A fine-grained comparison of pragmatic language understanding in humans and language models

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 4194–4213 July 9-14, 2023 ©2023 Association for Computational Linguistics A fine-grained comparison of pragmatic language understanding in humans and language models Jennifer Hu1, Sammy Floyd1,2, Olessia Jouravlev3, Evelina Fedorenko1,4, Edward Gibson1 1Department of Brain and Cognitive Sciences, Massachusetts Institute of Technology 2Department of Psychology, Sarah Lawrence College 3Department of Cognitive Science, Carleton University 4McGovern Institute for Brain Research, Massachusetts Institute of Technology {jennhu,samfloyd,evelina9,egibson}@mit.edu olessiajouravlev@cunet.carleton.ca Abstract Pragmatics and non-literal language under- standing are essential to human communica- tion, and present a long-standing challenge for artificial language models. We perform a fine- grained comparison of language models and humans on seven pragmatic phenomena, us- ing zero-shot prompting on an expert-curated set of English materials. We ask whether models (1) select pragmatic interpretations of speaker utterances, (2) make similar error pat- terns as humans, and (3) use similar linguistic cues as humans to solve the tasks. We find that the largest models achieve high accuracy and match human error patterns: within incor- rect responses, models favor literal interpreta- tions over heuristic-based distractors. We also find preliminary evidence that models and hu- mans are sensitive to similar linguistic cues. Our results suggest that pragmatic behaviors can emerge in models without explicitly con- structed representations of mental states. How- ever, models tend to struggle with phenomena relying on social expectation violations. 1 Introduction Non-literal language understanding is an essential part of communication. For example, in everyday conversations, humans readily comprehend the non- literal meanings of metaphors (My new coworker is a block of ice), polite deceits (I love the gift), indirect requests (It’s a bit cold in this room), and irony (Classy pajamas, dude!). These phenomena fall under the broad label of pragmatics, which encompasses the aspects of meaning that go beyond the literal semantics of what is said (Horn, 1972; Grice, 1975; Yule, 1996; Levinson, 2000). A long-standing challenge for NLP is to build models that capture human pragmatic behaviors. Code and data: https://github.com/jennhu/lm-pragmatics The remarkable abilities of modern language mod- els (LMs) have triggered a recent effort to investi- gate whether such models capture pragmatic mean- ing, both through philosophical arguments (Bisk et al., 2020; Bender and Koller, 2020; Potts, 2020; Michael, 2020) and empirical evaluations (Jeretic et al., 2020; Zheng et al., 2021; Tong et al., 2021; Liu et al., 2022; Ruis et al., 2022; Stowe et al., 2022). However, prior empirical studies have pri- marily evaluated LMs based on a binary distinction between pragmatic and non-pragmatic responses, providing limited insights into models’ weaknesses. A model could fail to reach the target pragmatic interpretation in multiple ways – for example, by preferring a literal interpretation, or by preferring a non-literal interpretation that violates certain social norms. Understanding these error patterns can sug- gest specific directions for improving the models, and foreshadow where pragmatics might go awry in user-facing settings (e.g., Saygin and Cicekli, 2002; Dombi et al., 2022; Kreiss et al., 2022). From a cognitive …

Content truncated to 500 words