ARTIFICIAL CONVERSATIONAL COMPANIONS

A Requirements Analysis

Sviatlana Danilava

, Stephan Busemann

and Christoph Schommer

Faculty of Science, Technology and Communication, University of Luxembourg

6 Rue Coudenhove-Kalergi, Luxembourg City, Luxembourg

Deutsches Forschungszentrum f¨ur K¨unstliche Intelligenz (DFKI) GmbH, Stuhlsatzenhausweg 3, Saarbr¨ucken, Germany

Keywords:

Artiﬁcial companions, Conversational agents, Human-machine relationship, Long-term interaction.

Abstract:

This work is based on several attempts to provide a deﬁnition and a design approach of Artiﬁcial Companions

that can be found in the referenced literature. We focus on computer agents that simulate human language

behaviour and are aimed to serve, to assist and to accompany their owner over a long period of time, that

we call Artiﬁcial Conversational Companions. Although accepted by the research community, the visions

set very high expectations of such agents, but they do not address the technical feasibility and the system

limitations. This is the ﬁrst approach to deﬁne a set of features that allow an artiﬁcial agent to be regarded

as an Artiﬁcial Conversational Companion. We describe relationships between the components and identify

systematic shortcomings of the current systems. We propose a scalable method for implementing the desired

capabilities of an Artiﬁcial Conversational Companion in a generic framework with reusable, customizable

and interdependent components.

1 INTRODUCTION

The term Artiﬁcial Companion (AC) has been intro-

duced in (Wilks, 2006) as “... an intelligent and help-

ful cognitive agent which appears to know its owner

and their habits, chats to them and diverts them, as-

sists them with simple tasks...” . The most important

characteristics of an AC are the absence of a central

task, a sustained discourse over a long time period,

a capability to serve interests of the user, and a lot

of personal knowledge about the main user (Wilks,

2010).

(Adam et al., 2010) deﬁne companions as “...

agents that are intelligent, and built to interact nat-

urally [...] with their user over a prolonged period

of time, personalising the interaction to them and de-

veloping a relationship with them.” In (St˚ahl et al.,

2009) an AC is “a computational agent that acts as a

conversational partner to its user, builds a long-term

relationship to the user, and learns about the user’s

needs and preferences.” (Webb et al., 2010) empha-

sise that “Companions are targeted as persistent, col-

laborative, conversational partners [which] can have

a range of tasks.”. (Pulman et al., 2010) see a con-

versation with an AC as “not necessarily connected

to any immediate task”. (Benyon and Mival, 2008)

describe an AC as a “... personalised conversational,

multimodal interface, one that knows its owner.” They

see a companionship as “...an accessible, pleasing re-

lationship with an interactive source in which there

has been placed a social and emotional investment

(Benyon and Mival, 2010).

Summarised, an AC is a personalised, multi-

modal, helpful, collaborative, conversational, lear-

ning, social, emotional, cognitive and persistent com-

puter agent that knows its owner, interacts with the

user over a long period of time and builds a relation-

ship to the user.

These visions of an AC raise the level of expec-

tations of such an agent quite high, but they do not

address the technical feasibility and the system limi-

tations. Requirements like “to know its owner”, “be

helpful” or “long-term relationship” are vague. These

requirements and their impacts must be clearly de-

ﬁned in order to implement an AC.

1.1 Previous Work on Companions

(Benyon and Mival, 2010) give an overview on pet

and anthropomorphic computer agents. All of them,

from Tamagochi to artiﬁcial woman, are referred to

as “companions”. The form of an AC inﬂuences all

the issues of interaction and possibilities for compan-

ionship (a cat needs only to be a cat, see also (Benyon

282

Danilava S., Busemann S. and Schommer C..

ARTIFICIAL CONVERSATIONAL COMPANIONS - A Requirements Analysis.

DOI: 10.5220/0003834702820289

In Proceedings of the 4th International Conference on Agents and Artiﬁcial Intelligence (ICAART-2012), pages 282-289

ISBN: 978-989-8425-96-6

 2012 SCITEPRESS (Science and Technology Publications, Lda.)

and Mival, 2010)). In this paper, we use the term Arti-

ﬁcial Conversational Companion (ACC) for compan-

ions that are aimed to simulate human language be-

haviour, in order to distinguish them from those, that

are not (e.g. artiﬁcial pet companions).

Recent contributionsin the domain of ACC are the

EU-funded Companions project

with the “How Was

Your Day” Companion (HWYD-Companion) (Pul-

man et al., 2010) talking about job-related topics,

the Senior Companion (SC) reminiscing about im-

ages (Wilks et al., 2011), and the Health and Fitness

Companion (HFC) planning daily exercises and di-

ets (Turunen et al., 2011). A pre-companion work

on Relational Agents (Bickmore, 2003) focuses on

social-emotional relationships between humans and

computer agents. The ALIZ-E project

concentrates

on robot companions for children in a hospital envi-

ronment (Baxter et al., 2011), and the Child Compan-

ion (Adam et al., 2010) is designed to engage a child

user with games and stories.

A large amount of research has been done up to

now on various questions related to ACCs. However,

the implementation of the distinguishing features that

are part of the above descriptions of an ACC (per-

sonalisation, sustained discourse, required knowledge

about the user and the learning mechanism), have not

been deeply analysed.

The research results available from the different

areas do not yet sufﬁce to fully support an ACC. We

will argue that not only domain-speciﬁcity but also

limited cognitive, social, and emotional competencies

will, in spite of the visions put forward, be a factor to

be dealt with in the foreseeable future.

1.2 Research Questions

(Benyon and Mival, 2008) introduce a general model

for designing technologies for relationships. The

model is based on ﬁve concepts: utility, form, emo-

tions, companion’s personality and trust as well as its

social attitudes. According to this model, long-term,

persistent interactions are part of the companion’s

personality and trust axis. However, modelling of

what the authors call “companion’s personality” can-

not be done independently from modelling its social

or emotional attitudes. The sets of the companion’s

capabilities shaped alongside these axes will overlap.

It will make the model more complex and lead to re-

dundancies.

Moreover, the model does not clearly deﬁne what

is meant by a “companion’s personality” and how it

http://www.companions-project.org/

http://www.aliz-e.org/

should be modelled. The authors mention assertive-

ness and submissiveness as properties of a person-

ality. However, we cannot observe the properties

attributed to a person, but only how the person or

agent interacts with another person or agent. This,

in turn, depends on how all the participants inﬂuence

the interaction. As will be discussed below, a per-

son interacts differently with different partners – even

within the same role distribution like, for instance, in

a teacher-student interaction. During long-term inter-

action, participants tend to adapt their behaviour to

the situation and the relationship expectations as well

as to the behaviour of the other participants. The use-

fulness of a notion of an ACC’s personality seems, in

our view, dubious.

This paper is the ﬁrst attempt to deﬁne a set of re-

quirements for an ACC that supports the practical im-

plementation. In this paper, we address the following

questions:

1. What is the set of mandatory requirements that

a computer agent must satisfy in order to be re-

garded as an ACC?

2. What is, in addition, required for a long-term

human-companion interaction?

3. Where are important technical limitations?

We take the view that

• mutual dependencies among the components

rather than a system of independent modules will

provide the desired functionality;

• the utility and the adaptivity of the system build a

basis for a long-term interaction;

• the design of an ACC should accommodate cer-

tain future research progress and support different

application cases.

We see a possibility to implement such an ACC in

a generic system of interdependent components that

can be accustomed.

The remainder of this paper is organised as fol-

lows: in Section 2 we analyse the requirements and

their implementation in the current companion pro-

totypes. We also point out shortcomings and re-

search gaps that may currently make the requirements

unattainable. In Section 3 we propose design princi-

ples for an ACC framework that takes research gaps

and technical limitations into consideration. This is

followed by conclusions in Section 4.

2 REQUIREMENTS

Which interactional resources are involvedin the con-

versation depends on the modalities involved. For an

ARTIFICIAL CONVERSATIONAL COMPANIONS - A Requirements Analysis

283

ACC interacting via instant messages, it will be only

text, possibly extended by emoticons etc. For a talk-

ing head avatar it will be speech, prosody, eye gaze,

facial expressions and head movements. For an em-

bodied agent, the body languageand body movements

must be added.

The entire interactional phenomena of a talk and

the connections among them related to each single

participant of an interaction are referred to as interac-

tional proﬁle (Spranz-Fogasy, 2002). The content, the

actions, the ﬂow of emotions and the relationships are

determined and co-constructed by all the participants

in each interaction and especially in each conversa-

tion. We discuss these aspects of the implementation

of a companion’s interactional proﬁles below.

2.1 Conversational abilities

Conversation is interactive, spontaneous, exchange of

ideas between two or more agents that follow rules

of etiquette and politeness, according to social dis-

tance and cultural norms. Thus, the conversational

part of an ACC is responsible for the understanding

and production of spontaneous utterances during the

interaction with the user and following rules of social

interaction using the available modalities.

Companions are also aimed at maintaining a sus-

tained discourse over a long period of time. This re-

quires an analysis of the conversational history and

different cognitive functions to work on it (e.g., asso-

ciative memory, learning or reasoning). The interac-

tion with an ACC cannot be modelled as just a simple

stimulus-response based exchange of utterances.

2.1.1 Language Understanding and Generation

In Natural Language Understanding (NLU), we have

the tradeoff between deep language understanding

reaching at very elaborate interpretations of utter-

ances at the cost of covering only a restricted do-

main and shallow language understanding that is un-

restricted with respect to the domain but inherently

limited in its understanding capabilities – very simple

techniques as keyword spotting or pattern matching

systematically ignores information available in the

text. Work on combining both has been carried out:

deep linguistic analysis can be enriched with shallow

techniques such as named entity recognition (NER)

(Sch¨afer, 2008). Current ACC prototypes use differ-

ent shallow techniques for NLU, depending on the

system application case. Each of them acts in a single

topic domain.

The main objective of the HWYD-Companion

was producing longer utterances that are still appro-

priate in terms of content and emotions. The HWYD-

Companion performs template-based information ex-

traction – it uses shallow syntactic and semantic pro-

cessing – to ﬁnd instantiations of event templates.

The dialogue manager questions the user until enough

slots are ﬁlled. Then a longer empathic response is

generated.

The HFC uses semantic interpretation for speech

recognition and domain speciﬁc grammars. “The

Cognitive Manager models the domain [...]. In con-

trast, the Dialogue Manager focuses on interaction

level phenomena, such as conﬁrmations, turn taking,

and initiative management.” (Turunen et al., 2011)

Conversational key features of the SC are read-

ing news from a few categories, telling jokes taken

from the Internet, and voice-based picture tagging.

The NLU module of the SC is based on GATE (Cun-

ningham et al., 1996). The components have been

improved for the SC system by gazetteers containing

locations and family relationships. The NER mod-

ule builds the key part, which is required for this ap-

plication scenario. The information obtained is then

passed to the Dialogue Manager and stored in the

knowledge base for later reference.

Understanding the goals underlying the user’s ut-

terances enables a system to decide when to produce

the next utterance (turn-taking), what content to con-

vey next, and how to express that content. Complex

interactions between dialogue manager, planner and

action selection are needed to meet interaction goals

like fulﬁlling some task or just “killing time”.

In classical dialogue applications, templates with

a ﬁxed number of slots deﬁne howthe dialogue should

be maintained. Dialogue managers for a free, inter-

active conversation are mostly improved by enlarg-

ing the state space, which leads to combinatorial ex-

plosions in planning tasks, as the dialogue develops.

This complexity can be managed to a certain extent

by policy activation (Kruijff and Lison, 2010). But

eventually, all these techniques have the disadvantage

that developers need to specify in advance the condi-

tions under which the system produces a certain de-

cision. Similarly, the subsequent system utterance is

selected from a set of all probably appropriate utter-

ances. This leads to a perceived repetitiveness of the

system, as it has been demonstrated in the evaluation

of Laura (Bickmore et al., 2005). Most of the par-

ticipants found the conversations repetitive at some

point. Annoyance and negative feelings reduced the

motivation of several participants.

All these systems act in a restricted domain in or-

der to make the NLU task manageable. However,

there is no guarantee for the system to always remain

within its domain. For instance, in a human-to-human

conversation about images, comments like “You look

ICAART 2012 - International Conference on Agents and Artificial Intelligence

284

so beautiful in this picture!” or “Good shot! This

dress suits you very well!” are also probable. More-

over, domains are usually not well-deﬁned. For in-

stance, “job-related topics” may include topics on any

job, including taxi driver, teacher, programmer, and

not only manager.

For an ACC, techniques are required that

smoothly open emergency exit doors whenever the

system’s coverage is left. Better ways than the in-

famous “Tell me more about your family!” of the

ELIZA system are easily available, but depend on

available cognitive abilities and social relationships.

The straightforward solution – if appropriate – would

be to let the user know what the system did not under-

stand.

2.1.2 Cognitive Abilities

With the vision of an ACC described in Section 1, an

ACC needs cognitive functions that allow it to have

personal knowledge about its main user, to learn new

facts about the user and user’s world, his needs and

preferences, and to use this knowledge in the interac-

tion as well. However, cognitive abilities of an ACC

do not necessarily have to coverall the cognitivefunc-

tions of humans. It is a subject for the future research

to identify the functions required for a personalised

long-term interaction, based on the particular appli-

cation case and desired functionality.

For example, the SC introduced in Section 1.1 per-

forms a real-time information search about places to

show that it has some knowledge of what is being

talked about while talking about images, in contrast

to HWYD-Companion and HFC, that do not have and

probably do not need a real time access to additional

knowledge and are well served with the facts they

have been initialised with.

SC asks the user about the names and the age of

the persons on images, and about the places where

the pictures were taken. The gathered information is

used to maintain the dialogue which is still focused

on image tagging. Simple inferences about family re-

lationships it already knows makes the dialogue more

natural.

The HWYD-Companion needs for each subtopic

a particular amount of information that is extracted

from user utterances. Then a longer response of the

system is generated.

Two sample interactions show how knowledge

about the user in combination with simple inferences

improves the quality of the conversation. The ﬁrst one

is by SC: SC: Where was this photo taken?

U: This photo was taken in Tanzania.

SC: Have you ever been to Zanzibar?

U: Yes, I have.

SC: Great place, I hope you saw the beaches. ...

Then the system goes to the next picture.

These simple inferences (Tanzania, Zanzibar,

beaches) make the dialogue more ﬂuent and natural.

In the HWYD-Companion’s video demonstration,

the conversation starts usually with Hello John. How

did your work day go today?, and the user tells the

ACC, how it was.

In the demonstration, the user says that he arrived

late because of the trafﬁc. The ACC replies:“You have

my sympathy. What happened next?” A more appro-

priate reaction would be to ask the user whether he

managed it to be on time in the meeting, produced by

the inference from the knowledge about the user (the

plan to have a meeting) and the context (user arrived

late because of the trafﬁc). The given system reac-

tion is emotionally adapted to the context (sympathy)

and can be applied in each situation where sympathy

would be appropriate reaction.

The existing systems have a predeﬁned informa-

tion need and just need to get those data from the

user that are declared as information need by the pro-

grammer. Systems cannot decide, whether or not they

need more information as long as there is no delib-

erate planning associated with the knowledge items

mandatory to carry out the plan.

While convincing examples are demonstrated in

today’s ACCs, it is completely unclear how the cogni-

tive abilities can be extended to keep the conversation

interesting for the user in many-facetted long-term in-

teraction. This requires an amount of inferencing and

learning that is not required for the usual task-oriented

dialogue systems.

2.1.3 Emotional Competence

The ability to address the emotional side of compan-

ionship may play a key part in their acceptance by the

users (Cowie, 2010).

A big progress in affective computing was

achieved by the HUMAINE network

. Two general

types of emotions are studied: pervasive (general per-

sonal attitude colouring long time periods) vs. emer-

gent (short, intensive affectivestates) emotions. In the

follow-up project SEMAINE

the focus was on non-

verbal emotional behaviour and different agent char-

acters (happy, angry or despondent) (Schr¨oder et al.,

2011).

Most of this work is not directly applicable to

ACCs. Insights on emotion recognition are manda-

tory but must be complemented by appropriate reac-

tions. Insights on signalling emotions are important,

http://emotion-research.net/

http://www.semaine-project.eu/

ARTIFICIAL CONVERSATIONAL COMPANIONS - A Requirements Analysis

285

but need in an ACC trigger methods requiring inter-

nal stimuli. Studies of emotional effects in human-

computer dialogue are, to our knowledge, not avail-

able yet.

Emotion handling in the HWYD-Companion is

implemented in form of two feedback loops. The

“short loop” provides an immediate backchannel that

aligns the companion’s response to the user’s attitude

showing, e.g. empathy. The “major loop” is responsi-

ble for emphatic utterance generation, typically ad-

vice or warning expressed in both verbal and non-

verbal behaviour, based on the gathered information.

The SC’s emotional behaviour is based on speech

recognition. Recognised emotions were mapped onto

a two-dimensional space. SC should be able to recog-

nise user’s emotion placed in this space, formulate a

belief about user’s emotional state and move itself in

this space for an appropriate response.

Both systems focus only on short and intensive af-

fective states. The recognition of emotions is mainly

based on the speech prosody, which is not available in

a text based chat. Pervasive emotions need also to be

taken into account when designing tools for long-term

interaction.

Issues similar to the gathering knowledge about

user raise also in emotion handling: the systems can

only recognise predeﬁned emotions in particular pre-

deﬁned states, and produce an emotional response in

a predeﬁned way. The complexity of emotion han-

dling increases, if issues like “different perception of

the same event being in different moods” are consid-

ered.

2.1.4 Socio-cultural Competence

We communicate as social and cultural entities. There

are different sets of rules for successful conversation

within different groups of people as this was notices

many decades ago (Knigge, 1805).

The existence of such rules is usually not observ-

able until two legal but contradictory rules are applied

by participants of an interaction and lead to a con-

ﬂict situation and misunderstanding, see for example

(Young, 2011; Tannen, 2001).

Currently neither socio-linguistic nor computa-

tional models of such interactional rule systems are

available, but there are research efforts on socio-

linguistic phenomena in discourse (Strzalkowski

et al., 2010; Agar, 1996; Scollon and Scollon, 2000).

Using small-talk as a form of social dialogue in con-

versational agents helps to establish a bond between

the user and the system (Bickmore et al., 2005). Fur-

ther research investigations in socio-linguistic phe-

nomena and social signal processing will allow to im-

prove the conversation with an ACC.

The selection and the use of the interactional re-

sources is subject of research in second language ac-

quisition domain. Recent work shows that there is -

besides the four common competences: reading, lis-

tening, speaking and writing comprehensions - a ﬁfth

competence, the interactional competence, also re-

ferred to as intercultural or transversal (Cook, 1992;

Hall et al., 2006). Interactional competence is not

ﬁxed, it can be developed in the process of the in-

teraction (Young, 2011).

Each communication has a content part and a re-

lationship part, in which the latter determines the for-

mer (Watzlawick et al., 2000). Social and cultural

norms and rules are expressed in socio-linguistic phe-

nomena during a conversation. Given this fact, we can

eliminate neither the relationship part nor the socio-

cultural part from our communication.

Particular implementations of interpersonal rela-

tionships between people are in each case unique, but

they are categorised in large classes like “friends” or

“colleagues”. The relationship between an ACC and

its main user will be also belong to one of the large

classes that will probably differ from all interpersonal

human relationship classes.

Prior to modelling relationship-related speech

acts, we need to decide what kind of relationship

we want to establish between the user and the ACC.

A teacher-learner relationship will necessarily differ

from e.g. a personal assistant-boss relationship. The

general requirement that an ACC must produce a re-

lational response in its user is vague.

2.2 Adaptivity

(Reeves and Nass, 1996) report that users prefer com-

puters that become more like them over time over

those which maintain a consistent level of similar-

ity, and that users prefer computers that are similar to

them. This property is already used in ALIZ-E, where

the robot adapts its behaviour to the user’s behaviour

(Baxter et al., 2011).

Several research results show that people adapt

their language by selecting the same words and gram-

mar constructions while interacting with other peo-

ple, but also while interacting with machines, see e.g.

(Dobroth et al., 1990). This process is also referred

to as convergence and denotes negotiation on vocabu-

lary and communication style among all conversation

participants. In this way, artiﬁcial agents should be

able to adapt their language behaviour to that of their

users, who in turn might be inﬂuenced by the linguis-

tic behaviour of the ACC.

Since the interaction is co-constructed, interac-

tional proﬁles (introduced at the beginning of this

ICAART 2012 - International Conference on Agents and Artificial Intelligence

286

section) need to be modelled as a cooperation of the

ACC and the user according to a stereotype based user

model in the initial implementation. A large amount

of personal knowledge is necessary to create a highly

adaptive user model of one particular user. The adap-

tivity mechanism will allow to use the information

from the conversations with the user and to adjust the

user model.

Besides language behaviour, also cognitive, emo-

tional, and social behaviour is subject to adaptation.

While systems that can both interpret and generate

such behaviour should in principle be able to adapt

themselves to their interlocutor, this requires a way

of analysing the (non-)linguistic behaviour and learn-

ing why it occurred. A simple mimicking strategy

such as using a certain word as frequently as the di-

alogue partner is bound to fail. We are not aware

of relevant studies of credible adaptivity in long-term

human-machine interaction, but from ALIZ-E, such

results may be expected.

2.3 Utility

Tools should be useful. Conversations with an AC re-

quire time investment at the expense of time the user

could have spent with her family or friends. The ser-

vices of an ACC compete against services of other

machines. There must be a reason why a user decides

to use an ACC for searching Internet for news instead

of a web browser.

The utility of a companion’s services can be taken

as a measure of relative satisfaction, which is in this

case the frequency of consumed companions services

or the cumulative length of the conversations. Exper-

iments with elderly people described in (Benyon and

Mival, 2010) show that the kind of services an ACC

can perform for its user is considered very important.

Desired tasks for a robot companion range from mak-

ing the tea to doing the ironing.

In contrast to the vision of Wilks, most current AC

prototypes have a central task. Bickmore’s Laura was

aimed to be a ﬁtness coach, the HWYD-Companion

is created to talk about job-related topics, and the SC

is designed for reminiscing about images. Not all

the user interfaces to computer programs can be re-

placed by an ACC or a voice control. Correspond-

ing tests have been made with speech-to-text (STT)

technology and people who were willing to buy an

STT computer, as described in (Savoia, 2011). After

a few hours of tests participants changed their mind

to the worse. People’s throat would get sore, it cre-

ated a noisy work environment, and it was not suitable

for conﬁdential material. However, in scenarios like

reading or writing messages, or changing navigation

options while driving a car, voice control may well be

desirable.

The utility of the system provides a basis for a

long-term interaction. Application scenarios need to

be elaborated, where an ACC may successfully com-

pete against other devices and conventional computer

programs. An ACC could be helpful in applications

where the user beneﬁts directly from the conversation,

such as conversation training in a foreign language, or

where a kind of long-term goal exists as in coaching,

teaching, or psychotherapy, when a task-oriented dia-

logue system will not perform well.

2.4 Long-term Interaction

Long-term interaction is characterised by the user’s

continuous motivation to interact with the ACC. Such

interaction is inﬂuenced by emotions, trust, sympa-

thy, positive emotional bond and/or utility. To keep

the user motivated, the interaction must be interest-

ing and stimulating, which can be supported by oc-

casional unexpected utterances or non-linguistic be-

haviour. Another, more technical requirement is the

ACC’s need to ensure the consistency and the persis-

tence of the mass data acquired over time in the inter-

actions, and its appropriate usage in, e.g., determining

what has been communicated before.

Long-term interactions cannot be enforced. It is

based on a continuous interest of the user to commu-

nicate. Consider an interaction on a ticket counter,

which involves a simple work ﬂow and few dialogue

steps to achieve the goal. Taking a few minutes of in-

teraction at most, this is clearly not a suitable ACC ap-

plication. The requirements differ widely with a per-

son who wants to learn a foreign language and inter-

acts with a teacher in a class at a language school. The

teacher needs to keep the person engaged, to respond

to personal needs of the learner, and to establish a pos-

itive emotional relationship with the learner in order

to keep the learner motivated. The teacher needs to

acquire much personal knowledge about each learner.

While phone ticket selling has been ﬁelded using

state-of-the-art dialogue systems, more complex ca-

pabilities as described above are necessary for the

teaching scenario.

The duration of long-term interaction has always

been left open. Certainly it is multi-session interac-

tion, and it extends over several days up to weeks,

months, or years. In practice, real-time behaviour of

the system will become an issue as mass data accu-

mulate over time, which need to be accessed and in-

terpreted by the ACC. As experiments over extended

periods of time have to our knowledge never taken

place so far, it is unclear whether a notion of forget-

ARTIFICIAL CONVERSATIONAL COMPANIONS - A Requirements Analysis

287

ting and shaded accessibility of information is manda-

tory in order to efﬁciently deal with the data. Results

on the functioning of the human brain from further

research areas – psycho-linguistics and cognitive sci-

ence – may be of help here.

3 TOWARDS DESIGNING AN

ACC SYSTEM

A huge amount of research effort has been invested in

different disciplines that are relevant for developing

an ACC. Existing ACC systems are not easily com-

parable due to different design approaches, system

architectures and components used. As proposing a

full-ﬂedged new ACC architecture would go beyond

the scope of this position paper, our notion of “com-

ponent” is quite vague. Components may correspond

to key functionality required for each of the subsec-

tions of Section 2. Each subsection may represent

multiple components.

Designing an ACC based on the requirements

identiﬁed above should take the limitations of cur-

rent systems and the research gaps sketched in the

previous sections into consideration. With the cur-

rent knowledge future changes of ACC system de-

signs and architectures are more than likely. To min-

imise the frequency with which ACCs must be (re-)-

built from scratch, we propose a highly modular and

customisable approach that will allow components to

be deﬁned and reused. Basically this design involves

a highly declarative approach to knowledge represen-

tation, and – as a consequence – the separation be-

tween interpreter componentsand the knowledge they

interpret. This has become a standard approach in

NLU, where grammars are kept separate from parsers

or generators interpreting them. This way, the knowl-

edge and the interpreters can be modiﬁed indepen-

dently of each other, allowing e.g. to use the same

parser for a different language.

The same approach can be applied to modelling

mutual dependencies of component behaviour, which

– as argued in Section 2.1 – are an integral part of

an ACC. First attempts to implement mutual depen-

dencies among the components of an ACC include

the NLU and the Cognitive Module in HFC, emotions

and language generation in HWYD-Companion, and

past history, simple inferences and dialogue manage-

ment in SC. These systems can already partially in-

tegrate the learned information about the user, emo-

tional analysis output, or simple cognitive functions

into the conversation.

We propose to generalise on this by virtue of

a framework embedding components, component

knowledge, and knowledge about interdependency,

through which all interdependent behaviour is mod-

elled in a declarative way. In such a framework, the

developercan determine most easily the impact of any

change of one element onto any others.

As the conversational competence of ACCs will

– in the foreseeable future – be limited to a certain

domain and task at a time, particular components

will have different complexity for different applica-

tion cases. For example, the capabilities necessary

for a pleasant conversation will be accustomed for a

particular service domain. Customisation strongly de-

pends on the modalities of interaction available. For

instance, if emotions must be expressed through text,

the NLU components must be able to convey appro-

priate signs. The components will be customised in

such a way that they cover the domain and task in

hand. We expect that the general architecture by and

large remains untouched by customisation, but the

knowledge required will very much differ across ap-

plication cases.

Building various ACCs within this same design

paradigm will allow us to compare the particular com-

ponents of the ACCs and the resulting systems in a

uniﬁed way.

4 CONCLUSIONS

In this paper for the ﬁrst time a set of features and re-

quirements was proposed that allow an artiﬁcial agent

to be regarded as an Artiﬁcial Conversational Com-

panion. We argue that a multitude of different fea-

tures must be considered that taken together allow for

a useful long-term interaction between an ACC and

its user. Since research gaps and technical limitations

will prevent the realisation of Wilks’ visionary ideas

in the near future, we proposed design principles for

the implementation of an ACC framework that is par-

ticularly well-suited to extensions in view of scien-

tiﬁc and technical progress. Moreover these princi-

ples support the customization of the components to

dedicated tasks and domains. Mutual dependencies

among the components will be explicitly modeled to

provide the desired functionality.

ACKNOWLEDGEMENTS

We would like to thank Prof. Dr. Gudrun Ziegler

(University of Luxembourg) for valuable discussions

and constructive criticism of an earlier version of this

work.

ICAART 2012 - International Conference on Agents and Artificial Intelligence

288

REFERENCES

Adam, C., Cavedon, L., and Padgham, L. (2010). ”Hello

Emily, how are you today?”: personalised dialogue in

a toy to engage children. In Proc. of the 2010 Work-

shop on CDS, pages 19–24. ACL.

Agar, M. H. (1996). Language Shock: Understanding The

Culture Of Conversation. Harper Paperbacks.

Baxter, P., Belpaeme, T., Canamero, L., Cosi, P., Demiris,

Y., and Enescu, V. (2011). Long-Term Human-Robot

Interaction with Young Users. In IEEE/ACM Human-

Robot Interaction 2011 Conf.

Benyon, D. and Mival, O. (2008). Landscaping personiﬁca-

tion technologies: from interactions to relationships.

In CHI EA ’08, pages 3657–3662.

Benyon, D. and Mival, O. (2010). From human-computer

interactions to human-companion relationships. In

Proc. of the IITM ’10, pages 1–9.

Bickmore, T. W. (2003). Relational Agents: Effect-

ing Change through Human-Computer Relationships.

PhD thesis, Massachusetts Institute of Technology.

Bickmore, T. W., Rosalind, and Picard, W. (2005). Estab-

lishing and maintaining long-term human-computer

relationships. ACM Transactions on Computer Hu-

man Interaction, 12:293–327.

Cook, V. (1992). Evidence for multi-competence. Lan-

guage Learning, 42(4):557 – 591.

Cowie, R. (2010). Companionship is an emotional busi-

ness. In Close Engagements With Artiﬁcial Compan-

ions: Key Social, Psychological, Ethical and Design

Issues. John Benjamins Publishing Company.

Cunningham, H., Wilks, Y., and Gaizauskas., R. (1996).

GATE - a general architecture for text engineering. In

Proceedings of COLING-96. ACL.

Dobroth, K., Karis, D., and Zeigler, B. (1990). The de-

sign of conversationally capable automated systems.

In Proceedings of the 13th Int. Symp. on Human Fac-

tors in Telecommunications, pages 389–396.

Hall, J. K., Cheng, A., and Carlson, M. T. (2006). Recon-

ceptualizing multicompetence as a theory of language

knowledge. Applied Linguistics, 27(2):220–240.

Knigge, A. F. (1805). Practical philosophy of social life:

or, The art of conversing with men: after the German

of Baron Knigge. Penniman & Bliss, O. Penniman,

printers, Troy.

Kruijff, G. J. and Lison, P. (2010). Policy activation

for open-ended dialogue management. In Proc. of

the AAAI 2010 Fall Symp. “Dialogue with Robots”.

AAAI.

Pulman, S. G., Boye, J., Cavazza, M., Smith, C., and de la

C´amara, R. S. (2010). “How was your day?”. In Proc.

of the 2010 Workshop on CDS, pages 37–42. ACL.

Reeves, B. and Nass, C. (1996). The Media Equation: How

People Treat Computers, Television, and New Media

Like Real People and Places (CSLI - Lecture Notes).

Cambridge University Press.

Savoia, A. (2011). Pretotype it. http://

pretotyping.blogspot.com/.

Sch¨afer, U. (2008). Shallow, deep and hybrid process-

ing with UIMA and Heart of Gold. In Proc. of the

LREC-2008 Workshop Towards Enhanced Interoper-

ability for Large HLT Systems: UIMA for NLP, 6th Int.

Conf. on Language Resources and Evaluation, pages

43–50.

Schr¨oder, M., Bevacqua, E., Cowie, R., Eyben, F., Gunes,

H., Heylen, D., ter Maat, M., McKeown, G., Pammi,

S., Pantic, M., Pelachaud, C., Schuller, B., de Sevin,

E., Valstar, M., and W¨ollmer, M. (2011). Building

autonomous sensitive artiﬁcial listeners. IEEE Trans-

actions on Affective Computing, 99(1).

Scollon, R. and Scollon, S. W. (2000). Intercultural Com-

munication: A Discourse Approach (Language in So-

ciety). Wiley-Blackwell.

Spranz-Fogasy, T. (2002). Interaktionsproﬁle: Die Heraus-

bildung individueller Handlungstypik in Gespr¨achen.

Radolfzell: Verlag f¨ur Gespr¨achsforschung.

St˚ahl, O., Gamb¨ack, B., Turunen, M., and Hakulinen,

J. (2009). A mobile health and ﬁtness companion

demonstrator. In EACL ’09, pages 65–68.

Strzalkowski, T., Broadwell, G. A., Stromer-Galley, J.,

Shaikh, S., Taylor, S., and Webb, N. (2010). Mod-

eling socio-cultural phenomena in discourse. In Proc.

of the 23rd Int. Conf. on Computational Linguistics,

pages 1038–1046. ACL.

Tannen, D. (2001). You Just Don’t Understand: Women and

Men in Conversation. William Morrow Paperbacks.

Turunen, M., Hakulinen, J., St˚ahl, O., Gamb¨ack, B.,

Hansen, P., Rodr´ıguez Gancedo, M., de la C´amara, R.,

Smith, C., Charlton, D., and Cavazza, M. (2011). Mul-

timodal and mobile conversational Health and Fitness

Companions. Comput. Speech Lang., 25:192–209.

Watzlawick, P., Beavin, J. H., and Jackson, D. D.

(2000). Menschliche Kommunikation. Formen, Strun-

gen, Paradoxien. Huber, Bern.

Webb, N., Benyon, D., Hansen, P., and Mival, O. (2010).

Evaluating human-machine conversation for appropri-

ateness. In Proceedings of LREC2010.

Wilks, Y. (2006). Artiﬁcial companions as a new kind of in-

terface to the future internet. Technical report, Oxford

Internet Institute.

Wilks, Y. (2010). Is a companion a distinctive kind of rela-

tionship with a machine? In Proc. of the 2010 Work-

shop on Companionable Dialogue Systems, pages 13

– 18. ACL.

Wilks, Y., Catizone, R., Worgan, S., Dingli, A., Moore, R.,

Field, D., and Cheng, W. (2011). A prototype for a

conversational companion for reminiscing about im-

ages. Computer Speech & Language, 25(2):140 – 157.

Young, R. F. (2011). Interactional competence in language

learning, teaching, and testing. In Hinkel, E., editor,

Handbook of research in second language teaching

and learning, volume 2, pages 426–443. London &

New York: Routledge.

ARTIFICIAL CONVERSATIONAL COMPANIONS - A Requirements Analysis

289