Download Full PDF

莫子良

王艺洁

吴继峰 

首都师范大学，中国

 

摘要

近年来，元话语分析被广泛用于考察学术文本的语言特征及其读者导向功能。随着以 ChatGPT为代表的生成式人工智能（ GenAI）的迅速发展，学界开始关注人工智能生成文本中的元话语使用情况。然而，在学术汉语领域，相关实证研究仍相对不足。为探讨 GenAI 生成学术汉语文本中元话语的使用特征，本文基于自建配对语料库，对比分析了 ChatGPT-4o 生成与人类撰写的理论语言学和应用语言学领域学术汉语摘要中的元话语使用差异。研究结果表明 ：在引导式元话语方面， ChatGPT 生成摘要与人类摘要在总体使用频率上呈现一定共性，但在篇章整体结构组织上仍弱于人类写作。在互动式元话语方面， ChatGPT 生成的应用语言学摘要相较理论语言学摘要表现更具优势，说明其更擅长生成应用类文本，但整体上仍存在公式化写作、作者意识不足等问题。本研究为自然语言生成与处理研究提供了实证依据，助力优化相关模型的文本生成机制，同时有助于提升研究者对 GenAI 生成文本的识别能力，促进其在学术研究中的合理使用。此外，研究结论也可为学术汉语写作教学提供一定参考。

 

关键词

ChatGPT，学术语篇，元话语，理论语言学，应用语言学

 

A Comparative Study of Metadiscourse in ChatGPT-generated and Human-written Academic Chinese Abstracts
 

Ziliang Mo

Yijie Wang

Jifeng Wu

Capital Normal University, China

 

Abstract

In recent years, metadiscourse analysis has been widely employed to examine the linguistic features of academic texts and their reader-oriented functions. With the rapid development of generative artificial intelligence (GenAI), represented by ChatGPT, increasing scholarly attention has been directed toward the use of metadiscourse in AI-generated texts. However, empirical research in the field of academic Chinese remains relatively limited. To investigate the characteristics of metadiscourse use in AI-generated academic Chinese texts, this study draws on a self-compiled corpus to conduct a comparative analysis of metadiscourse in academic Chinese abstracts generated by ChatGPT-4o and those written by humans in the fields of theoretical linguistics and applied linguistics. The results show that, with regard to interactive metadiscourse, ChatGPT-generated abstracts and human-written abstracts exhibit certain similarities in overall frequency of use. Nevertheless, AI-generated texts remain weaker than human writing in terms of global discourse organization. Concerning interactional metadiscourse, the applied linguistics abstracts generated by ChatGPT demonstrate relatively stronger performance than those in theoretical linguistics, suggesting that the model is more adept at producing application-oriented texts. However, issues such as formulaic writing patterns and insufficient authorial presence still persist. This study provides empirical evidence for research on natural language generation and processing, contributing to the optimization of text generation mechanisms in related models. It also helps enhance researchers’ ability to identify AI-generated texts and supports their appropriate use in academic research. In addition, the findings offer potential implications for the teaching of academic Chinese writing.

 

Keywords

ChatGPT, academic discourse, metadiscourse, theoretical linguistics, applied linguistics