Answers are grounded in this paper and cite their sources.
Chinese classical poetry is a carrier for Chinese culture, moving people with deep emotional meanings and rich aesthetic conceptions through its brief language form. There are also problems with existing texts generated by relevant algorithms: monotonous styles and poor cultural applicability. Against such a background, the deep integration of artificial intelligence technology and classical poems, to realize the reciprocal application of the cultural inheritance with the technological innovation, has important interdisciplinary research value. It can not only enrich the research approach in the field of computational linguistics, but also provide solid foundation to strengthen the digital humanities as well as educational application. More innovative aspects of this study are reflected in three points: First, a generation mechanism of keywords is established, and a keyword module for integration of allusions and metaphors. So can the model create poetry with a fixed number of couplets given by the user, keyed by keywords chosen by the user or given by the first couplet’s upper line. Users can also adjust type parameters to generate five-character/seven-character regulated verses or quatrains. Additionally, a knowledge base containing historical and literary allusions from ancient documents is established, and semantic embedding technology and knowledge graphs are used to enhance the depth of metaphor expression. Meanwhile, a "Style" parameter is introduced to call corpus of seven poetry stylesto achieve precise style control. Second, an attention auxiliary loss function is proposed, combined with customized poetry identifiers and rhyme embedding technology. Information of 106 rhyme categories in Pingshui Rhymes is integrated into word vectors to guide the model in capturing formal features such as metrical rules and rhyming, alleviating the contradiction between content optimization and format compliance. Third, the generative capability of GPT-2 is integrated with a BERT-based poetry evaluation model. The BERT model, fine-tuned as a regression task, outputs a quality score in the range [0, 1] to evaluate poetry from dimensions including emotional resonance and use of literary devices, improving the accuracy and polishing quality of generated texts. Experiments were conducted based on the Chinese-poetry dataset. Compared with models such as Qwen and GPT-3.5, the proposed model achieved a 12% improvement in Format Accuracy over the best general large language model, with a Manual Evaluation score of 94.29, and could correctly adapt allusions like "silver saddle shines on white horse". Ablation experiments verified that the attention auxiliary loss significantly improved metrical accuracy. This study provides an effective path for the inheritance and innovation of classical poetry in the digital age.
No takes yet. Share an insight, caveat, or question.
Yao et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: