November 24, 2003

Conversational game theory: the cartoon version

Deadlock: a funny exploration of why the logic of communication is hard, illustrating the thought processes of an interpersonally-sensitive Asperger's sufferer adult male human.

Like other actions, communicative choices have consequences. As shown in this strip, it's really hard to work out what choice leads to the best outcome a few moves down the road, especially when the other participants may not even be playing the same game.

This reminds me of a visit to SRI in the mid-70's, where I saw a demo of a principled conversational system. As I recall, my host typed in his conversational opening, and then we went to lunch while their KL10 churned away, trying to prove a theorem about what the optimal response to "hello" might be. I think we got back well before the machine had calculated its next move. This experience left me with a a completely illogical feeling that the machine, clueless and ungrounded as it was, still somehow really meant what it said, purely by virtue of the effort that it appeared to put into choosing its responses. But I also acquired another (more rational?) conviction: as a metaphor (or a system design) for on-line control of conversation, it would be better to pick a stochastic finite automaton instead of a theorem-prover. It might not do the right thing, but it would do something. A still more plausible conclusion, however, might be that no one had (has?) yet invented a formalism that does a good job of modeling human communication.

[via johnny logic].

Posted by Mark Liberman at November 24, 2003 02:44 PM