Learning about Word Meaning from Pragmatically Enriched Data

Anonymous

Abstract: The meaning of a natural language utterance can vary greatly depending on the context of the communication. An artificial agent interpreting natural language needs to be able to integrate models of the human speaker and the communicative goal in order to arrive at the correct interpretation. This paper introduces an approach integrating pragmatic reasoning about the conversational partner while learning representations from scratch. This leads to significant improvements over prior work that only considers pragmatics during inference or builds on fixed representations of literal meaning. Our artificial language learner is situated in a referential game about images, where we show that equipping the agent with explicit reasoning about the speaker and the shared observations, leads to faster learning, higher communicative success, and better generalization to changes in the environment.

Paper Type: long

Research Area: Discourse and Pragmatics

Contribution Types: NLP engineering experiment, Theory

Languages Studied: English