Attention conservation notice: Log-rolling promotion of a paper written by two friends.
Some weeks ago, I directed noises of harrumphing approval towards Mark Liberman's contention that it is not enough to explain why humans evolved language, one must also explain why every other species failed to do so. Carl Bergstrom kindly brought to my attention a paper where he and Michael Lachmann do so at least partially: language allows us to lie.
Michael and Carl --- excuse me, Dr. Lachmann and Prof. Bergstrom --- consider a two-player signalling game, where one player, the sender, gets to observe the actual state of the world and send a signal about it to the receiver, whose payoff depends both on the actual state of the world and the action they chose to take. In equilibrium, the expected value of the information in the signal, averaging over all signals, is non-negative. However, for combinatorial signalling systems, but not monolithic ones, "the value of information conditional on a particular signal can be negative at equilibrium". Now, for any particular signal, you could patch this by applying a "monolithic" meaning to it, one not derived from the general combinatorial rules (as, e.g., learning "Of course I'll respect you in the morning" means "I don't respect you now"). But this doesn't really make the difficulty go away:
The problem is not that one can never assign monolithic meanings to phrases. It is simply that one does not encounter most phrases often enough to assign them monolithic meanings and as a result those phrases can be used in ways that confer negative value of information.... Combinatorial communication can efficiently facilitate large numbers of messages because novel messages can be interpreted simply from a familiarity with the message components. Unfortunately, this also means that receivers will assign meanings to messages without first-hand experience of the circumstances of their use — and thus certain messages can be consistently used to the detriment of signal receivers.Let me repeat that all of this holds at equilibrium, when neither the sender nor the receiver can do better by unilaterally switching to a different strategy. (Out of equilibrium, the value of information can be negative over-all, whether the signalling system is combinatorial or monolithic.) Of course this doesn't amount to a full model of the non-evolution of language, but I think it's a real insight. Now, if somebody would just come up with a good explanation for why the benefits of combinatorial communication outweighed this cost in our case, but not for cephalopods or lions or chimpanzees, I'd be happy...
Posted by crshalizi at July 28, 2005 11:55 | permanent link