Session "Conversational Competence"

• Participants: David, Jana, Kerstin, Kris, Kristinn
• Notes taken by Kerstin

What is conversational competence (CC)?

We first made a distinction between functional CC and human-like CC, yet then accumulated a broad set of phenomena that suggest that CC has to be human-like in order to be functional:

• users have subconscious expectations that they carry into interactions
  • one of these expectations concerns the fact that humans develop their competences in a particular order, in general from simple to complex; systems, in contrast, can be so designed as to be good at isolated capabilities that do not rest upon more basic behaviors;
  • if the system uses artificiality as a cue, it may remind people that they cannot apply their expectations from human-human interaction;
• matching these expectations reduces users’ cognitive load;
• in case of uncertainty, people fall back on what they know, which is conversations among humans

We also identified one area in which human-likeness cannot be the goal: while human disfluencies result from certain cognitive and social processes (and are ambiguous between them), system utterances are generated in very different ways that do not naturally result in disfluencies if problems occur. This can be a problem in those instances in which a disfluency is expected for politeness reasons, as in A: ‘can you help me move on Saturday?’ B: ‘no’ – where the expected human response would be disfluent and delayed.

We then named some competences that contribute to CC:

• turn-taking
• understanding the task
• situation awareness
• repairs
  • verification
  • clarification
• picking the right style depending on topic, partner, situation
• adjusting behavior to shared goal

We also discussed the relationship between CC and predictability and found

• CC means being predictable
• conversational incompetence is even worse if it is not predictable
  • for example, if you know that your car navigation system is insensitive to the current driving situation, it is easier for you to filter out its messages when you need to concentrate on the traffic
• while personas may be a good way to make a system predictable to the user, it can backfire if it is not congruent with the task

Edit this page -- TimoBaumann -- 06 Oct 2012