Dialogue Systems for Synthetic Characters

Language as used is wonderfully messy. We use metaphor, we allude, we sing and we swear.   As Wittgenstein discovered, FOPC is not going to tell us much about language in use. What we need is a model of Language as Action in a Social Setting.   Representation and information based models of language are okay as far as they go, but something more is needed.  Franco was an ECA (Embodied Conversational Agent) in a data cave for which we did some preliminary Wizard of Oz style experiments (more). The conclusion then was that we needed to know more about politeness.

Dialog at Sheffield

I was initially drawn to Sheffield to work with Yorick Wilks - someone I had admired since my PhD days and working with LDOCE.  I worked on the proposal for Companions (and it's unsuccessful predecessor Copain) and spend a few months working for Yorick before deciding that it was all too hard. I did however do some preliminary work with him as part of the Humaine NoE. Nigel the Cuttlefish is a chatbot that uses colour to express emotion.  He uses the Belief, Desire and Intention agent architecture to balance reactive behaviour (answering questions, greetings etc) with deliberative behaviour (planning to get compliments for example).  Nigel uses Information extraction techniques to understand the user's utterances, and templates to generate text - he is not that sophisticated, but using BDI as the dialogue manager means he is more believable than many chatbots. (see the discussion of believability here).


In 2005/6 some of us thought to look at why people abuse chatbots [Interact'04, CHI'06].   My take on it is that language in use is social interaction, and that politeness and abuse are two extremes in the maintenance of normative systems [Wallis'05].


Language has of course been studied for a long time by people outside AI :-) and Mark Hepple and I set up an EPSRC project to look at Conversation Analysis.   The conclusions from the CA4NLP project (EPSRC: Engineering Natural Language Interfaces: can CA help?) were as follows: These results are unpublished, but not from lack of trying.   I have also done some independent work re-evaluating the DARPA Communicator data from a CA perspective.
The scene in which the dog attacks the AIBO ( movie )

The key thing to come out of the CA-EM literature for me was however Paul Seedhouse's overview of lessons learned from CA, and his overview provides an explanation for why people swear at chatbots.   The blury picture to the right is taken from a movie of dog attacking an AIBO.   The critical point for me is that the dog warns the robot twice before throwing it across the room (the yelling is apparently the researchers worried about their AIBO - it is okay by the way if you were worried to).   Why does the dog warn, and how is the AIBO meant to know that it is being warned?   The hypothesis is that the dog is using a species specific hard-wired behavioural norm to socialize the AIBO.   If the AIBO had been a puppy, the puppy would be hard-wired to understand the significance of the bared teeth and the growl, and would learn to not interrupt a dominant dog's meal.  

Are there human equivalents of these species specific hard-wired behavioural norms?   Yep. Seedhouse summarises the last 50 years of CA research with the high level observation that an utterance in a conversation will either go seen but unnoticed - people answer your questions and greet when greeted, or noticed and accounted for - you can figure out why he or she said what they said, or risks sanction. When people abuse chatbots, the chatbot is being sanctioned. Like the Dog in the move, we humans are hard-wired to socialize the socially inept.

And now ...

The latest is the SERA project, which attempts to look at the very broad picture of social engagement of humans with synthetic characters, and give us a theoretical framework on which to pin politeness and roles, abuse and normative action...