Probabilistic Semantics and Pragmatics - Approximation, Hyperboles, Quantity Implicatures, – ProbaSem

The interpretation of linguistic utterances depends on both their linguistically encoded meaning and on pragmatic inferences that derive from a reasoning about the speaker’s communicative intentions. Formal pragmatics, in the last decades, has involved two different approaches: a) formal semantics, and, b) decision-theoretic and probabilistic, Bayesian approaches to meaning, which tend to rely on different research strategies.

Using tools from formal logic, formal semanticists construct theories that predict certain qualitative generalizations, and have a transparent deductive structure, which makes it possible to falsify them by uncovering critical test-cases. Formal semantics and pragmatics have given rise to predictive and empirically fine-grained characterization of many phenomena, and to a rich and principled typology of semantic and pragmatic inferences. However, in this tradition, the probabilistic dimension of pragmatic inferences has been mostly ignored, and the field has few contact points with other theoretical branches of cognitive science.

Bayesian approaches, on the other hand, are able to ground accounts of linguistic inferences in independently motivated approaches to human inferences (Bayesian reasoning in particular), and can model the probabilistic nature of these inferences, in line with other developments in cognitive science. However, they tend to be less concerned with uncovering fine-grained generalizations than formal semantics/pragmatics approaches.

A major goal of this project is to bridge the gap between these two research philosophies. We will take advantage of the power and conceptually motivated methods of the new Bayesian approaches to pragmatics, while preserving and expanding what is particularly valuable in the culture and results of formal semantics/pragmatics.

One model of pragmatic meaning that has gained momentum and will play a central role in our proposal is the Rational Speech Act Model (RSA model, Goodman & Stühlmuller 2013). The model rests on an interactive theory of the inferences speaker and listener make about each other. Speakers choose their messages based on a model of the listener, with a specific goal: maximize the informativity of the message while minimizing its cost. When processing a message, listeners use Bayes rule to update their beliefs, using both their prior beliefs and the likelihood that the speaker would have chosen this message in various situations. While the RSA framework has been used to model many pragmatic phenomena, so far the models deal with relatively simple cases and do not engage with the full phenomenology of the phenomena under study. Moreover, they involve a lot of free parameters, which makes it hard to isolate clear falsifiable predictions.

The current project will focus on three main topics for which it is expected that integrating formal semantics approaches with Bayesian game-theoretic pragmatics will result in significant theoretical breakthroughs, namely: expressions of approximation (meaning and use of expressions like “about” and “around”), hyperboles (meaning and use of exaggeration as in “the shirt cost me a billion”), quantity implicatures (as in the strengthening of “some” to “some but not all”). In each case, the project includes both a theoretical and an experimental component. The project is guided by the following question: how does prior probabilistic knowledge influence interpretation and production?