社会–认知语用视角下大语言模型识别与生成反讽能力的探究

Exploring the Ability of Large Language Models to Recognize and Generate Irony: A Sociocognitive Pragmatic Perspective

摘要: 本研究以社会认知语用学为理论基础，选取《脱口秀大会》中的反讽话语为语料，考察大语言模型DeepSeek识别与生成反讽的能力。通过一系列提示语，探究DeepSeek如何利用语境机制、凸显机制与合作–自我中心机制来处理反讽表达。并分析其基于社会认知语用学生成反讽语料的表现。研究发现，DeepSeek能够借助上述机制识别反讽话语，但在处理涉及文化背景语境的表达可能出现偏差；该模型能依据提示语设计出符合社会认知语用学特点的反讽语料，但其表达可能较为直白。针对这些局限，本文进一步探讨了通过改进训练数据、调整模型架构及优化提示工程策略等可操作的优化路径。鉴于反讽是一种复杂且多功能的现象，本研究不仅有助于提升公众对人工智能技术使用的认识，也为评估大语言模型提供了新的理论视角，并探讨了优化其语用能力的可能路径。

Abstract: Grounded in sociocognitive pragmatics, this study selects ironic utterances from the stand-up comedy show Rock & Roast as its corpus to examine the ability of the Large Language Model DeepSeek to recognize and generate irony. Through a series of prompts, the study explores how DeepSeek processes ironic expressions by leveraging contextual salience, salience, and the cooperation-egocentrism mechanisms, and analyzes its performance in generating ironic utterances informed by sociocognitive pragmatics. The findings reveal that DeepSeek can identify ironic utterances with the help of these mechanisms, though deviations may occur when cultural context contextual information is involved. The model is also able to generate ironic utterances that align with sociocognitive pragmatics based on the prompts, but its output tends to be overly explicit. In response to these limitations, this paper further discusses actionable optimization pathways through improving training data, adjusting model architecture, and refining prompt engineering strategies. Given that irony is a complex and multifaceted phenomenon, this study not only helps enhance public awareness of the application of artificial intelligence technologies, but also offers a new theoretical perspective for evaluating large language models and explores possible pathways for optimizing their pragmatic competence.