Binding by structure, distracted by cues: Principle A in Language Models

• 발행기관: 경희대학교 언어정보연구소
• 저자명: Jieun Kim
• 간행물 정보: 『언어연구』제43권 제2호, 427~450쪽, 전체 24쪽
• 주제분류: 인문학 > 언어학
• 파일형태: PDF
• 발행일자: 2026.06.30

국문 초록

Binding Principle A provides a well-studied testing ground for examining how hierarchical syntactic constraints regulate dependency formation (Chomsky 1981, 1986). Psycholinguistic research has long debated whether reflexive interpretation relies on structural filtering that excludes illicit antecedents in advance (Nicol and Swinney 1989; Sturt 2003), or on cue-based retrieval in which structural information competes with graded similarity cues (Lewis and Vasishth 2005; Dillon et al. 2014). This study revisits that debate in transformer-based language models. Using controlled reflexive configurations from the Pythia family (410M, 2.8B, 6.9B), we systematically dissociate structural accessibility from feature-based similarity by independently manipulating c-command, locality, antecedent number, distractor animacy, and model scale. We evaluate models using surprisal at the reflexive themselves. Across models and sentence types, a highly consistent ordering emerges: TRUE ANT < FALSE ANT < CORRUPTED. Structurally licensed plural antecedents yield the lowest surprisal, whereas structurally illicit but number-matching distractors produce only partial facilitation. This asymmetry provides direct evidence that model predictions are guided by structural accessibility rather than surface plurality alone. At the same time, the intermediate status of FALSE ANT reveals reliable distractor interference, indicating that feature-matching competitors are not fully ignored. Additional analyses show that animate distractors generate stronger interference than inanimate distractors, and that smaller models are substantially more susceptible to such interference than larger models. These findings suggest that transformer language models encode a hybrid retrieval profile: reflexive prediction is strongly constrained by structural information, yet remains probabilistically sensitive to competing semantic and featural cues.

목차

1. Introduction

2. Theoretical background

3. Experiment

4. Results and discussions

5. General discussion and conclusion

References

키워드

reflexive binding Binding Principle A transformer-based language models syntactic structure computational model processing cue-based retrieval

참고문헌

• Psychonomic Bulletin & Review
• Dordrecht
• Cambridge, MA
• Journal of Memory and Language
• Frontiers in Psychology
• Journal of Experimental Psychology: Learning, Memory, and Cognition
• Journal of Memory and Language
• Journal of Experimental Psychology: Learning, Memory, and Cognition
• Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies
• Second Meeting of the North American Chapter of the Association for Computational Linguistics
• Cognition
• Cognitive Science
• Transactions of the Association for Computational Linguistics
• Journal of Experimental Psychology: Learning, Memory, and Cognition
• Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing
• Journal of Psycholinguistic Research
• Journal of Memory and Language
• Journal of Psycholinguistic Research
• Linguistic Inquiry
• Cambridge, MA
• Bell System Technical Journal
• Cognition
• Journal of Memory and Language
• Journal of Memory and Language
• Journal of Memory and Language