Entailment Semantics Can Be Extracted from an Ideal Language Model

TL;DR

This paper demonstrates that entailment semantics can be extracted from an ideal language model trained on pragmatically grounded data, revealing how semantic information can be decoded from language models.

Contribution

It proves that entailment judgments can be derived from an ideal language model trained on Gricean data, bridging the gap between language modeling and semantic inference.

Findings

Entailment judgments can be extracted from an ideal language model.

Decoding entailment is possible from models trained on pragmatically grounded data.

The results suggest a framework for understanding semantics in language models.

Abstract

Language models are often trained on text alone, without additional grounding. There is debate as to how much of natural language semantics can be inferred from such a procedure. We prove that entailment judgments between sentences can be extracted from an ideal language model that has perfectly learned its target distribution, assuming the training sentences are generated by Gricean agents, i.e., agents who follow fundamental principles of communication from the linguistic theory of pragmatics. We also show entailment judgments can be decoded from the predictions of a language model trained on such Gricean data. Our results reveal a pathway for understanding the semantic information encoded in unlabeled linguistic data and a potential framework for extracting semantics from language models.