Scholarlink

Most current techniques to automatically compose traditional Vietnamese poems sometimes have different challenges in managing material and structure, leading to limitations in the poem’s coherence. Typically, these approaches are mainly based on keyword-focused inputs, resulting in a lack of vocabulary diversity and not sufficiently reviewing the context and importance of the generated poems. This work introduces a novel approach, RAG-ViVerse, which integrates sophisticated deep-learning methods with semantic analysis to generate ancient Vietnamese poetry. We use a multi-tiered contextual control system, using rhyme and semantic cues to maintain structural coherence and thematic consistency across various poetry genres. Experimental results demonstrate that the proposed paradigm significantly improves lyrical diversity, contextual depth, and semantic complexity compared to current standards. Our model achieves a 20% higher accuracy in response rate when utilizing poetic rules compared to direct responses from GPT versions. The model will simultaneously focus on using the elements included in the poem and the diversity of word classes.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

RAG-ViVerse: Enhancing Poetry Generation with Retrieval Augmented Generation and Large Language Models

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Son T. Huynh,
  • Phu G. Lam,
  • Hung P. Tran,
  • Binh T. Nguyen

摘要

Most current techniques to automatically compose traditional Vietnamese poems sometimes have different challenges in managing material and structure, leading to limitations in the poem’s coherence. Typically, these approaches are mainly based on keyword-focused inputs, resulting in a lack of vocabulary diversity and not sufficiently reviewing the context and importance of the generated poems. This work introduces a novel approach, RAG-ViVerse, which integrates sophisticated deep-learning methods with semantic analysis to generate ancient Vietnamese poetry. We use a multi-tiered contextual control system, using rhyme and semantic cues to maintain structural coherence and thematic consistency across various poetry genres. Experimental results demonstrate that the proposed paradigm significantly improves lyrical diversity, contextual depth, and semantic complexity compared to current standards. Our model achieves a 20% higher accuracy in response rate when utilizing poetic rules compared to direct responses from GPT versions. The model will simultaneously focus on using the elements included in the poem and the diversity of word classes.