国际中文教师一直在探索如何更有效地提高汉语二语写作产出的效果。此前，“互动协同”教学模式已经受到关注。但是，传统的“互动协同”教学模式无法保证互动对象的即时性。2023年3月，OpenAI公司发布了ChatGPT的进阶版本ChatGPT Plus，即GPT-4。所以，基于互动对象即时性的问题，本研究探讨了生成式大语言模型GPT-4在辅助汉语二语写作产出方面的提升效果，对比了GPT-4与汉语母语语伴在辅助二语写作教学方面的异同，并深入探究了学习者的互动体验等因素对写作产出的影响。

本研究设计了一个行为实验，以23名HSK水平在4级至6级的汉语二语者为实验对象，选取了3个难度水平相当的作文话题（话题A《绿色食品与健康》、话题B《如何解决“代沟”问题》、话题C《如何面对挫折》），使实验对象分别在无交互、GPT-4辅助与汉语语伴辅助三种条件下进行写作，并从三元组比例、三元组多样性、依存距离、语法点比例与语法点密度5个方面的句法维度对实验对象的写作产出进行评价。此外，本研究还设计了李克特量表，通过量表追踪实验对象与GPT-4及汉语语伴互动的互动体验效果。

研究结果显示，在句法维度的评价上，相较于无交互条件，在GPT-4辅助条件下，写作产出中的主谓三元组比例、动宾三元组多样性、并列依存距离、高级句法语法点比例与高级句法语法点密度表现更为优异；而相较于汉语语伴辅助条件，在GPT-4辅助条件下，写作产出中的动宾三元组多样性、述补依存距离表现更为优异。此外，根据李克特量表，相较于无交互条件，在GPT-4辅助条件下，实验对象对作文话题的熟悉程度有所提升，书写的困难程度有所下降，书写的信心程度有所提升。而相较于汉语语伴辅助条件，在GPT-4辅助条件下，实验对象能够获得更加良好的交互体验感受。这都说明GPT-4为代表的生成式大语言模型具备辅助汉语二语写作产出的优越能力，能够进入传统“师-生”互动协同的教学模式，进而打造更新的“师-机-生”的教学模式。

同时，本研究也探究了GPT-4在辅助汉语二语写作产出方面的不足。基于此，研究团队可以对GPT-4为代表的生成式大语言模型进行针对性地升级，以期实现生成式大语言模型向精准化的汉语二语学习产品或教学机器人的转变。

International Chinese teachers have been exploring how to improve the writing output of Chinese as a second language more effectively. Previously, the "interactive and collaborative" teaching model has attracted attention. However, the traditional "interactive and collaborative" model cannot guarantee the immediacy of the interactive object. In March 2023, OpenAI company released ChatGPT Plus, an advanced version of ChatGPT, namely GPT-4. Therefore, based on the immediacy of the interactive object, this study explores the improvement effect of the generative large language model GPT-4 in assisting the writing output of Chinese as a second language, compares the similarities and differences between GPT-4 and Chinese native speakers in assisting the writing teaching of the second language, and delves into the influence of learners' interactive experience and other factors on the writing output.

This study designs a behavioral experiment, taking 23 Chinese second speakers with HSK level between level 4 and level 6 as the experimental objects, and selects three composition topics with similar difficulty levels (topic A "Green Food and Health", topic B "How to solve the ‘generation gap’ problem", topic C "How to face frustrations"), so that the experimental objects can write under three conditions: no interaction, GPT-4 assisted and Chinese native speaker assisted. The experimental objects' compositions are evaluated from five syntactic dimensions: triple proportion, triple diversity, dependency distance, grammatical point proportion and grammatical point density. In addition, the Likert scale was also designed to track the interaction effect of the experimental subjects with GPT-4 and Chinese language partner.

The results show that, in the evaluation of syntactic dimension, compared with the no interaction condition, under the GPT-4 auxiliary condition, the proportion of subject-predicate triples, the diversity of verb-object triples, the parallel dependency distance, the proportion of high-level syntactic grammatical points and the density of high-level syntactic grammatical points in the writing output are better; and compared with the Chinese language partner auxiliary condition, under the GPT-4 auxiliary condition, the diversity of verb-object triples and the  predicate-complement dependency distance in the writing output are better. In addition, according to the Likert scale, compared with the no interaction condition, under the GPT-4 auxiliary condition, the experimental subjects' familiarity with the topic of the composition is improved, the difficulty of writing is reduced, and the confidence of writing is improved. Compared with the Chinese language partner auxiliary condition, under the GPT-4 auxiliary condition, the experimental subjects can obtain better interactive experience. This shows that the generative large language model represented by GPT-4 has the superior ability to assist the writing output of the second language of Chinese, which can enter the traditional "teacher-student" interactive and collaborative teaching mode, and then create the "teacher-machine-student" teaching mode.

At the same time, this study also found the deficiencies of GPT-4 in assisting the writing output of the second language of Chinese. Based on this, the research team can upgrade the generative large language model represented by GPT-4, in order to realize the transformation of generative large language model to accurate Chinese second language learning products or teaching robots.