Topic management for an engaging conversational agent

https://doi.org/10.1016/j.ijhcs.2018.07.007Get rights and content

Highlights

  • A Topic Manager for a conversational virtual agent is proposed to favour the user’s engagement in human-agent interaction.

  • It manages what topic to introduce, when and how, based upon user engagement, the agent's mental state, and dialogue history.

  • The Topic Manager adapts the interaction topics on the fly to any user, without pre-entering information about the user.

  • Third-party human observers perceive the actions of the Topic Manager in the way they are modelled.

Abstract

We present an engagement-driven Topic Manager that enables a conversational agent to personalise the topics of interaction in human-agent information-giving chat. The Topic Selection component of this computational model decides what the agent should talk about and when. For this it takes into account the agent’s dynamically updated perception of the user’s engagement as well as the agent’s own mental state. The Topic Transition component of the Topic Manager computes how the agent should introduce the topics in the ongoing interaction without loosing the coherence of the interaction. We have implemented the Topic Manager in a virtual agent, endowing it with the ability to adapt the topics of the interaction on the fly to promote the user’s engagement. By means of an evaluation study we have found that third party observers perceive the actions of the Topic Manager in the agent’s behaviour.

Introduction

To keep a human-agent interaction going it is essential for the agent to engage the user. This is necessary for the agent to deliver its messages and/or to complete the objective of the interaction (Bickmore et al., 2010). Because the topic of the interaction can have an influence on the user’s engagement (Glas and Pelachaud, 2015c) we propose a Topic Manager (TM) that aims at enhancing the user’s engagement by adapting the topics of the interaction to the user’s preferences.

The TM gives the agent the ability to select dialogue topics that are adapted to the user’s engagement; it computes what the agent should talk about, when, and how the agent should introduce the topics in the interaction. By taking into account (1) the goal of wanting to promote the user’s engagement, (2) the agent’s perception of the user’s engagement, and (3) the agent’s own mental state including the agent’s preferences and associations, the agent is not merely a user-oriented system, but modelled as a human-like interaction participant.

Fig. 1 illustrates that the engagement of the user forms the input of the TM. User engagement could be deducted from the user’s verbal and non-verbal behaviour, such as backchannels (Rich et al., 2010), postures (Peters, 2005), and gaze (Peters, 2005). The detection of user engagement itself is however outside the scope of this paper. This paper focuses on the TM itself. We evaluate the TM by verifying the perception of the generated agent behaviour.

Most dialogue models fall into two broad types. Often conversational virtual agents employ a strict task-oriented dialogue structure where the order and selection of the topics of the interaction are predefined by the particular task for which the agent is built. Chat-based systems on the other hand, allow for less rigid interaction but the agent has less control over the topics of the interaction. We are, however, interested in dialogue that falls in between these categories: Where there is not a clear task to achieve and where the interaction is not completely open either, but where there is freedom of topic choice within a certain domain. Concretely, we propose topic management for information-giving chat where the agent aims at “transferring a maximum amount of information while the exact choice of information does not matter” (Glas et al., 2015). In this type of interaction the interaction does not need to be strictly task-driven locally but can be driven by social variables of the interaction, such as the user’s engagement. While the TM can be applied to multiple domains we will illustrate it for the information-giving chat for which we developed it originally: one-to-one face-to-face interaction between a human and a virtual agent in a museum where the agent tries to engage the user in order to transfer some amount of cultural information.2

In the following section we first discuss related work. We subsequently describe the TM’s components that are responsible for selecting what topics to talk about (Topic Selection Model, Section 4), when (Engagement Detection Model, Section 5) and how (Topic Transition Model, Section 6). We propose an implementation of the TM in Section 7 and an evaluation in Section 8. In Section 9 we conclude our findings.

Section snippets

Engagement and topic management

The term engagement in human-agent interaction has a wide range of definitions and uses (see Glas and Pelachaud, 2015a for an overview). These definitions reflect the focus of studies on a particular aspect or interpretation of engagement without the studies being able to cover the entire range of interpretations. We employ the definition of Poggi (2007) where engagement is defined as:

“The value that a participant in an interaction attributes to the goal of being together with the other

Overall architecture and dialogue example

Our work is embedded in a human-agent interaction system (Fig. 2) that allows a human to communicate with a virtual agent by verbal and nonverbal means. The agent makes also use of communicative signals. A camera and a microphone detect the acoustic and visual signals of the user that are interpreted in terms of verbal content, engagement level, and emotional states. This information is sent to the intent planner which contains several modules (e.g. an emotion model, dialogue manager, and Topic

What: Topic Selection Model

The Topic Selection component of the TM decides that whenever the agent needs to introduce a new topic in the interaction, which one to choose. This selection is performed by taking into account (1) the goal of wanting to promote the user’s engagement, (2) the agent’s perception of the user’s engagement, and (3) the agent’s own mental state including the agent’s preferences and associations. Before explaining the calculations that incorporate these concepts we first describe how these concepts

When: Engagement Detection

As mentioned above, the detected engagement of the user serves to decide when to change a topic (Appendix B, Eq. (5)), and to update variables that play a role in the selection of the topics (Appendix B, Eqs. (6) and (7)). In this section we describe how the TM calculates the values that are used for these procedures. We assume that the TM receives measurements of detected user engagement with a value between 0 and 1, accompanied by a certainty score (between 0 and 1) of the measurement. How

How: Topic Transition Model

In order to obtain human-like agent discourse, the (sub)topics that are selected by the topic selection component of the TM to engage the user have to be initiated by the agent in a natural and coherent way in the ongoing interaction. By coherence we mean that there is a form of connection that is explicit between successive topics of discussion. We assume that coherence in discussed topics makes the interaction appear more natural.

On a subtopic level this is straightforward; all subtopics

Implementation in Greta

To use the TM in a conversational virtual agent we integrated it in the agent platform Greta (Pecune et al., 2014). Greta enables the creation of ECA’s (Embodied Conversational Agents) by handling all the steps required to transform an agent’s intention to the rendering of the agent’s multimodal behaviour. It is SAIBA compatible (Heylen et al., 2008) and made of three main components (see Fig. 6):

  • Intent planner: It takes as input the utterances of the user and the level of observed user

Evaluation

We evaluate the Topic Manager in an application where a virtual agent plays a visitor in a museum who tries to pass on a maximum amount of cultural information to the user. The agent and the user are both visitors but the agent is a visitor that has already seen the museum while the user is discovering it. We already know that the user’s engagement is promoted if an agent talks about artworks that the user prefers (Section 2.1). What is left to verify is if the TM enhances how the agent manages

Conclusion

We have developed, implemented and evaluated a Topic Manager (TM) for a conversational agent that tries to engage the user while conveying information. The TM is developed to adapt the topics of the interaction on the fly by deciding what (sub)topics to talk about, when, and how to introduce the topics in the ongoing interaction.

The Topic Selection component of the TM takes into account the agent’s dynamically updated perception of the user’s engagement, the agent’s own preferences and its

Acknowledgements

We would like to thank Ken Prepin and Hervé Marie-Nelly for valuable comments, and Brice Donval, David Panou, and Charles Rich for their help in implementing the Topic Manager. Thank goes also to the anonymous volunteers who participated in the evaluation study.

This work was supported by the national project A1:1 and the European project H2020 Aria-Valuspa.

References (55)

  • T. Bickmore et al.

    Maintaining engagement in long-term interventions with relational agents

    Appl. Artif. Intell.

    (2010)
  • D. Bohus et al.

    Models for multiparty engagement in open-world dialog

    Proceedings of the Annual Meeting of the Special Interest Group on Discourse and Dialogue (SIGDIAL)

    (2009)
  • G. Brown et al.

    Discourse Analysis

    (1983)
  • J. Burgoon et al.

    Relational messages associated with nonverbal behaviors

    Hum. Commun. Res.

    (1984)
  • G. Castellano et al.

    Detecting engagement in hri: an exploration of social and task-based context

    Proceedings of the International Conference on Social Computing (SocialCom), Privacy, Security, Risk and Trust (PASSAT)

    (2012)
  • G. Castellano et al.

    Detecting user engagement with a robot companion using task and social interaction-based features

    Proceedings of the International Conference on Multimodal Interfaces

    (2009)
  • R.T. Clemen et al.

    Making hard decisions with decisiontools

    (2013)
  • M. De Jong et al.

    Politeness and alignment in dialogues with a virtual guide

    Proceedings of the International Joint Conference on Autonomous Agents and Multiagent systems (AAMAS)

    (2008)
  • D. Dellarosa

    The psychological appeal of connectionism

    Behav. Brain Sci.

    (1988)
  • M. Garber-Barron et al.

    Adaptive storytelling through user understanding

    Proceedings of the Artificial Intelligence and Interactive Digital Entertainment Conference

    (2013)
  • N. Glas et al.

    Definitions of engagement in human-agent interaction

    Proceedings of the ACII International Workshop on Engagement in Human Computer Interaction (ENHANCE)

    (2015)
  • N. Glas et al.

    Topic transition strategies for an information-giving agent

    Proceedings of the European Workshop on Natural Language Generation (ENLG)

    (2015)
  • N. Glas et al.

    User engagement and preferences in information-giving chat with virtual agents

    Proceedings of the IVA Workshop on Engagement in Social Intelligent Virtual Agents (ESIVA)

    (2015)
  • Glas, N., Prepin, K., Pelachaud, C., 2015. Engagement driven topic selection for an information-giving agent....
  • Gratch, J., Kang, S., 2015. Rapport Measures. In: Gratch Lab, p....
  • D. Heylen et al.

    The next step towards a function markup language

    Proceedings of the International Conference on Intelligent Virtual Agents (IVA)

    (2008)
  • A. Kaukiainen et al.

    Peer-estimated Social Intelligence (PESI)

    (1995)
  • Cited by (0)

    1

    Present address: CNRS - ISIR, Sorbonne Université, 4 Place Jussieu, 75005 Paris, France.

    View full text