大语言模型的语用能力探索——从整体评估到反语分析  

Exploring the Pragmatic Capabilities of Large Language Models:From Overall Assessment to Irony Analysis

作　　者：刘海涛[1] 亓达 LIU Haitao;QI Da(Zhejiang University)

机构地区：[1]浙江大学

出　　处：《现代外语》2024年第4期439-451,共13页Modern Foreign Languages

基　　金：教育部人文社科重点研究基地重大项目“数据驱动的外语能力发展研究”(22JJD740018)的阶段性成果。

摘　　要：本文探讨大语言模型在处理语用现象时的表现,特别是它们对反语现象的处理能力。结果表明,新一代大语言模型在语用能力方面已有显著提升,但在处理反语和幽默等语用现象时,依然依赖于字面意义的直接理解,对复杂语境的推理能力有限。通过进一步分析GPT-4的具体表现,本文揭示了大模型在识别和生成反语时的认知动机与语用学理论框架之间的关系,为语用学理论和大语言模型的未来发展提出了建议。This paper analyzes the performance of large language models(LLMs)in understanding and generating pragmatic expressions,with a particular focus on their ability to handle irony.The findings suggest that the latest LLMs have made significant advancements in pragmatic capabilities.However,when dealing with pragmatic phenomena such as irony and humor,they still rely heavily on literal meanings and thus exhibit limited capacity for complex contextual reasoning.Through a further analysis of GPT-4’s performance,this study reveals the relationship between the cognitive motivations of LLMs in recognizing and generating irony and the theoretical frameworks of pragmatics.This study offers suggestions for future developments of both pragmatic theories and LLMs.

关 键 词：大语言模型 语用能力 反语 

分 类 号：H030[语言文字—语言学]