT: Large-Scale Generative Pre- It is an open-domain pre-trained model trained on
training for Conversational Response Reddit dataset and uses the mu lti-layer transformer
Generation [25]
architecture which handles issues like content
inconsistency, loss of contextual data and lack of
emotion encountered in other models.

Wu et al.

2019 A Sequential Matching Framework for A sequential matching framewo rk (SMF) for contextMulti-Turn Response Selection in response matching was introduced that can handle

Retrieval-Based Chatbots [26]
important info rmation in a context as well as model
the utterance relationships. The models were based
on convolution-pooling technique and an attention
mechanism

Gu et al.

2019 Dually Interactive Matching Network An IMN models the matching degree between a
(IMN) fo r Personalized Response context constituting mu ltiple utterances and a
Selection in Retrieval-Based Chatbots candidate response. The DIM model adopts a dual
[27]
matching arch itecture. It interactively matches
responses to contexts and personas respectively for
ranking
response candidates, giving
more
personalized responses

Yuan et al.

2019 Multi-hop Selector Network (MSN) for An MSN utilizes a mu lti-hop selector to select the
Multi-turn Response Selection in relevant utterances as context after wh ich the model
Retrieval-based Chatbots [28]
matches the filtered context with the candidate
response and returns a matching score. Th is helps
level down the problem of too much context in a
sentence.

Knowledge Grounded chatbots converse using

internal and external knowledge on a subject. Welldesigned attention mechanis ms were proposed to
reflect context without topic deviation

76

Electronic copy available at: />

International Conference on IoT based Control Networks and Intelligent Systems (ICICNIS 2020)

Su et al.

2018 Response selection and automatic
message-response
expansion
in
retrieval-based QA systems using
semantic dependency pair model
(SDPM) [29]

For response selection, an SDPM is constructed from
a matrix representing correlations between the
semantic dependencies of the messages and
responses. For database expansion, the unstructured
data from the message board of various
psychological
consultation
websites
are
automatically collected, fed to the QA system to find
the best matched response segment and scored to

verify if the new pair is good enough to be included
in the database.

Wu et al.

2018 Response selection with topic clues for A Topic aware attentive recurrent neural network
retrieval-based chatbots [30]
(TAARNN) was proposed in which topic information
is used to facilitate representations of the message
and response. RNNs are good at capturing the local
structure of a word sequence (syntactic and semantic)
and topic awareness reduces the issue of not being
able to handle long term dependencies

Wu et al.

2018 Learn ing Matching Models with Weak An unlabelled data set was created by retrieving
Supervision for Response Selection in response candidates and a weak annotator( preRetrieval-based Chatbots [31]
trained fro m large scale unlabelled hu man-hu man
conversations) was employed to provide matching
signals for the unlabelled input response pairs, which
helped supervise the learning of matching models
Table 3 - Datasets for Sentiment Recognition

Name of Dataset

Size

IMDB dataset (Sentiment
analysis) in CSV format

Amazon Reviews
Sentiment Analysis

Usability

Contains long movie reviews (>200 words) labelled with the sentiment
associated with the movie review

10

for 493.13MB Contains 3.6M A mazon product reviews, separated into two classes for
positive and negative reviews (1,2 stars-negative; 4,5 stars- positive)
for learning how to train fastText for sentiment analysis.

6.9

Stanford
Sentiment
Treebank v2 (SST2) [32]
Sentiment140 dataset

62.81MB

Description

46.5MB

Contains 11k+ sentences fro m movie rev iews parsed using Stanford
parser. All phrases were labelled by Mechanical Turk for sentiment

9.4

227.74MB Contains 1.6M t weets extracted using twitter API, labelled on a scale
of 0 to 4, indicating polarity of the tweet.

8.8

Emotions dataset for NLP

1.97MB

Contains list of documents with corresponding emotion flag

10

Sentiment Lexicons for
81 Languages [33]

1.96MB

Contains positive and negative sentiment lexicons for 81 languages.

7.6

77

Electronic copy available at: />

International Conference on IoT based Control Networks and Intelligent Systems (ICICNIS 2020)

Table 4 - Datasets for Response Generation Training
Name of Dataset
DialoGPT-large

Size

Description

4.52GB Contains trained model for a large-scale pretrained dialogue response
generation model. The model is trained on 147M mu lti-turn d ialogue fro m
Reddit discussion thread.

Ubuntu Dialogue Corpus 2.71GB Contains 26 million turns from natural two-person dialogues

Usability
8.8

7.6

5. Discussion
5.1. Principal findings
Chatbots have been developed since 1966 starting with Eliza, the very first chatbot created by Joseph Weizenbaum. The
program was designed to mimic a hu man conversation. In the decades that followed, chatbot developers have built upon this
model to strive for mo re hu man-like interactions. These were all based on pattern matching and substitution methodology to
simu late conversation. The 21st century has brought about new and exciting chatbots which showcase machine learn ing and other
advanced algorith ms, due to wh ich they are able to learn fro m their interactions with hu mans. Since then, these chatbots have
always attracted people. They have been used as personal assistants that can perform tasks on behalf of their users. They hav e
advanced so much that people even talk to chatbots for their o wn amusement. Chatbots have been used for assistance in
depression for a long time now. Chatbots possess both reactive i.e. instant response to an input and proactive behavior i.e.
constant healthy notifications, which is why they are apt as an alternative for face to face counselling. They can be developed to

be companions capable of identifying emotions and playing the role of a therapist. Wo mbat is one such chatbot developed for
Cognitive Behavioral Therapy (CBT). However, it has been found to lack empathy. Research is still being conducted to create a
human like chatbot which can provide users with an interaction wh ich is as close to a conversation with another human as
possible. Using the methodologies reviewed in this paper, it is possible to achieve more human ization of chatbots to give natural
and emotionally accurate responses.

5.2. Assumptions and limitations
Despite massive train ing, a bot is still a mach ine and it is difficu lt for a machine to capture and comprehend human emot ions in
its entirety. Th is can beco me harmful to the user if they beco me too attached to the chatbot and consider it as another human and
not as a bot. Emotional dependence on a non-living entity can become quite dangerous especially when the user’s mental stability
relies on it. The usage must be monitored and regulated by a dedicated institution. The uncomplicated limitat ions of chatbots
mainly includes the follo wing. Even after incorporating contextual awareness, it is hard to enable the chatbot to recognize sarcasm
and other such feelings which require good world knowledge and very good understanding of the user. The chatbot may also not
be able to p rocess long poetic statements which a good number of depressed people tend to use. Lack of high quality hu manhuman interaction dataset for train ing and the fact that most of the existing ones are biased in terms of gender, race, sexua l
orientations etc. which may generate biased conversations that do more harm than help is another chief reason which sets
researchers back in their path to develop a universally deliverable counselor bot. The chatbot won’t be able to give a proper
med ical diagnosis for users with serious mental health issues and rather detects it as stress or a temporary ill feeling. A chat bot
therefore, should not be developed with the intention to replace therapy but as a means to encourage a person to be more

78

Electronic copy available at: />

International Conference on IoT based Control Networks and Intelligent Systems (ICICNIS 2020)

conscious and aware of their moods, behavioral patterns, and potential trigger points, suggesting healthy options for dealing with
stress and depressive episodes, and reinforcing techniques learned in therapy like CBT, SFBT approaches.

5.3. Areas of improvement
The current research on emotionally aware chatbots can be extended into an interdisciplinary research, combin ing knowledge

in the fields of mental health, artificial intelligence and computational linguistics. More wo rk may be performed focused on
bringing about visible, tangible behavioral changes rather than merely provid ing the user with suggestions to tackle th eir
problems. Provision to handle the issue of mu ltiple languages and dialects of the locals majorly in developing countries. M ore
datasets may be created from actual counselling/therapy sessions while maintaining anonymity of patients involved. There’s also
a scope to utilize user engagement with smartphones to increase accuracy and comprehension of human behavior. A dedicated
regulating institution is desired for mon itoring the users, handling emergencies decisively, measuring effect iveness of the b ots and
being accountable for the well-being of the society. So me mechanisms to detect user satisfaction after using these chatbots may be
resourceful in better training of the deep learning models and hence maximize the functionality of the chatbots.

6. Conclusion
This work proposes ways to enable a chatbot system to understand the user’s emotion fro m their input utterance with emotion
recognition, by employing deep learning and Natural Language Understanding (NLU). Using the most efficient methodologies,
the chatbot will be equipped to generate natural, empathic responses which are well suited to the user’s input on an emotional
level. The chatbot can have constructive conversations with the user and in performing continuous comprehensive observation of
the user’s emotional and behavioral patterns. Such a chatbot can be deployed in situations which call for immed iate attention to
mental health post the Covid-19 outbreak and help people deal with adapting to the new normal. In future, such chatbots can be
further improved on by considering the mult ilingual aspect as well, making it available to the population in developing nations
with limited supply of mental health professionals. Research is also being conducted to incorporate Speech Emotion Recognit io n
techniques to identify the user’s emotion fro m the non-verbal characteristics of speech, even creating a mu ltimodal emot ion
recognition system and investigate clinical results and user satisfaction. This in turn will enable the development of many c hatbots
that can help counselors and therapists reach out to a larger number of patients.

REFERENCES
[1]

Pietrabissa, G., & Simpson, S. G. (2020). Psychological consequences of social isolation during COVID-19 outbreak. Frontiers in Psychology, 11, 2201.
Clement, S., Schauman, O., Graham, T., Maggioni, F., Evans-Lacko, S., Bezborodovs, N., ... & Thornicroft, G. (2015). What is the impact of mental health related stigma on help-seeking? A systematic review of quantitative and qualitative studies. Psychological medicine, 45(1), 11-27
[3]
Durkin, J., Jackson, D., & Usher, K. (2020). Touch in times of COVID‐19: Touch hunger hurts.
[4]

Loades, M. E., Chatburn, E., Higson-Sweeney, N., Reynolds, S., Shafran, R., Brigden, A., ... & Crawley, E. (2020). Rapid Systematic Review: The Impact of
Social Isolation and Loneliness on the Mental Health of Children and Adolescents in the Context of COVID-19. Journal of the American Academy of Child &
Adolescent Psychiatry.
[5]
Evans, M. L., Lindauer, M., & Farrell, M. E. (2020). A Pandemic within a Pandemic—Intimate Partner Violence during Covid-19. New England journal of
medicine.
[6]
Sher, L. (2020). T he impact of the COVID-19 pandemic on suicide rates. QJM: An International Journal of Medicine, 113(10), 707-712.
[7]
Pamungkas, E. W. (2019). Emotionally-aware chatbots: A survey. arXiv preprint arXiv:1906.09774.
[8]
Prakash, A. V., & Das, S. (2020). Intelligent Conversational Agents in Mental Healthcare Services: A Thematic Analysis of User Perce ptions. Pacific Asia
Journal of the Association for Information Systems, 12(2), 1.
[9]
Spring, T ., Casas, J., Daher, K., Mugellini, E., & Abou Khaled, O. (2019). Empathic Response Generation in Chatbots. In SwissT ext.
[10]
Ekman, P. (1999). Basic emotions. Handbook of cognition and emotion, 98(45-60), 16.
[2]

79

Electronic copy available at: />

International Conference on IoT based Control Networks and Intelligent Systems (ICICNIS 2020)

[11]

Mohsin, M. A., & Beltiukov, A. (2019, May). Summarizing Emotions from Text Using Plutchik’s Wheel of Emotions. In 7th Scientific Conference on
Information Technologies for Intelligent Decision Making Support (ITIDS 2019). Atlantis Press.
[12] Parrott, W. G. (2000). Emotions in social psychology: Key readings.

[13]
Song, C., Wang, X. K., Cheng, P. F., Wang, J. Q., & Li, L. (2020). SACPC: A framework based on probabilistic linguistic terms for short text sentiment
analysis. Knowledge-Based Systems, 105572.
[14]
Sailunaz, K., & Alhajj, R. (2019). Emotion and sentiment analysis from Twitter text. Journal of Computational Science, 36, 101003.
[15]
Singh, S. K., & Sachan, M. K. (2019). SentiVerb system: classification of social media text using sentiment analysis. M ultimedia Tools and Applications,
78(22), 32109-32136.
[16]
Abbasi, M. M., & Beltiukov, A. P. (2019). Summarizing Emotions from Text Using Plutchik’s Wheel of Emotions. In Advances in Intelli gent systems
research, In Proceedings of the 7th Scientific Conference on Information Technologies for Intelligent Decision-Making Support (ITIDS). Atlantis press (Vol. 166,
p. 291).
[17]
Hasan, M. R., Maliha, M., & Arifuzzaman, M. (2019, July). Sentiment Analysis with NLP on T witter Data. In 2019 International Conference on Computer,
Communication, Chemical, Materials and Electronic Engineering (IC4ME2) (pp. 1 -4). IEEE.
[18]
Zhong, X., Liu, J., Li, L., Chen, S., Lu, W., Dong, Y., ... & Zhong, L. (2020). An emotion classification algorithm based on SPT -CapsNet. Neural Computing
and Applications, 32(7), 1823-1837.
[19]
Chatterjee, A., Gupta, U., Chinnakotla, M. K., Srikanth, R., Galley, M., & Agrawal, P. (2019). Understanding emotions in text using deep learning and big
data. Computers in Human Behavior, 93, 309-317.
[20]
Xu, G., Meng, Y., Qiu, X., Yu, Z., & Wu, X. (2019). Sentiment analysis of comment texts based on BiLST M. Ieee Access, 7, 51522 -51532.
[21]
Jurek, A., Mulvenna, M. D., & Bi, Y. (2015). Improved lexicon-based sentiment analysis for social media analytics. Security Informatics, 4(1), 1-13.
[22]
Kanakaraj, M., & Guddeti, R. M. R. (2015, March). NLP based sentiment analysis on T witter data using ensemble classifiers. In 2015 3rd International
Conference on Signal Processing, Communication and Networking (ICSCN) (pp. 1-5). IEEE.
[23]
Zhang, L., Yang, Y., Zhou, J., Chen, C., & He, L. (2020). Retrieval-Polished Response Generation for Chatbot. IEEE Access, 8, 123882-123890.

[24]
Kim, S., Kwon, O. W., & Kim, H. (2020). Knowledge-Grounded Chatbot Based on Dual Wasserstein Generative Adversarial Networks with Effective
Attention Mechanisms. Applied Sciences, 10(9), 3335.
[25]
Zhang, Y., Sun, S., Galley, M., Chen, Y. C., Brockett, C., Gao, X., ... & Dolan, B. (2019). DialoGPT : Large-scale generative pre-training for conversational
response generation. arXiv preprint arXiv:1911.00536.
[26]
Wu, Y., Wu, W., Xing, C., Xu, C., Li, Z., & Zhou, M. (2019). A sequential matching framework for multi-turn response selection in retrieval-based chatbots.
Computational Linguistics, 45(1), 163-197.
[27]
Gu, J. C., Ling, Z. H., Zhu, X., & Liu, Q. (2019). Dually interactive matching network for personalized response selection in retrieval-based chatbots. arXiv
preprint arXiv:1908.05859.
[28]
Yuan, C., Zhou, W., Li, M., Lv, S., Zhu, F., Han, J., & Hu, S. (2019, November). Multi-hop selector network for multi-turn response selection in retrievalbased chatbots. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on
Natural Language Processing (EMNLP-IJCNLP) (pp. 111-120).
[29]
Su, M. H., Wu, C. H., Huang, K. Y., & Lin, W. H. (2018). Response selection and automatic message-response expansion in retrieval-based QA systems using
semantic dependency pair model. ACM Transactions on Asian and Low-Resource Language Information Processing (T ALLIP), 18(1), 1-24.
[30]
Wu, Y., Li, Z., Wu, W., & Zhou, M. (2018). Response selection with topic clues for retrieval-based chatbots. Neurocomputing, 316, 251-261.
[31]
Wu, Y., Wu, W., Li, Z., & Zhou, M. (2018). Learning matching models with weak supervision for response selection in retrieval-based chatbots. arXiv preprint
arXiv:1805.02333.
[32]
Socher, R., Perelygin, A., Wu, J., Chuang, J., Manning, C. D., Ng, A. Y., & Potts, C. ( 2013, October). Recursive deep models for semantic compositionality
over a sentiment treebank. In Proceedings of the 2013 conference on empirical methods in natural language processing (pp. 1631-1642).
[33]
Chen, Y., & Skiena, S. (2014). Building Sentiment Lexicons for All Major Languages. In ACL (2) (pp. 383-389).

80

Electronic copy available at: />