Entailment Semantics Can Be Extracted from an Ideal Language Model

Proceedings of the 26th Conference on Computational Natural Language Learning (CoNLL), pages 176 - 193 December 7-8, 2022 ©2022 Association for Computational Linguistics Entailment Semantics Can Be Extracted from an Ideal Language Model William Merrill New York University Alex Warstadt ETH Zürich {willm,linzen}@nyu.edu alexanderscott.warstadt@inf.ethz.ch Tal Linzen New York University Abstract Language models are often trained on text alone, without additional grounding. There is debate as to how much of natural language se- mantics can be inferred from such a procedure. We prove that entailment judgments between sentences can be extracted from an ideal lan- guage 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 commu- nication from the linguistic theory of pragmat- ics. 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 seman- tic information encoded in unlabeled linguistic data and a potential framework for extracting semantics from language models. 1 Introduction Recent advances in building computational models of language have been powered by distributional semantics: the idea that a text span’s surrounding context encodes its meaning (Firth, 1957). In par- ticular, large pretrained language models (LMs; Peters et al., 2018; Devlin et al., 2019; Brown et al., 2020) have become an integral part of NLP sys- tems: the representations that emerge from training to predict missing words in a text are empirically useful for natural language understanding tasks. Despite this empirical progress, Bender and Koller (2020) argue LMs cannot learn to under- stand the semantics of sentences. This is because of a mismatch between the LM training objective— predicting missing words in text (“form”)—and Bender and Koller’s conception of meaning as the relation of a sentence to the external world. Thus Bender and Koller claim “that the language model- ing task, because it only uses form as training data, cannot in principle lead to learning of meaning.” In this paper, we argue meaning can be learned from form because the communicative goals of hu- man authors encode semantic information in unla- beled text. We show how this semantic information can be extracted to resolve semantic relations be- tween sentences (e.g., whether one sentence entails another): in this inferentialist sense, ideal LMs en- code the meaning of sentences. This argument has been raised speculatively by others (Michael, 2020; Potts, 2020; Bommasani et al., 2021), but we will rigorously justify it here with formal results. To give the simplest (and least general) illustra- tion of our argument, we first assume training data is generated by overly idealized uniformly truthful speakers: agents who decide what to say by picking sentences they consider true uniformly at random.1 This very coarsely captures human authors’ goal of being informative (rather than misleading) to their listeners (Grice, 1975). In Theorem 1, we prove a sentence x entails sentence y if and only if, after uttering x, a uniformly truthful speaker is just …

Content truncated to 500 words