The Practice and Reflection of Large-Language Models in Regulated Verse Composition: A Human-AI Interaction Analysis Based on Deepseek

Abstract

This study investigates the application of the large language model Deepseek in classical regulated verse (Jintishi) by analyzing its interactions with both teachers and students. Using a case study approach, five sets of human-AI dialogue records were examined to assess the model’s understanding of prosodic rules and its ability to support poetic creation. The results show that Deepseek demonstrates basic knowledge of poetic structure and the ability to generate rule-based content, responding appropriately to questions about rhyme, metre, and style. However, it remains prone to errors in tonal marking and rhyme classification. Student interactions with AI reveal the feasibility of assisted poem composition, with some learners showing initiative in revision and cultural contextualization, though others exhibited over-reliance on AI and a weak grasp of form. The study highlights human-AI dialogue as a dual evaluative space—for assessing both AI cognition and student literary competence—offering implications for educational use and model development. Recommendations are made for educators, learners, and AI designers to enhance the integration of AI in classical poetry learning.