=Paper=
{{Paper
|id=Vol-2338/paper2
|storemode=property
|title=ADELE: Care and Companionship for Independent Aging
|pdfUrl=https://ceur-ws.org/Vol-2338/paper2.pdf
|volume=Vol-2338
|authors=Brendan Spillane,Emer Gilmartin,Christian Saam,Benjamin R. Cowan,Vincent Wade
|dblpUrl=https://dblp.org/rec/conf/atal/SpillaneGSCW18
}}
==ADELE: Care and Companionship for Independent Aging==
https://ceur-ws.org/Vol-2338/paper2.pdf
ADELE: Care and Companionship for Independent Aging
Brendan Spillane Emer Gilmartin
ADAPT Centre ADAPT Centre
Trinity College Dublin Trinity College Dublin
brendan.spillane@adaptcentre.ie gilmare@tcd.ie
Christian Saam Benjamin R. Cowan Vincent Wade
ADAPT Centre University College Dublin ADAPT Centre
Trinity College Dublin benjamin.cowan@ucd.ie Trinity College Dublin
saamc@cs.tcd.ie vwade@adaptcentre.ie
Abstract talk (Allen et al., 2000), and initial prototypes con-
centrated on practical tasks such as travel book-
Dialogue system technology offers inter- ings or logistics (Walker et al., 2001; Allen et al.,
esting prospects for services to seniors liv- 1995). Implementation of artificial task-based di-
ing independently. A spoken or text dialog alogues is facilitated by a number of factors. In
system can support services such as moni- these tasks, the lexical content of utterances drives
toring medication adherence, fall preven- successful completion of the task, conversation
tion and reporting, exercise and wellbe- length is governed by task-completion, and par-
ing coaching, entertainment, companion- ticipants are aware of the goals of the interac-
ship and the maintenance of social net- tion. Such dialogues have been modelled as fi-
works. Such systems would amalgamate nite state and later slot-based systems, first using
several different types of conversation hand-written rules and later depending on data-
or speech-exchange systems, from well- driven stochastic methods to decide the next ac-
defined tasks to engaging talk. Knowl- tion. Task-based systems have proven invaluable
edge of how these types of talk function in many practical domains. However, the cre-
is essential to system design. The ADELE ation of social companion application for health-
project aims to build an agent, ADELE, care or senior use entails the ability to wrap nec-
which can provide a range of services to essary tasks and recommendations in a matrix of
seniors. Below, we outline plans for the social conversation. Although the aim of such
system and describe challenges we antici- agents is often described as conversational social
pate in implementing the system. companions, the ability for dialog systems to chat
realistically lags their success in task based dia-
1 Introduction logue. Below, we describe a typical use case for
Spoken and text-based dialogue technology is the our proposed system, ADELE. ADELE will be ca-
focus of increasing interest in the domains of geri- pable of monitoring medication, providing well-
atric care and coaching. These domains present ness advice and positive motivation, monitoring
interesting challenges, as they rely not only on exercise and daily habits, storing and prompting
the formulaic instrumental exchanges used in task- general reminders, and engaging in companion-
based systems, but also on an ability to per- able, social dialogue. We then outline relevant cur-
form human-like casual and social talk. Instru- rent approaches in dialogue systems and identify
mental or task-based conversation is the medium key challenges for companionable social talk, and
for practical activities such as service encounters highlight the challenges of supporting social talk
(shops, doctor’s appointments), information trans- interaction. We outline early progress on ADELE,
fer (lectures), or planning and execution of busi- and describe future work.
ness (meetings). A large proportion of daily talk
2 ADELE Use Case
does not seem to contribute to a clear short-term
task, but builds and maintains social bonds, and The overall research goal of the ADELE project is
is described as ‘interactional’, social, or casual to explore the use of personalisation in improving
conversation. Early dialogue system researchers the efficacy of a digital companion that can com-
recognised the complexity of dealing with social municate through informal, yet informed social di-
18
alogue, on a variety of topics of interest to a user care domain. Below, we briefly overview some
over a prolonged time scale. A key point for the existing systems.
development of ADELE is that social spoken di-
alogue requires knowledge of the user to inform 3 Current Elder Care Systems and
topic, style, and timing of conversation. This will Applications
be achieved by adapting the system persona and Academic work on companion applications and
interaction based on the user’s interests and pro- agents for the elderly is well established, and sev-
file to manage how and when the interactions oc- eral commercial products have come to market.
cur, their content, and the means by which they are The work of Bickmore’s Relational Agents Group,
conveyed to the user to aid in comfortable, spoken which includes several applications of hybrid so-
delivery. The following is an example use case cial and task-based dialogue, is especially perti-
scenario between a future iteration of ADELE and nent to the ADELE project. Their Senior Exercise
Emma, a 77 year old woman. Agent uses dialogue to encourage users to do more
ADELE: ”Emma, the next episode of that med- exercise, with moderate success with health liter-
ical TV show you like should be on soon.” ate adults (Bickmore et al., 2013). Their Virtual
Emma: ”Oh great, I’ll check it out.” Nurse system takes patients through the transition
from in-patient to discharge and aftercare through
ADELE: ”In the meantime, you have time to do
a dialogue. Users liked the system more than hu-
your blood pressure daily check. I can then add
man doctors and nurses, citing the unrushed qual-
the results to your file.”
ity of interaction with the agent, and the possibility
Emma: ”Fine, let me just get the blood pressure of re-checking every step without embarrassment.
monitor.” The group have also created agents which explain
ADELE: ”Your son Mark will also be here to- health documentation and provide counselling on
morrow morning at 10 to collect you for your doc- a range of topics. Their early work on REA, a vir-
tors appointment.” tual estate agent which combined property view-
ing tasks with social talk, provided foundation re-
Emma: ”Oh great, I’d forgotten about that. It’s
search for hybrid task/social systems (Bickmore
his birthday next Thursday - can you remind me of
and Cassell, 2001). The Serroga system is an ex-
that?”
ample of a non-speaking robot companion for do-
ADELE: ”Ok, would you like me to remind you mestic health care assistance which assists senior
to give him a call on the day?” users in tasks from their day to day schedule and
Emma: ”Thanks ADELE, that would be great!” health care (Gross et al., 2015). The system em-
phasises social-emotional functions; in user trials
This interaction contains many elements which participants accepted the non-speaking robot as a
demonstrate what an everyday conversation en- real social companion or relational agent. This
tails, with topics flowing naturally. ADELE would was largely due to the successful establishment of
use the time between the reminder of the TV show co-presence. There is interesting recent work on
(a schedule reminder) and the actual start time to multimodal systems providing basic care for the
recommend to Emma to check her blood pressure, elderly and migrants (Wanner et al., 2016). Com-
which must be done daily. This, like other user mercial examples of social care robots include
recommendations and their responses, could be ElliQ, Jibo, and GeriJoy1 . The ADELE system
marked as critical, resulting in regular reminders will be based on dialog system and recommender
and/or notification of a third party. ADELE could system technology. Below, we briefly review dia-
also provide additional reminders such as a doc- log system design most relevant to our purposes
tor’s appointment. It could also reference events and discuss challenges to the creation of casual
such as Mark’s birthday, confirming with its cal- talk.
endar and asking if Emma wanted a reminder set.
The use case scenario above raises a number of re- 4 Spoken Dialogue Systems
search challenges, which will be explored during
the development of ADELE. There has been much Dialogue systems have predominantly focussed on
progress in dialog technology and there is a body practical tasks. Classic dialogue systems are usu-
1
of work on the use of such systems in the elder www.elliq.com, www.jibo.com, and www.gerijoy.com
19
ally based on a division into several modules that based on a range of models (Brusilovsky, 1998).
handle the different problems of natural language Recent work on personalisation and dialogue
dialogue (Jokinen and McTear, 2009). The Natu- agents includes a customer service agent (Verha-
ral Language Understanding component converts gen et al., 2014), a social robot tutor (Gordona
input into an internal representation that can be et al., 2015), and an eLearning agent (Peeters
reasoned on. The Dialogue Manager decides on et al., 2016). Dialogue management in digital
a next action and supplies the Natural Language agents has long adopted personalisation strategies.
Generation with a specification of the next output Recent contributions include Ultes et al, who ex-
which it converts to natural language. Dialogue tended dialogue management to adapt to user sat-
management was initially based on handwritten isfaction (Ultes et al., 2016), Litman and Pan,
rules, but this approach is severely limited in in- and San-Segundo et al. who developed meth-
teractions other than simple question/answer se- ods to adapt the overall dialogue strategy based
quences. More recently, research has concentrated on the performance of the speech recognizer (Lit-
on stochastic or machine learning methods, to bet- man and Pan, 2002; San-Segundo et al., 2005),
ter handle the uncertainty, noise, and variability in- and Nothdurft et al. and Gnjatović and Rösner
herent in spoken interaction. Stochastic dialogue have developed approaches to adaptive dialogue
systems have relied on the availability of large (Nothdurft et al., 2012; Gnjatović and Rösner,
quantities of relevant data. While several dialogue 2008). An early approach to combine a person-
corpora exist, these are generally collections of alised companion with adaptive dialogue was that
task-based dialogues, and not of casual talk (Ser- of André and Rist who developed personalised in-
ban et al., 2015a). Traditional approaches require formation assistants for accessing information on
this data to be labelled, adding significant cost to the web (André and Rist, 2002). More recently,
research as corpus annotation is a time and labour SARA was developed as a multifunctional con-
intensive undertaking. The ground truth provided versational agent capable of personalised recom-
by labelled data furnishes a good training signal mendations (Niculescu et al., 2014). To ensure a
to supervised learning for task-based interactions social and personalised interaction, the type and
which tend to be short and relatively predictable in source of content that each user is recommended
form and content. However, social talk can touch should be based on their history and content that
on virtually any topic and exhaustive annotation they have explicitly expressed an interest in dur-
of corpora that capture realistic variation is pro- ing speech interaction. This content may include
hibitively expensive. Therefore, approaches where news stories, conversational search terms or in-
labelling is not necessary are highly relevant. End- dividuals from social feeds (Garcin et al., 2012).
to-End systems can be trained directly from user Each user also has preferred sources for such con-
input to system output without additional levels of tent. These may also include preferred social con-
training data annotation. This advantage comes tacts such as close friends or family. A person-
at the price of requiring even greater amounts of alised intelligent companion should be capable of
training data. Sequence-to-Sequence (Sutskever accessing and recommending content in a similar
et al., 2014) models are a popular Neural Net- fashion. This is a complex process and includes
work architecture which can achieve this goal. topic extraction, weighting, mapping, longevity,
These models have proven effective in a variety of context etc. The nuances of people’s individual
tasks. Early systems based on this approach essen- interests and preferences are difficult to model and
tially translated from a user turn to a system turn require complex approaches to be accurately cap-
(Vinyals and Le, 2015). Some current versions use tured, if they are to allow personalised services to
an additional level of hierarchy that allows the sys- better meet the user’s needs. Efforts have been
tem to take into account longer histories by aggre- made to include this ability in agents. Garcin et
gating state information over previous turns (Ser- al. developed a personalised news recommenda-
ban et al., 2015b). tion system which demonstrated that collaborative
filtering provided the best results for recommen-
5 Challenges for Companionable Talk dations (Garcin et al., 2012). The content from
these favoured sources could be used to select in-
Personalised systems are concerned with adapt- formation to recommend when initiating or tak-
ing and personalising services to individual users
20
ing part in social talk. The time, location and ADELE’s user model needs to account for physi-
context of delivery of interruptions are important cal and demographic attributes, usage preferences,
in social agent interaction. A conversational so- context and temporal requirements, etc. It will
cial agent like ADELE needs to be able to insti- also need to model additional attributes such as
gate conversation which would involve interrup- personality, conversational preferences, and criti-
tion, but without excessively annoying or disturb- cal care needs, such as medication requirements,
ing the user. It is also important to ensure that each and long term care needs such as memory mon-
interruption or conversation has a point and is de- itoring. Many of these challenges and potential
livered in a concise manner. Likewise, it is impor- solutions, have been detailed by De Carolis et al.
tant that recommendations delivered within these (Carolis et al., 2013). Current chat-oriented sys-
interruptions are not overly repeated or become tems often exhibit a lack of variability in system
tiresome. The comparison by Bickmore et al. of output. One approach to counter this is the use of
strategies for interrupting the user to instigate ex- Reinforcement Learning to learn the production of
ercise or take medication is highly relevant to this utterances that are more beneficial to the long-term
research (Bickmore et al., 2008). There are sig- goal of the conversation (Li et al., 2016). A further
nificant challenges around the time at which a so- step is to drive this learning by adversarial training
cial agent should interrupt a user’s current task (to which seeks to make system output indistinguish-
initiate conversation) as well as how to interleave able from human-generated conversation (Li et al.,
with the user’s utterances to provide an effective 2017). Latent Variable models are another tech-
conversation. The urgency of the information or nique that can sample from a learnt stochastic vari-
request to be delivered by the agent should also able to introduce randomness (Serban et al., 2017).
impact the agent’s interruption strategy, especially An extension is to make this process controllable
in elderly care and social care contexts. Research by making the distribution conditional on a vari-
is also needed on personal adaptation of this in- able (Sohn et al., 2015). This can facilitate the ad-
terruption pattern to suit individual tasks and user justment of a dimension that needs to be adapted
preferences. There are also significant challenges such as friendliness. To work over longer contexts
to be faced in the personalisation and recom- such as those found in casual interpersonal conver-
mendation of content based on granularity, word sation which may lapse and restart over the course
choice, comprehensibility etc. Much of the rele- of hours or even days, memorization will need to
vant work is in personalised eLearning where the be addressed. Research into Memory Networks
delivery of content has received considerable at- has shown potential in the management of intra-
tention (Daradoumis et al., 2013). It is also impor- session context (Bordes et al., 2016), and may
tant to consider personality traits such as friend- prove applicable to the maintenance of longer con-
liness, chattiness and professionalism of ADELE, texts in ADELE. One of the goals of the ADELE
and the personality of the user. The importance system is to efficiently interleave different ‘sub-
of personality development for personal and vir- dialogues’ and sub-tasks - the system should be
tual agents has been documented previously (Doce able to break off from story-telling or casual chat
et al., 2010). Research has also highlighted the im- to perform a task such as medication checking and
portance of matching a system’s personality with then return to the previous activity. Both the chat
that of its users (Lee and Nass, 2003). There are and task elements will vary between users depend-
also several contributions on modelling person- ing on their care model and circumstances. A per-
ality traits, such as the framework of McQuig- tinent question is how to manage these different
gan et al. for modelling empathy (McQuiggan sub-dialogues to promptly intervene to perform
et al., 2008). Further work is required to identify subtasks? There has been some success with rein-
and understand how these traits may be perceived forcement learning in this context, which is being
in social agent dialogue, the delivery of person- explored by the ADELE project (Yu et al., 2017).
alised recommendations, and the extent to which
they need to be personalised to individuals. Per- 6 Current Focus
sonalised recommendation and dialogue. based
on user personality has recently been investigated Initially the ADELE project focused on investigat-
(Braunhofer et al., 2015; Vail and Boyer, 2014). ing the greeting and leavetaking phases of an in-
formal conversation. The project now aims to in-
21
vestigate how to generate the body of a friendly Marc Light, Nathaniel Martin, Bradford Miller,
conversation (both interactive chat and longer Massimo Poesio, and David R. Traum. 1995. The
TRAINS project: a case study in building a conver-
chunks). To this end the project is focussing on
sational planning agent. Journal of Experimental &
topic shift and shading, the mechanisms which Theoretical Artificial Intelligence 7, 1 (1995), 7–48.
underpin the development of such conversations
Elisabeth André and Thomas Rist. 2002. From adap-
(Ries, 2001; Lambrecht, 1996). Consequently, it
tive hypertext to personalized web companions.
will be necessary for ADELE to be able to iden- Commun. ACM 45, 5 (2002), 43—46.
tify, strategise, render, and initiate topic shift and
Timothy Bickmore and Justine Cassell. 2001. Rela-
topic shading in a conversation. There are several tional agents: a model and implementation of build-
reasons for this. Firstly, it will allow ADELE to ing user trust. In Proceedings of the SIGCHI confer-
change the course of a conversation. Secondly, it ence on Human factors in computing systems. ACM,
will enable ADELE to more easily and more nat- 396–403.
urally follow a conversation strategy, such as rec- Timothy Bickmore, Daniel Mauer, Francisco Cre-
ommending a television show. Thirdly, it will en- spo, and Thomas Brown. 2008. Negotiat-
able ADELE to form more natural dialogue. The ing Task Interruptions with Virtual Agents for
Health Behavior Change. International Foun-
ADELE project is currently identifying and anno-
dation for Autonomous Agents and Multiagent
tating examples of topic shift and topic shading in Systems, 1241—1244. http://dl.acm.org/
Switchboard (Godfrey et al., 1992), a corpus of citation.cfm?id=1402821.1402841
2,400 two-sided telephone conversations. A Wiz-
Timothy W. Bickmore, Rebecca A. Silliman, Kerrie
ard of Oz experiment based on a social care sce- Nelson, Debbie M. Cheng, Michael Winter, Lori
nario is being conducted to generate additional so- Henault, and Michael K. Paasche-Orlow. 2013. A
cial dialogues for training the Neural Network. Randomized Controlled Trial of an Automated Ex-
ercise Coach for Older Adults. Journal of the
American Geriatrics Society 61, 10 (Oct 2013),
7 Conclusion 1676—1683. DOI:http://dx.doi.org/10.
1111/jgs.12449
ADELE will be a virtual speech dialogue agent ca-
pable of informal, yet informed social dialogue. Antoine Bordes, Y.-Lan Boureau, and Jason We-
The importance of social dialogue to create social ston. 2016. Learning End-to-End Goal-Oriented
Dialog. arXiv:1605.07683 [cs] (May 2016).
bonds cannot be underestimated to support inde-
http://arxiv.org/abs/1605.07683
pendent living and companionship while provid- arXiv: 1605.07683.
ing a means for task reminders to promote bene-
Matthias Braunhofer, Mehdi Elahi, and Francesco
ficial self care. By treating the dialogue manage- Ricci. 2015. User Personality and the New User
ment as a personalisation task, the dialogue system Problem in a Context-Aware Point of Interest Rec-
can dynamically adapt to the users’ preferences to ommender System. Springer International Publish-
enhance a more socially beneficial and comfort- ing, 537549. DOI:http://dx.doi.org/10.
1007/978-3-319-14343-9_39
able conversational interaction. This will require
an increased focus on personalisation strategies. Peter Brusilovsky. 1998. Methods and techniques of
adaptive hypermedia. In Adaptive hypertext and hy-
Acknowledgements permedia. Springer, 1—43.
Berardina De Carolis, Irene Mazzotta, Nicole Novielli,
This research is supported by the Science Founda- and Sebastiano Pizzutilo. 2013. User Modeling in
tion Ireland (Grant 13/RC/2106) and the ADAPT Social Interaction with a Caring Agent. Springer,
Centre (www.adaptcentre.ie) at Trinity College, London, 89116. DOI:http://dx.doi.org/
Dublin. 10.1007/978-1-4471-4778-7_4
T. Daradoumis, R. Bassi, F. Xhafa, and S. Caballé.
2013. A Review on Massive E-Learning (MOOC)
References Design, Delivery and Assessment. 208—213. DOI:
http://dx.doi.org/10.1109/3PGCIC.
J. Allen, D. Byron, M. Dzikovska, G. Ferguson, L. 2013.37
Galescu, and A. Stent. 2000. An architecture for a
generic dialogue shell. Natural Language Engineer- Tiago Doce, Joäo Dias, Rui Prada, and Ana Paiva.
ing 6, 3&4 (2000), 213—228. 2010. Creating Individual Agents through Per-
sonality Traits. Springer Berlin Heidelberg,
James F. Allen, Lenhart K. Schubert, George Ferguson, 257—264. DOI:http://dx.doi.org/10.
Peter Heeman, Chung Hee Hwang, Tsuneaki Kato, 1007/978-3-642-15892-6_27 00041.
22
F. Garcin, K. Zhou, B. Faltings, and V. Schickel. 2012. Denmark, 2157–2169. https://www.aclweb.
Personalized News Recommendation Based on org/anthology/D17-1230
Collaborative Filtering. Vol. 1. 437—441. DOI:
http://dx.doi.org/10.1109/WI-IAT. Diane J. Litman and Shimei Pan. 2002. Designing
2012.95 and Evaluating an Adaptive Spoken Dialogue Sys-
tem. User Modeling and User-Adapted Interaction
Milan Gnjatović and Dietmar Rösner. 2008. Adap- 12, 2—3 (Jun 2002), 111—137. DOI:http://
tive Dialogue Management in the NIMITEK dx.doi.org/10.1023/A:1015036910358
Prototype System. Springer, Berlin, Heidel-
berg, 14—25. DOI:http://dx.doi.org/ Scott W. McQuiggan, Jennifer L. Robison, Robert
10.1007/978-3-540-69369-7_3 Phillips, and James C. Lester. 2008. Mod-
eling Parallel and Reactive Empathy in Virtual
John J Godfrey, Edward C Holliman, and Jane Mc- Agents: An Inductive Approach. International
Daniel. 1992. SWITCHBOARD: Telephone speech Foundation for Autonomous Agents and Multiagent
corpus for research and development. In Acoustics, Systems, 167—174. http://dl.acm.org/
Speech, and Signal Processing, 1992. ICASSP-92., citation.cfm?id=1402383.1402411
1992 IEEE International Conference on, Vol. 1.
IEEE, 517—520. Andreea I. Niculescu, Ridong Jiang, Seokhwan Kim,
Kheng Hui Yeo, Luis F. D’Haro, Arthur Niswar, and
Goren Gordona, Samuel Spauldinga, Jacqueline Kory Rafael E. Banchs. 2014. SARA: Singapore’s Auto-
Westlunda, Jin Joo Leea, Luke Plummera, mated Responsive Assistant, a multimodal dialogue
Marayna Martineza, Madhurima Dasa, and Cynthia system for touristic information. (2014).
Breazeala. 2015. Affective Personalization of a
Social Robot Tutor for Children’s Second Language Florian Nothdurft, Frank Honold, and Peter Kurzok.
Skills. (2015). 2012. Using Explanations for Runtime Dialogue
Adaptation. ACM, 63—64. DOI:http://dx.
H. M. Gross, S. Mueller, C. Schroeter, M. Volkhardt, doi.org/10.1145/2388676.2388694
A. Scheidig, K. Debes, K. Richter, and N. Doer-
ing. 2015. Robot companion for domestic health Marieke M.M. Peeters, Karel van den Bosch, John-
assistance: Implementation, test and case study Jules Ch. Meyer, and Mark A. Neerincx. 2016.
under everyday conditions in private apartments. Agent-Based Personalisation and User Modeling
5992—5999. DOI:http://dx.doi.org/10. for Personalised Educational Games. ACM,
1109/IROS.2015.7354230 303—304. DOI:http://dx.doi.org/10.
1145/2930238.2930273
Kristiina Jokinen and Michael McTear. 2009. Spoken
Dialogue Systems. Synthesis Lectures on Human Klaus Ries. 2001. Segmenting Conversations by
Language Technologies 2, 1 (2009). http: Topic, Initiative, and Style. Springer, Berlin, Hei-
//www.morganclaypool.com/doi/abs/ delberg, 51—66. DOI:http://dx.doi.org/
10.2200/S00204ED1V01Y200910HLT005 10.1007/3-540-45637-6_5
Knud Lambrecht. 1996. Information Structure and Rubén San-Segundo, Juan M. Montero, Javier Macı́as-
Sentence Form: Topic, Focus, and the Mental Guarasa, Javier Ferreiros, and José M. Pardo. 2005.
Representations of Discourse Referents. Cam- Knowledge-combining methodology for dialogue
bridge University Press. Google-Books-ID: bsXL- design in spoken language systems. International
CgAAQBAJ. Journal of Speech Technology 8, 1 (2005), 45—66.
Kwan Min Lee and Clifford Nass. 2003. Designing Iulian Serban, Alessandro Sordoni, Ryan
Social Presence of Social Actors in Human Com- Lowe, Laurent Charlin, Joelle Pineau, Aaron
puter Interaction. ACM, 289—296. DOI:http: Courville, and Yoshua Bengio. 2017. A Hi-
//dx.doi.org/10.1145/642611.642662 erarchical Latent Variable Encoder-Decoder
Model for Generating Dialogues. (2017).
Jiwei Li, Will Monroe, Alan Ritter, Dan Jurafsky, https://www.aaai.org/ocs/index.
Michel Galley, and Jianfeng Gao. 2016. Deep php/AAAI/AAAI17/paper/view/14567
Reinforcement Learning for Dialogue Generation.
In Proceedings of the 2016 Conference on Em- Iulian Vlad Serban, Ryan Lowe, Peter Henderson,
pirical Methods in Natural Language Processing. Laurent Charlin, and Joelle Pineau. 2015a. A
Association for Computational Linguistics, Austin, Survey of Available Corpora for Building Data-
Texas, 1192–1202. https://aclweb.org/ Driven Dialogue Systems. arXiv:1512.05742 [cs,
anthology/D16-1127 stat] (Dec. 2015). http://arxiv.org/abs/
1512.05742 arXiv: 1512.05742.
Jiwei Li, Will Monroe, Tianlin Shi, Sábastien Jean,
Alan Ritter, and Dan Jurafsky. 2017. Adversar- Iulian Vlad Serban, Alessandro Sordoni, Yoshua Ben-
ial Learning for Neural Dialogue Generation. In gio, Aaron C. Courville, and Joelle Pineau. 2015b.
Proceedings of the 2017 Conference on Empirical Hierarchical Neural Network Generative Models for
Methods in Natural Language Processing. Associ- Movie Dialogues. CoRR abs/1507.04808 (2015).
ation for Computational Linguistics, Copenhagen, http://arxiv.org/abs/1507.04808
23
Kihyuk Sohn, Honglak Lee, and Xinchen Yan. 2015.
Learning Structured Output Representation using
Deep Conditional Generative Models. In Ad-
vances in Neural Information Processing Systems
28, C. Cortes, N. D. Lawrence, D. D. Lee,
M. Sugiyama, and R. Garnett (Eds.). Curran Asso-
ciates, Inc., 3483–3491.
Ilya Sutskever, Oriol Vinyals, and Quoc V. Le. 2014.
Sequence to Sequence Learning with Neural Net-
works. CoRR abs/1409.3215 (2014). http://
arxiv.org/abs/1409.3215
Stefan Ultes, Hüseyin Dikme, and Wolfgang Minker.
2016. Dialogue Management for User-Centered
Adaptive Dialogue. Springer, Cham, 51—61.
DOI:http://dx.doi.org/10.1007/
978-3-319-21834-2_5
Alexandria Vail and Kristy Boyer. 2014. Adapting
to Personality Over Time: Examining the Effec-
tiveness of Dialogue Policy Progressions in Task-
Oriented Interaction. Association for Computa-
tional Linguistics, 4150. http://www.aclweb.
org/anthology/W14-4306
Tibert Verhagen, Jaap van Nes, Frans Feldberg, and
Willemijn van Dolen. 2014. Virtual Customer Ser-
vice Agents: Using Social Presence and Personal-
ization to Shape Online Service Encounters. Jour-
nal of Computer-Mediated Communication 19, 3
(Apr 2014), 529—545. DOI:http://dx.doi.
org/10.1111/jcc4.12066
Oriol Vinyals and Quoc V. Le. 2015. A Neural Con-
versational Model. CoRR abs/1506.05869 (2015).
http://arxiv.org/abs/1506.05869
Marilyn A. Walker, Rebecca Passonneau, and Julie E.
Boland. 2001. Quantitative and Qualitative Eval-
uation of Darpa Communicator Spoken Dialogue
Systems. In Proceedings of the 39th Annual Meet-
ing on Association for Computational Linguistics
(ACL ’01). Association for Computational Linguis-
tics, Stroudsburg, PA, USA, 515–522.
Leo Wanner, Josep Blat, Stamatia Dasiopoulou,
Mónica Domı́nguez, Gerard Llorach, Simon
Mille, Federico Sukno, Eleni Kamateri, Stefanos
Vrochidis, Ioannis Kompatsiaris, and others. 2016.
Towards a multimedia knowledge-based agent
with social competence and human interaction
capabilities. In Proceedings of the 1st International
Workshop on Multimedia Analysis and Retrieval for
Multimodal Interaction. ACM, 21–26.
Zhou Yu, Alexander Rudnicky, and Alan Black.
2017. Learning Conversational Systems that In-
terleave Task and Non-Task Content. (2017),
4214—4220. https://www.ijcai.org/
proceedings/2017/589
24