How is turn-taking handled in a virtual agent conversation? – Tech Episode 4

If you thought taking part in a group conversation was daunting, wait till you hear how difficult it is for our virtual agents to engage in a group conversation. Group conversation is a prominent form of human communication in order to make decisions or get ideas through exchanging information with each other. In these conversations, we see various aspects of group dynamics, i.e., the action, process, and changes that occur within the group.

So what actually simulates group conversational dynamics? It is difficult enough to bring two agents in one room but then, how do they actually collaborate?

If you thought taking part in a group conversation was daunting, wait till you hear how difficult it is for our virtual agents to engage in a group conversation. Group conversation is a prominent form of human communication in order to make decisions or get ideas through exchanging information with each other. In these conversations, we see various aspects of group dynamic, i.e., the action, process, and changes that occur within the group.

So what actually simulates group conversational dynamics? It is difficult enough to bring two agents in one room but then, how do they actually collaborate? 

As the goal of Council of Coaches is to create a council of virtual coaches, with various domain expertise and individual goals, it is important for our agents to be able to handle the differences in their individual goals and to overcome these differences to decide and achieve the group goal. Several existing works in literature have associated group performance, team satisfaction, and adherence to group cohesion. So, on one hand, we decided to build our multiparty conversational model based on a higher level group phenomenon i.e., group cohesion.  It describes the tendency of group members to share a bond, the attraction to stay together and the willingness to work together. It is prominent when the main goal of the group is decision making or problem solving. Further, literature on cohesion detection has shown a strong correlation with turn taking behaviours. Therefore, we will also be focusing on including turn-taking behaviours with the emphasis on non-verbal behaviours, i.e., gaze and feedback. With this type of a model, it should be easier for our agents to decide on which goal to pursue and how to present varying information in a cohesive manner.

On the other hand, our focus will be on how to present differences in opinions between the coaches. As all coaches will have different backgrounds, areas of expertise, and personalities, conflicts of opinion will naturally arise. How do our coaches navigate these conflicts, and how do we involve the user? Would it be helpful to have one of the coaches mediate between the other coaches?

To answer these kinds of questions, the focus of this part of our research will be on group discussion and the impact of different ways to present differences of opinion between the coaches. Stay tuned…

4 thoughts on “How is turn-taking handled in a virtual agent conversation? – Tech Episode 4”

  1. The problem reminds me of the question how to handle turn-taking in a conversation. Is there a scheduler that assings turns or do the agents manage this themselves? But what is themselves if we are talking about software agents that are designed? So on second thought the difference between the two might not be that big as it seems on first sight?

    1. Hello Rieks,

      Thank you for the interesting questions. Currently we have a scheduler that assigns turns to agents after we have determined what agents can make dialogue moves, and which moves those are. You are correct in saying that the difference is less large than one might think. Even if the agents did turn-management themselves, this would still come down to a central component through which the agents would have to negotiate with each other regarding who gets the turn. An interesting difference in presentation does come up if this negotation would be acted out in front of the user. With several agents trying to take a turn simultaneously, we could show some agents waiting their turn, some agents trying to grab the talking turn away from another agent, and agents interrupting each other.

  2. Hi Gerwin
    I wrote a text on turn-taking in Dutch on my web site rieksopdenakker.nl
    Good luck with your thesis.

Leave a Reply to Rieks Cancel reply

Your email address will not be published. Required fields are marked *