Grice's Strength

Speranza

In producing and comprehending language, speakers and listeners engage in rich inference processes that involve reasoning about the (often noisy) linguistic signal's literal meaning, world knowledge, and the context of utterance.

A fully-fledged theory of linguistic meaning requires formally describing and analyzing these inference processes, which has proven challenging because of the complexity of establishing which information is used and how it is integrated.

Rational theories of language use starting with Herbert Paul Grice in Oxford have attempted to bring some order to this difficult analytical situation by positing conversational principles that listeners expect cooperative speakers to abide by.

In particular, speakers are expected to include just enough -- but not too much -- information in their utterances for listeners to correctly infer their intended meaning from the literal meaning of their utterance, world knowledge, and the context of utterance.

But such theories have come under attack from psycholinguistics: speakers have been shown to produce both over- and underinformative utterances.

One may demonstrate that the assumption of (boundedly) rational linguistic agents can not only be upheld but even be explicitly modeled computationally to great explanatory effect.

Using a combination of behavioral lab-based and web-based experiments, corpus analyses, and computational modeling, I will show that speakers are best modeled as trading off the contextual informativeness and production cost of their utterances. Listeners in turn are best modeled as integrating their beliefs about such speakers with their prior beliefs about likely meanings via Bayesian inference. The modeled data come from two prima facie very different phenomena -- the production of overinformative referring expressions and the interpretation of underinformative scalar expressions.

I conclude that, rather than being wastefully overinformative, speakers systematically add information to the extent that doing so reduces their uncertainty about whether the listener will correctly infer their intended meaning.

Similarly, rather than being uselessly underinformative, speakers systematically provide less information because listeners can be relied on to make good use of the available contextual information. Together, these findings implicate a linguistic system geared towards communicative efficiency.