Language of Thought Models for the Learning of Multiple Concepts: to which Degree is Pragmatic Reasoning Involved?

Abstract

This work explores the interaction between pragmatic reasoning and concept learning. Specifically, it examines how reasoning about a speaker’s intentions influences concept learning, as well as how learners’ beliefs about the meanings of novel words interact when those words are known to have distinct meanings.

The study comprises an experiment where participants learn two concepts simultaneously, involving the elicitation of conversational implicatures. The experimental results are then used to fit cognitive models. The model for pragmatics being used is the Lexical Uncertainty Model, which is based on the Rational Speech Act framework. This is a Bayesian model, where hypotheses about concept meanings need to be clearly defined in order to evaluate whether experimental observations (such as seeing trials’ feedback) support them. This is where Languages of Thought prove useful, as they account for concepts being realized through compositions of symbols, thereby allowing hypotheses about concept meanings to be represented as logical expressions.

The experimental results provide no clear indication that participants engage in conversational implicatures, whereas the cognitive models including pragmatic reasoning do not exhibit a significantly better fit with the experimental data. Nevertheless, these results should be considered preliminary, as the participant pool was very limited, and individuals capable of the required pragmatic reasoning for this task are uncommon. Furthermore, there is still a vast margin of improvement on the methodologies, given the complexity of the study.