Nikita Mattar

Small talk is more than chit-chat: exploiting structures of casual conversations for a virtual agent

An approach of improving the small talk capabilities of an existing virtual agent architecture is presented. Findings in virtual agent research revealed the need to pay attention to the sophisticated structures found in (human) casual conversations. In particular, existing dialogue act tag sets lack of tags adequately reflecting the subtle structures found in small talk. The approach presented here structures dialogues on two different levels. The micro level consists of meta information (speech functions) that dialogue acts can be tagged with. The macro level is concerned with ordering individual dialogue acts into sequences. The extended dialogue engine allows for a fine-grained selection of responses, enabling the agent to produce varied small talk sequences.

Let’s Get Personal

Agents that are able to build relationships with the people they are interacting with are envisioned to be more successful in long-term interactions. Small talk about impersonal topics has been found an adequate tool in human-agent interactions for manipulation of such relationships. We suspect that an agent and the interaction with it will be evaluated even more positively when the agent talks about personal information it remembers about its interlocutor from previous encounters. In this paper a model of person memory that provides virtual agents with information needed in social conversations is presented. An interaction study demonstrates the impact of personal information in human-agent conversations and validates the performance of our model.

An Interaction Game Framework for the Investigation of Human–Agent Cooperation

Success in human–agent interaction will to a large extent depend on the ability of the system to cooperate with humans over repeated tasks. It is not yet clear how cooperation between humans and virtual agents evolves and is interlinked with the attribution of qualities like trustworthiness or competence between the cooperation partners. To explore these questions, we present a new interaction game framework that is centered around a collaborative puzzle game and goes beyond commonly adopted scenarios like the Prisoner’s dilemma. First results are presented at the conference.

Strangers and Friends

We demonstrate how an artificial agents conversational style can be adapted to different interlocutors by using a model of Person Memory. While other approaches so far rely on adapting an agents be- havior according to one particular factor like personality or relationship, we show how to enable an agent to take diverse factors into account at once by exploiting social categories. This way, our agent is able to adapt its conversational style individually to reflect interpersonal relationships during conversation.

Adapting a virtual agent’s conversational behavior by social strategies

Interpersonal encounters are a complex phenomenon in human/human interaction. As social encounters with virtual agents become more important, these agents have to cope with problems of social perception, as well. To account for tasks concerned with the acquisition, utilization, and recall of social information, we earlier proposed to equip virtual agents with a Person Memory. In this paper we present how information available through a Person Memory enables the conversational agent Max to tackle different interpersonal situations.

Prototyping User Interfaces for Investigating the Role of Virtual Agents in Human-Machine Interaction

To investigate how different levels of embodiment or variations of a task affect the performance or perception of an interaction, researchers need tools that enable them to effortlessly modify such aspects. We demonstrate the MultiPro framework that allows for fast and easy prototyping of applications that are to be used in the context of human-machine interaction research. In conjunction with a newly developed cooperative game scenario we show how different configurations with and without an embodied agent can be configures and how the agent’s behavior can be adapted.