Automated Comment Generation Based on the Large Language Model

Kaiwei Cai, Junsheng Zhou, Li Kong, Dandan Liang, Xianzhuo Li

Abstract: Automated essay comment generation is important for writing in education, as it can reduce the burden on teachers while providing students with rapid feedback to improve their writing. When writing, students tend to take writing thought ideas in several excellent essays as references, which inspire them to be more creative in writing. For teachers, they can refer to them to make suggestions on how to improve the quality of essays. Inspired by this behaviour of writing psychology, we guide the Large Language Model (LLM) to also imitate this behaviour. In this paper, we apply the essay comment generation task in application, which aims to generate comments on how to amplify the writing thought ideas for the given English expository essay. To tackle this task, we propose the two-stage comment generation framework, in which we first search for some cases of the chain of writing thought ideas, and then use them as evidence to guide the LLM to learn from these references and generate more concrete comments. What’s more, we manually collect some English expository essays to build the knowledge base and some essay-comment pairs (The source code is available at https://github.com/CarryCKW/EssayComGen). Extensive experiments show that our method outperforms strong baselines significantly.