Building Pedagogical Conversational Agents, Affectively Correct

Michalis Feidakis

, Panagiotis Kasnesis

, Eva Giatraki

, Christos Giannousis

Charalampos Patrikakis

and Panagiotis Monachelis

Department of Electrical & Electronics Engineering, University of West Attica, Thivon 250 & Petrou Ralli street,

Egaleo GR 12244, Greece

Primary School of Egaleo (K. Kavafis), 6 Papanikoli St, Egaleo GR 12242, Greece

giannousis@uniwa.gr, bpatr@uniwa.gr, pmonahelis@uniwa.gr

Keywords: Conversational Agents, Chatbots, Pedagogical Agents, Affective Feedback, Artificial Intelligence, Deep

Learning, Affective Computing, Reinforcement Learning, Sentiment Analysis, Core Cognitive Skills.

Abstract: Despite the visionary tenders that emerging technologies bring to education, modern learning environments

such as MOOCs or Webinars still suffer from adequate affective awareness and effective feedback

mechanisms, often leading to low engagement or abandonment. Artificial Conversational Agents hold the

premises to ease the modern learner’s isolation, due to the recent achievements of Machine Learning. Yet, a

pedagogical approach that reflects both cognitive and affective skills still remains undelivered. The current

paper moves towards this direction, suggesting a framework to build pedagogical driven conversational agents

based on Reinforcement Learning combined with Sentiment Analysis, also inspired by the pedagogical

learning theory of Core Cognitive Skills.

1 INTRODUCTION

During the last decade, ubiquitous learning (MOOCs,

Webinars, Mobile Learning, RFIDs & QRcodes, etc.)

has made significant steps towards the sought

democratization of education (Caballé and Conesa,

2019), also drifted by the latest ICT advancements

(i.e. IoT, Cloud Computing, 5G) (El Kadiri et al.,

2016). However, the auspicious new educational

paradigms fail to encounter distance learning (d-

learning) or e-learning common “headaches” such as

low students’ engagement and poor immersion rates

(Afzal et al., 2017). The rising technology of

Affective Computing (AC) promises to contribute

significantly towards the motivation and engagement

of the isolated remote learners (Graesser, 2016).

AC has emerged as a new research field in

Artificial Intelligence (AI) designated to “sense and

respond” to user’s affective state (Picard, 1997) in

various domains, and especially in education, in

which addresses three key issues (Feidakis, 2016):

1. Detect and recognize the learner’s emotion/

affective state with high accuracy.

2. Display emotion information through effective

visualisations for self, peers and others’ emotion

awareness.

3. Provide cognitive and affective feedback to

improve task & cognitive performance, as well

as learning outcomes.

The latter (affective feedback) remains

unexplored, still constituting a big challenge: How to

contend with the learner’s isolation encountered in

Virtual Learning Environments (VLEs), often leading

to dropouts, while system’s feedback suffers from:

 Empathy – The system is usually unaware of the

learner’s affective state and affect transitions, i.e.

the student’s confusion steadily increases, and

possibly transit to frustration (Afzal et al., 2017).

 Brevity – Response is outdated because there is

no need to address the problem anymore, or

worse, the student left the building (Feidakis,

2016).

 Sociality – New learning tools (MOOCs) often

isolate the learner who ultimately ends up in

cognitive deadlocks (Caballé & Conesa, 2018).

The recent findings of Deep Learning –especially

when applied to the fields of Natural Language

Processing (NLP) (Devlin et al., 2018) and

Reinforcement Learning (RL) (Silver et al., 2018) –

premises the capacity to develop smart, artificial

agents able to manipulate the above-mentioned

issues. Nevertheless, the following questions arise:

100

Feidakis, M., Kasnesis, P., Giatraki, E., Giannousis, C., Patrikakis, C. and Monachelis, P.

Building Pedagogical Conversational Agents, Affectively Correct.

DOI: 10.5220/0007771001000107

In Proceedings of the 11th International Conference on Computer Supported Education (CSEDU 2019), pages 100-107

ISBN: 978-989-758-367-4

1. How we built agents that are triggered

consistently? –in terms of transparency in

affect detection and brevity in response.

2. How we train agents to reply adequately? –

regarding the impact of response.

In the current paper, we present a framework

model to build agents based on Deep RL models,

empowered by Question-Answering (QA)

pedagogical learning strategies, also showing some

respect to the respondent’s feelings through

Sentiment Analysis of short dialogues. We first

review current State-of-the-Art of Artificial/

Conversational Agents, Chatbots and Affective/

Empathetic Agents, together with Sentiment Analysis

of short texts and posts. Then, we present our

proposal towards the enquiries set. This new

framework will guide the implementation of a new

model in our future steps, also prompting for more

contributions and collaborations.

2 LITERATURE REVIEW

2.1 Artificial Conversational Agents

and Chatbots

There have been almost three decades since intelligent

agents have been introduced, and AI researchers

started to look at the “whole agent” problem (Russell

et al., 2010). Nowadays, “AI systems have become so

common in Web-based applications that the -bot suffix

has entered everyday language” (Russell et al., 2010,

p. 26). The consistent evolution of AI together with the

prompt establishment of IoT have coined many new

areas to invest such as smart homes, smart cities and

recently smart advisors or virtual assistants (Walker,

2019). Artificial Conversational Agents are software

agents trying to answer a particular question through

information retrieval techniques and by engaging the

user into understanding the nature of the problem

behind the question.

Chatbots are one category of Conversational

Agents (Radziwill and Benton, 2017), hence,

applications that use AI techniques to communicate

and process data provided by the user. Chatbots look

for keywords into user’s questions, trying to

understand what the user really wants. With the help

of AI, Chatbot applications keep evolving based on

historical data to cope with new input information.

Chatbots such as Amazon Alexa (“Amazon Alexa,”

2019), Google Assistant (“Google Assistant,” 2019),

Apple’s Siri (“Apple Siri,” 2019), and Facebook

Jarvis (Zuckerberg, 2016) are steadily increase their

market share in people’s everyday applications.

There are several tools based on information

retrieval techniques to create a Chatbot agent that

come with respective developer’s tools. Dialogflow

(“Dialogflow,” 2019) consists of 2 related compo-

nents: (i) intents, to build arrays of various questions

(e.g. Where is the Gallery?), and (ii) entities, sets of

words (tokens) that help the agent to analyse spoken

or written text (e.g. product names, street names,

movie categories). It also provides an API to integrate

the application in certain media or services such as

Facebook Messenger, WhatsApp, Skype, SMS etc.

IBM’s Watson (“IBM Watson,” 2019) works

similarly to Dialogflow, allowing the user to connect

to many other applications through Webhook or

APIs, however it requires subscription (the free

version is limited up to 1000 intents per month).

In education, Chatbots have been recently applied

to various disciplines such as physics, mathematics,

languages and chemistry, usually inspired by

gamification pedagogical strategies. For instance,

through machine learning-oriented Q&As, a Chatbot

is able to evaluate the learner’s level each time and

fine-tunes the game’s level of difficulty (Benotti,

Martínez, & Schapachnik, 2014).

2.2 Affective Agents

The provision of affective feedback to users, in

response to their implicit (automatically and

transparently) or explicit recognition of affective

state, remains prominent topic. Learners need to

perceive a reaction from the system, in agreement

with their emotion sharing, immediately or after a

short period (Feidakis et al., 2014), triggering an

affective loop of interactions (Castellano et al., 2013).

Since 2005, “James the butler” (Hone et al.,

2019), an Affective Agent developed in Microsoft

Agents and Visual Basic, managed to reduce negative

emotions’ intensity (despair, sadness, boredom).

Similarly, Burleson (2006)) developed an Affective

Agent able to mimic facial expressions. The agent

managed to help 11-13 years old students to solve

cognitive problems (i.e. Tower of Hanoi) by

providing affective scaffolds in case i.e. of despair.

AutoTutor comes with a long list of published results

(D’Mello et al., 2011) successfully demonstrating

both cognitive and affective skills (empathy). Moridis

and Economides (Moridis and Economides, 2012)

report on the impact of Embodied Conversational

Agents (ECAs) on the respondent’s affective state

(sustain or modify) though corresponding empathy,

either parallel (express harmonized emotions) or

reactive (stimulate different or even contradictory

emotions). EMOTE is another tool that can be

Building Pedagogical Conversational Agents, Affectively Correct

101

integrated in existing agents, enriching their emotiona-

lity towards their improved learning performance and

emotion wellbeing (Castellano et al., 2013).

In (Feidakis et al., 2014), a virtual Affective

Agent provided affective feedback enriched with

task-oriented scaffolds, according to fuzzy rules. The

agent managed to improve student cognitive

performance and emotion regulation. A visionary

review of artificial agents that simulate empathy in

their interactions with humans is provided by Paiva et

al. (2017), delivering sufficient evidence about the

significant role of affective or empathetic responses

when (or where) humans, agents, or robots are

collaborating and executing tasks together.

2.3 Sentiment Analysis

Sentiment Analysis plays a significant role on

products marketing, companies’ strategies, political

issues, social networking, as well as on education.

Main goal is to mine for opinions in textual data e.g.

a web source, and extract information about author’s

sentiments. The derived data are oriented mainly to

polarity depiction of the sentiments using data mining

and NLP techniques (Cambria et al., 2013).

There are 3 main approaches to implement

Sentiment Analysis:

 Linguistic Approach: A sentiment is

exported from text analysis according to a

lexicon. In this case, three levels of text

(document-level, sentence-level, aspect-level)

are analysed and algorithms based on lexicons

produce the sentiment score (Feldman, 2013).

 Machine Learning Approach: Employs both

supervised and unsupervised learning towards

the classification of texts. It has become quite

popular since it is used for many applications,

such as movie review classifier (Cambria,

2016). Different algorithms have been applied

over the years, such as Support Vector

Machines, Deep Neural Networks, Naïve

Bayes, Bayesian Network and Maximum

Entropy MaxEnt, with Deep Learning

approaches appearing to give better results

(Howard and Ruder, 2018).

 Hybrid Approach: A reconciliation of the

two above methods (Cambria, 2016).

Nowadays, State-of-the-Art Sentiment Analysis

techniques are based on supervised machine learning

approaches and in particular Deep Learning

algorithms. Most of these approaches use word

embeddings, such as word2vec (Mikolov et al.,

2013), which are vectors for representing the words,

and are trained in an unsupervised way. Afterwards,

they are fed to a deep Recurrent Neural Network

(RNN) (Rumelhart et al., 1987) – a family of neural

networks used for processing sequence values–

empowered with a mechanism named LSTM (Long

Short-Term Memory) (Hochreiter and Schmidhuber,

1997) to enhance the memory of the network. In

addition, recent advances (Devlin et al., 2018)

exploiting the attention mechanism (Vaswani et al.,

2017) –focusing more on some particular words in a

sentence– seem to achieve higher accuracy in the

Sentiment Analysis task.

In education context, students’ textual data could

be analysed and provide valuable feedback to an

agent, who can act interactively, either by rewarding

a successful task accomplishment or encouraging the

students after a failed attempt. Τhe application of the

aforementioned techniques to short posts extracted

from educational platforms (i.e. LMS, MOOC), could

provide useful information regarding the correlation

between students’ sentiments and performance

(Tucker et al., 2014).

3 ARCHITECTURAL APPROACH

According to (Russell et al., 2010), the definition of

an AI agent involves the concept of rationality: “a

rational agent should select an action that is expected

to maximize its performance measure, given the

evidence provided by the percept sequence and

whatever built-in knowledge the agent has” (p. 37). A

learning agent comprises two main components: (i)

the learning element, for making improvements, and

(ii) the performance element, for selecting external

actions. In other words, the performance element

involves the agent decisions, while the learning

element deals with the evaluation of those actions.

Designing an agent requires the selection of an

appropriate and effective learning strategy to train the

agent. Learning strategies and models that have been

validated for years, in real education settings, could

extend the design horizons of artificial pedagogical

agents. In next paragraphs we unveil this perspective.

A popular strategy to guide a learner in alternative

learning paths is to use Question-Answer (QA)

models, which are quite easy to implement, especially

when short default answers are deployed (Yes/No,

Multiple choices, Likert-scales, etc.). In learning

settings, QA models constitute a common way to

deploy formative assessment i.e. according to a rubric.

Also, they can be nicely shaped as valuable scaffolds

to unlock preexisting knowledge or semi-completed

conceptual schemas (Vygotsky, 1987). In all cases, the

overload shifts to formulate the right questions.

CSEDU 2019 - 11th International Conference on Computer Supported Education

102

In their Academically Productive Talk (APT)

model, Tegos and Demetriadis (2017) emphasize the

orchestration of teacher-students talks and highlight a

set of useful discussion practices that can lead to

reasoned participation by all students, thus increasing

the probability of productive peer interactions to

occur. Moreover, ColMOOC project (Caballé and

Conesa, 2019) constitutes a recent exertion towards a

new pedagogical paradigm that integrates MOOCs

with Conversational Agents and Learning Analytics,

according to the Conversation Theory (Pask, 1976).

Next paragraphs provide our framework to build

agents, (i) to give the right answers, (ii) based on

pedagogical models, (iii) also considering affective

factors. Our design is grounded on Deep Learning,

thus, generating responses through the imitation of

training datasets.

3.1 Goal-Oriented Question-Answering

(QA) Model

In supervised learning the algorithm given as input a

labeled dataset X tries to predict the output Y. For

example, in Sentiment Analysis the algorithm is

given as input a short text (e.g. review) and its goal is

to predict if it is negative or positive. Since, during

the training process the label is given (i.e. 0 for

negative and 1 for positive) the algorithm learns from

its mistakes and updates its weights in every iteration,

until it converges.

However, in RL the algorithm (i.e. AI agent) has

as input a set of data X, but the desired output is

defined as a reward r. In particular, the input is

considered to be the state of the agent s. To maximize

the reward function, the agent selects an action a from

the action space A. It should be noted, that there can

be two types of rewards, long-term and short-term

(instant). Let’s take for example, the Pac-Man game:

 The state of an agent (i.e. Pac-Man) is the

locations of the ghosts, its own location and the

existing dots in the maze.

 The action space is consisted of the moves the

agent can make (i.e. up, down, left, right).

 The instant rewards can be 1 for eating a dot, -

100 for been eaten by a ghost and 0 for doing

none of them. However, there can be a long-term

reward equal to 100 for completing the level.

 Finally, the strategy of the agent (i.e. which move

it should select given its current state) is called

policy, and is denoted with π.

In the current paper we advocate that RL

combined with Sentiment Analysis, can be used to

develop a pedagogical driven conversational agent. In

this case the whole task of the agent can be formulated

as follows:

 Goal: Help the student to complete the test

successfully without providing the direct answer

to him/her.

 State: The text input given by the student and

his/her current affective state.

 Action: The hint that is provided to the student in

text format.

 Reward: When the student will successfully

complete the task (e.g. if the hint will be helpful).

 Policy: What strategy the agent should select,

given the affective state of the student (e.g.

frustration) and the answer(s) he/she provided.

Therefore, the learning task is expressed in a goal-

oriented way, meaning that the agent tries to achieve

the student’s comprehension through dialog. Of

course, this is not the first goal-oriented approach in

a QA task (see Rajendran et al., 2018). The proposed

method includes two phases: (i) the agent tries to

learn how to perform dialog with the student and

detect his/her affective state given an annotated

dataset (supervised learning), and (ii) the agent learns

through trial and error –based on the incoming

rewards– to assist the student (reinforcement

learning) (Figure 1).

Figure 1: Goal-Oriented Question-Answering Agent.

In (Figure 1), the student sends a text answer in

the agent, which converts it into embeddings (i.e.,

numerical vectors). Next, the embeddings are

processed by an LSTM layer, which is both used to

encode the incoming data and recognizes the

students affective state (i.e., confusion). The outputs

are fused together using again a LSTM layer and

produce the output word embeddings; these are

mapped into actual words so that a helpful hint is

sent to the user. During the whole process the agent

receives a reward based on the student performance.

Building Pedagogical Conversational Agents, Affectively Correct

103

3.2 Pedagogical Supervision

Human reasoning constitutes a synergistic association

of ideas, elaborating a “non-stop” mining of reasonable

correlations between existing and input data, in a

continuous restructuring, or reforming of cognitive

schemas (assimilation-accommodation-adaptation,

known from the Piaget's Theory of Cognitive

Development (Piaget et al., 1969). Reforming

comprises two main operations: the convergent critical

and the divergent creative thinking.

Elaborate creative thinking on machines, seems

intangible, nevertheless, it is already in the AI agenda.

In (“What’s next for AI,” 2019) it is mentioned as the

“ultimate moonshot for artificial intelligence”

highlighting the fact that AI researchers should not

confine to machines that only think or learn, but also

create (Russell et al., 2010). The integration of

emotional and affective appraisals in computer

intelligence, constitutes a big step, since creativity

depend heavily on emotional weighted decisions

related to motivation, perversion, insight, intuition, etc.

Critical thinking sounds more tangible. For

instance, the IBM Watson (“IBM Watson,” 2019)

cognitive platform managed to analyse visuals, sound

and composition of hundreds of existing horror film

trailers and accomplished an AI-created movie trailer

for 20th Century Fox’s horror flick, Morgan (“What’s

next for AI,” 2019).

The vision and challenge as well, involves an agent

holding the capacity, not only to beat a chess

grandmaster (Silver et al., 2018), but also to guide the

learner in a collaborative mining of creative solutions,

out and beyond the respective problem. Q&As

methods – i.e. the Socratic method – are moving

towards that direction by breaking down a subject into

a series of questions, the answers to which gradually

distill the answer a person would seek.

In the exemplary work of Marzano (1988), 24

Core Cognitive Skills (CCS) have been classified

and grouped into four basic categories of data

processing, namely collection, organisation,

analysis and transcendence (Table 1). Based on this

work, Kordaki et al. (2007) proposed an architecture

for a Cognitive Skill-based Question Wizard.

It is in our intentions to build, evaluate and evolve

a framework that will employ both supervised and RL

approaches to simulate tutor-learner conversations.

The intelligent tutor will guide the learner to solve

learning tasks, evolving Cognitive Skills based on the

24 CCS. In Table 2 we provide examples of questions

that could be used to develop basic Cognitive Skills in

agents.

Table 1: Marzano 24 CCS.

Data

Collection

1. Observation

2. Recognition

3. Recall

Data

Organization

4. Comparison

5. Classification

6. Ordering

7. Hierarchy

Data Analysis

8. Analysis

9. Recognition of Relationships

10. Pattern Recognition

11. Separation of Facts from

Opinions

12. Clarification

Data

Transcendence

13. Explanation

14. Prediction

15. Forming Hypotheses

16. Conclusion

17. Validation

18. Error detection

19. Implementation-Improvement

20. Knowledge organization

21. Summary

22. Empathy

23. Assessment /Evaluation

24. Reflection

Table 2: Examples of question-models (adjusted from

Kordaki et al., 2007).

List of Basic Cognitive Skills

Examples of question-

models

Data

collection

Observation

A is a list of integers

A = [2, 1, 7, 0]

0 1 2 3



index

List index starts from 0

List increments by 1

What is the value of A[2]?

Recall

Remember A index starts

from 0.

Print (A[4])

Is this correct?

Data

organization

Comparison

A[0] > A[1]

Is this correct?

Classification

a is integer, b is real

c = a+b

c is real

Is this correct?

Data

analysis

Recognition

Relationships

a >b, b>c

a>c Is this correct?

Pattern

recognition

A = [2, 1, 7, 0], B = [3, 5,

C = [A, B]

C[1][0]=3 Is this correct?

Data

transcendence

Prediction

a=2x+y 1, if p<0.1

if x=2, y=

0, if p>=0.1

What is the value of a?

Forming

Hypotheses

If (it rains) and (I won’t

take umbrella)

then I get wet

CSEDU 2019 - 11th International Conference on Computer Supported Education

104

3.3 Affective Factor

As already mentioned (subsection 2.2), an Affective

Agent seeks for affective cues to appraise the

respondent’s sentiments. Next step involves the

agent’s decision to provide an answer affectively

correct. An affective response should be able to

change learning paths when i.e. boredom or

frustration is recognized (Feidakis, 2016).

Zhou et al. (2017) provide a dataset of 23,000

sentences collected from the Chinese blogging

service Weibo and manually annotated using 5 labels:

anger, disgust, happiness, like, and sadness. Such

datasets provide a starting point to enrich agent’s

short dialogs with emotion hues.

In previous work (Feidakis et al., 2014), a

Sentiment Analysis mechanism was implemented

classifying short posts in Web forums and Wikis,

according to 6 states (Figure 2), based on Machine

Learning and NLP.

Figure 2. Sentiments’ classifiers (Feidakis et al., 2014).

In this work, we need to extend our tasks to

integrate emotion or affective oriented lemmas

(inferences), following a machine learning approach.

In Table 3 we provide an example of agent’s

responses, while detecting first frustration and then

interest in user’s response:

Our contribution lies also on a new conceptual

model reflecting learners’ affective states in e-

learning context deploying both (i) dimensional, and

(ii) label representation models. In the former,

candidate dimensions already evaluated in learning

settings (Feidakis, 2016) are valence (positive-

negative), arousal (high-low) and duration (short-

long). The latter involves affective states that have

been classified in educational context such as

inspiration(high-five), excitement, frustration, anger,

stress (more emotive) or, engagement(interest),

boredom, fatigue, confusion (more affective). Both

approaches will be prototyped in a hybrid conceptual

model addressing emotive/affective states together

with their tentative transitions in time (Feidakis,

2016).

Table 3: An example of affective responses.

Agent

User

A = [2, 1, 7, 0]

B = [3, 5, 4]

C = [A, B]

C[1][0]=3

Is this correct?

Remember first index is 0.

I am not sure I got it

OK. Do you want to try

something else and come

back later?

I like that! Here is a hint for

you: Lists can take also other

lists as values.

What is the value of C[1]?

A is a list

[3, 5, 4]

Nice! And the [0] of A

Great! That’s the correct

answer! C[1][0]=3

4 CONCLUSIONS AND FUTURE

WORK

The proliferation of digital assistants, which at the

present is targeting the domains of productivity and

domotic automation (i.e Google home, Amazon’s

Alexa), has paved the ground for the introduction of

digital assistants in several other domains (e.g.

personal training). The education domain, presents

particularities and needs that arise from the need to

integrate the assessment, monitoring and exploitation

of cognitive and affective skills of the trainees. At the

same time, RL inherently supports the level of

interactivity required, so that the process of

integrating machine learning techniques in the

process of supporting the educational process is

performed in line with Sentiment Analysis and

Cognitive Skills assessment.

The current State-of-the-Art of Deep Learning

can provide simple models and exploit large amounts

of real data to train agents to act as valuable assistants,

customizing them to the needs of each case through

specialized datasets like bAbI (Weston et al., 2015)

or Emotional STC (ESTC) (Zhou et al., 2017). In this

paper, the above approach is proposed, towards the

provision of a framework for supporting pedagogical

driven conversational agents. The model of course

needs a lot of data towards its realistic use, since the

agent need to perform natural language understanding

and Sentiment Analysis, before responding to the user

based on the optimal policy (both state and action

Building Pedagogical Conversational Agents, Affectively Correct

105

spaces i.e., input and output text, are quite big, if we

consider how many word combinations exist).

Though the potential that the proposed approach

is quite high, leading to the introduction of highly

interactive tools of (pedagogically) added value, the

introduction of such a framework will introduce new

challenges, especially when it comes to the protection

of personal information. For this, both the

legal/regulatory and the technical frameworks which

will ensure the protection of personal and sensitive

information should be taken under serious

consideration, and implementations adopting privacy

by design approaches should be adopted. As for the

former, the recently applied GDPR in EU (679/2016)

provides the basis, over which compliance to ethical

and legal standards of any implementation can be

evaluated. As for the latter, the ability of increased

capabilities of terminal devices (Edge Computing –

Shi et al., 2016) is a promising candidate for limiting

the range over which collected data for applying the

machine learning techniques are used.

REFERENCES

Afzal, Shazia, Bikram Sengupta, Munira Syed, Nitesh

Chawla, G. Alex Ambrose, and Malolan Chetlur. “The

ABC of MOOCs: Affect and Its Inter-Play with

Behavior and Cognition.” In 2017 Seventh

International Conference on Affective Computing and

Intelligent Interaction (ACII), 279–84. San Antonio,

TX: IEEE, 2017. https://doi.org/10.1109/ACII.2017.

8273613.

Amazon Alexa [WWW Document], URL https://deve

loper.amazon.com/alexa (accessed 3.21.19).

Apple Siri [WWW Document], URL https://www.

apple.com/siri/ (accessed 3.21.19).

Benotti, Luciana, María Cecilia Martínez, and Fernando

Schapachnik. “Engaging High School Students Using

Chatbots.” In Proceedings of the 2014 Conference on

Innovation & Technology in Computer Science

Education - ITiCSE ’14, 63–68. Uppsala, Sweden:

ACM Press, 2014. https://doi.org/10.1145/2591708.

2591728.

Burleson, Winslow. “Affective Learning Companions:

Strategies for Empathetic Agents with Real-Time

Multimodal Affective Sensing to Foster Meta-

Cognitive and Meta-Affective Approaches to Learning,

Motivation, and Perseverance.” PhD Thesis,

Massachusetts Institute of Technology, 2006.

Caballé, Santi, and Jordi Conesa. “Conversational Agents

in Support for Collaborative Learning in MOOCs: An

Analytical Review.” In Advances in Intelligent

Networking and Collaborative Systems, edited by Fatos

Xhafa, Leonard Barolli, and Michal Greguš, 384–94.

Springer International Publishing, 2019.

Cambria, Erik. “Affective Computing and Sentiment

Analysis.” IEEE Intelligent Systems 31, no. 2 (March

2016): 102–7. https://doi.org/10.1109/MIS.2016.31.

Cambria, Erik, Bjorn Schuller, Yunqing Xia, and Catherine

Havasi. “New Avenues in Opinion Mining and

Sentiment Analysis.” IEEE Intelligent Systems 28, no.

2 (March 2013): 15–21. https://doi.org/10.1109/MIS.

2013.30.

Castellano, Ginevra, Ana Paiva, Arvid Kappas, Ruth

Aylett, Helen Hastie, Wolmet Barendregt, Fernando

Nabais, and Susan Bull. “Towards Empathic Virtual

and Robotic Tutors.” In Artificial Intelligence in

Education, edited by H. Chad Lane, Kalina Yacef, Jack

Mostow, and Philip Pavlik, 7926:733–36. Berlin,

Heidelberg: Springer Berlin Heidelberg, 2013.

https://doi.org/10.1007/978-3-642-39112-5_100.

D’Mello, Sidney K., Blair Lehman, and Art Graesser. “A

Motivationally Supportive Affect-Sensitive

AutoTutor.” In New Perspectives on Affect and

Learning Technologies, edited by Rafael A. Calvo and

Sidney K. D’Mello, 113–26. New York, NY: Springer

New York, 2011. https://doi.org/10.1007/978-1-4419-

9625-1_9.

Devlin, Jacob, Ming-Wei Chang, Kenton Lee, and Kristina

Toutanova. “BERT: Pre-Training of Deep Bidirectional

Transformers for Language Understanding.”

ArXiv:1810.04805 [Cs], October 10, 2018. http://arxiv.

org/abs/1810.04805.

Dialogflow [WWW Document], URL http://dialog

flow.com (accessed 3.21.19).

El Kadiri, Soumaya, Bernard Grabot, Klaus-Dieter Thoben,

Karl Hribernik, Christos Emmanouilidis, Gregor von

Cieminski, and Dimitris Kiritsis. “Current Trends on

ICT Technologies for Enterprise Information Systems.”

Computers in Industry 79 (June 2016): 14–33.

https://doi.org/10.1016/j.compind.2015.06.008.

Feidakis, M. “A Review of Emotion-Aware Systems for e-

Learning in Virtual Environments.” In Formative

Assessment, Learning Data Analytics and

Gamification, 217–42. Elsevier, 2016. https://doi.org/

10.1016/B978-0-12-803637-2.00011-7.

Feidakis, Michalis, Santi Caballé, Thanasis Daradoumis,

David Gañán Jiménez, and Jordi Conesa. “Providing

Emotion Awareness and Affective Feedback to

Virtualised Collaborative Learning Scenarios.”

International Journal of Continuing Engineering

Education and Life-Long Learning 24, no. 2 (2014):

141. https://doi.org/10.1504/IJCEELL.2014.060154.

Feldman, Ronen. “Techniques and Applications for

Sentiment Analysis.” Communications of the ACM 56,

no. 4 (April 1, 2013): 82. https://doi.org/10.1145/

2436256.2436274.

Google Assistant [WWW Document], 2019. URL

https://assistant.google.com/ (accessed 3.21.19).

Graesser, Arthur C. “Conversations with AutoTutor Help

Students Learn.” International Journal of Artificial

Intelligence in Education 26, no. 1 (March 2016): 124–

32. https://doi.org/10.1007/s40593-015-0086-4.

Hochreiter, Sepp, and Jürgen Schmidhuber. “Long Short-

Term Memory.” Neural Computation 9, no. 8

CSEDU 2019 - 11th International Conference on Computer Supported Education

106

(November 1997): 1735–80. https://doi.org/10.1162/

neco.1997.9.8.1735.

Hone, Kate, Lesley Axelrod, and Brijesh Parekh.

Development and Evaluation of an Empathic Tutoring

Agent, 2019.

Howard, Jeremy, and Sebastian Ruder. “Universal

Language Model Fine-Tuning for Text Classification.”

ArXiv:1801.06146 [Cs, Stat], January 18, 2018.

http://arxiv.org/abs/1801.06146.

IBM Watson [WWW Document], URL http://www.ibm.

com/watson (accessed 3.21.19).

Kordaki, Maria, Spyros Papadakis, and Thanasis

Hadzilacos. “Providing Tools for the Development of

Cognitive Skills in the Context of Learning Design-

Based e-Learning Environments.” In Proceedings of E-

Learn: World Conference on E-Learning in Corporate,

Government, Healthcare, and Higher Education 2007,

edited by Theo Bastiaens and Saul Carliner, 1642–

1649. Quebec City, Canada: Association for the

Advancement of Computing in Education (AACE),

2007. https://www.learntechlib.org/p/26585.

Mikolov, Tomas, Ilya Sutskever, Kai Chen, Greg Corrado,

and Jeffrey Dean. “Distributed Representations of

Words and Phrases and Their Compositionality.”

ArXiv:1310.4546 [Cs, Stat], October 16, 2013.

http://arxiv.org/abs/1310.4546.

Moridis, Christos N., and Anastasios A. Economides.

“Affective Learning: Empathetic Agents with

Emotional Facial and Tone of Voice Expressions.”

IEEE Transactions on Affective Computing 3, no. 3

(July 2012): 260–72. https://doi.org/10.1109/T-

AFFC.2012.6.

Paiva, Ana, Iolanda Leite, Hana Boukricha, and Ipke

Wachsmuth. “Empathy in Virtual Agents and Robots:

A Survey.” ACM Transactions on Interactive

Intelligent Systems 7, no. 3 (September 19, 2017): 1–

40. https://doi.org/10.1145/2912150.

Pask, Gordon. Conversation Theory: Applications in

Education and Epistemology. Amsterdam; New York:

Elsevier, 1976.

Piaget, Jean, Bärbel Inhelder, and Helen Weaver. The

Psychology of the Child. Nachdr. New York: Basic

Books, Inc, 1969.

Picard, Rosalind W. Affective Computing. Cambridge,

Mass: MIT Press, 1997.

Radziwill, Nicole M., and Morgan C. Benton. “Evaluating

Quality of Chatbots and Intelligent Conversational

Agents.” ArXiv:1704.04579 [Cs], April 15, 2017.

http://arxiv.org/abs/1704.04579.

Rajendran, Janarthanan, Jatin Ganhotra, Satinder Singh,

and Lazaros Polymenakos. “Learning End-to-End

Goal-Oriented Dialog with Multiple Answers.”

ArXiv:1808.09996 [Cs], August 24, 2018.

http://arxiv.org/abs/1808.09996.

Rumelhart, David E, James L McClelland, San Diego

University of California, and PDP Research Group.

Parallel Distributed Processing. Explorations in the

Microstructure of Cognition, 1987.

Russell, Stuart J., Peter Norvig, and Ernest Davis. Artificial

Intelligence: A Modern Approach. 3rd ed. Prentice Hall

Series in Artificial Intelligence. Upper Saddle River:

Prentice Hall, 2010.

Shi, Weisong, Jie Cao, Quan Zhang, Youhuizi Li, and

Lanyu Xu. “Edge Computing: Vision and Challenges.”

IEEE Internet of Things Journal 3, no. 5 (October

2016): 637–46. https://doi.org/10.1109/JIOT.2016.

2579198.

Silver, David, Thomas Hubert, Julian Schrittwieser, Ioannis

Antonoglou, Matthew Lai, Arthur Guez, Marc Lanctot,

et al. “A General Reinforcement Learning Algorithm

That Masters Chess, Shogi, and Go through Self-Play.”

Science 362, no. 6419 (December 7, 2018): 1140–44.

https://doi.org/10.1126/science.aar6404.

Stergios Tegos, and Stavros Demetriadis. “Conversational

Agents Improve Peer Learning through Building on

Prior Knowledge.” Journal of Educational Technology

& Society 20, no. 1 (2017): 99–111.

Tucker, Conrad S., Barton K. Pursel, and Anna Divinsky.

“Mining Student-Generated Textual Data in MOOCs

and Quantifying Their Effects on Student Performance

and Learning Outcomes.” Computers in Education

Journal 5, no. 4 (January 1, 2014): 84–95.

Vaswani, Ashish, Noam Shazeer, Niki Parmar, Jakob

Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz

Kaiser, and Illia Polosukhin. “Attention Is All You

Need.” ArXiv:1706.03762 [Cs], June 12, 2017.

http://arxiv.org/abs/1706.03762.

Vygotsky, L. S. The Collected Works of L. S. Vygotsky, Vol.

1: Problems of General Psychology. The Collected

Works of L. S. Vygotsky, Vol. 1: Problems of General

Psychology. New York, NY, US: Plenum Press, 1987.

Walker, M., 2019. Hype Cycle for Emerging Technologies

https://www.gartner.com/doc/3885468/hype-cycle-

emerging-technologies- (accessed 3.21.19).

Weston, Jason, Antoine Bordes, Sumit Chopra, Alexander

M. Rush, Bart van Merriënboer, Armand Joulin, and

Tomas Mikolov. “Towards AI-Complete Question

Answering: A Set of Prerequisite Toy Tasks.”

ArXiv:1502.05698 [Cs, Stat], February 19, 2015.

http://arxiv.org/abs/1502.05698.

What’s next for AI [WWW Document], n.d. URL

http://www.ibm.com/watson/advantage-reports/future-

of-artificial-intelligence/ai-creativity.html (accessed

3.21.19).

Zhou, Hao, Minlie Huang, Tianyang Zhang, Xiaoyan Zhu,

and Bing Liu. “Emotional Chatting Machine:

Emotional Conversation Generation with Internal and

External Memory.” ArXiv:1704.01074 [Cs], April 4,

2017. http://arxiv.org/abs/1704.01074.

Zuckerberg, M., 2016. Building Jarvis [WWW Document].

URL https://www.facebook.com/notes/mark-zucker

berg/building-jarvis/10103347273888091/?pnref=

story (accessed 3.21.19).

Building Pedagogical Conversational Agents, Affectively Correct

107