Cross-Linguistic Syntactic Parsing in GPT Models: Evidence from English and Korean Ambiguities
4
조회
0
다운로드
0
인용
0
참고문헌
키워드
초록
This study investigates how three GPT-based large language models—ChatGPT-3.5, ChatGPT-4o, and ChatGPT-5.0—resolve globally ambiguous structures in English and Korean, focusing on relative clause (RC) attachment and temporal adverbial adjunct (TAA) attachment ambiguities. Across both constructions, all three models produced near-categorical low-attachment interpretations in English. Although English speakers generally prefer low attachment, human judgments are gradient; thus, the models’ uniformly categorical responses reflect an overgeneralized heuristic rather than human-like syntactic parsing. In Korean, where human readers are known to favor high attachment, the models again exhibited strong low-attachment preferences for both RC and TAA structures, with ChatGPT-4o and ChatGPT-5.0 showing especially rigid, near-ceiling low-attachment behavior. These cross-linguistic patterns indicate that GPT models do not acquire language-specific attachment strategies but instead extend an English-derived low-attachment heuristic across typologically distinct languages. The findings underscore key limitations in current autoregressive LLMs’ cross-linguistic syntactic generalization and highlight the need for training and modeling approaches that more effectively capture typological variation in structural interpretation.
Abstract
This study investigates how three GPT-based large language models—ChatGPT-3.5, ChatGPT-4o, and ChatGPT-5.0—resolve globally ambiguous structures in English and Korean, focusing on relative clause (RC) attachment and temporal adverbial adjunct (TAA) attachment ambiguities. Across both constructions, all three models produced near-categorical low-attachment interpretations in English. Although English speakers generally prefer low attachment, human judgments are gradient; thus, the models’ uniformly categorical responses reflect an overgeneralized heuristic rather than human-like syntactic parsing. In Korean, where human readers are known to favor high attachment, the models again exhibited strong low-attachment preferences for both RC and TAA structures, with ChatGPT-4o and ChatGPT-5.0 showing especially rigid, near-ceiling low-attachment behavior. These cross-linguistic patterns indicate that GPT models do not acquire language-specific attachment strategies but instead extend an English-derived low-attachment heuristic across typologically distinct languages. The findings underscore key limitations in current autoregressive LLMs’ cross-linguistic syntactic generalization and highlight the need for training and modeling approaches that more effectively capture typological variation in structural interpretation.