Scholarlink

<p>Pragmatic competence, the ability to infer implied meaning, recognize presuppositions, and interpret speech acts, has traditionally been tied to embodied experience and social interaction. Whether large language models (LLMs) can approximate these abilities has become an increasingly pressing question. This study compares the performance of 70 Uzbek undergraduate students and two state-of-the-art LLMs (GPT-4 and Gemini 2.0) across three domains: conversational implicature, presupposition, and speech acts, using a controlled set of 60 stimuli. While the models handled some conventionalized pragmatic cues, they consistently underperformed relative to human participants, particularly in tasks requiring contextual inference, presupposition accommodation, and recognition of indirect communicative intent. Error patterns revealed systematic tendencies toward literal interpretation, accommodation failures, and misidentification of illocutionary force. The results indicate that current LLMs approximate the surface form of pragmatic behavior without reliably engaging the inferential and social mechanisms that underlie it in human communication.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Pragmatic competence without embodiment? Evaluating LLM performance on implicature, presupposition, and speech acts

  • 期刊论文

  • 发表日期：

    2026年4月4日

查看全文
  • Dilyorjon Solidjonov

摘要

Pragmatic competence, the ability to infer implied meaning, recognize presuppositions, and interpret speech acts, has traditionally been tied to embodied experience and social interaction. Whether large language models (LLMs) can approximate these abilities has become an increasingly pressing question. This study compares the performance of 70 Uzbek undergraduate students and two state-of-the-art LLMs (GPT-4 and Gemini 2.0) across three domains: conversational implicature, presupposition, and speech acts, using a controlled set of 60 stimuli. While the models handled some conventionalized pragmatic cues, they consistently underperformed relative to human participants, particularly in tasks requiring contextual inference, presupposition accommodation, and recognition of indirect communicative intent. Error patterns revealed systematic tendencies toward literal interpretation, accommodation failures, and misidentification of illocutionary force. The results indicate that current LLMs approximate the surface form of pragmatic behavior without reliably engaging the inferential and social mechanisms that underlie it in human communication.