| 摘 要: 为推动“河州花儿”的传承与发展,提出了一种句法和主题增强的唱词生成模型(Syntactic-Theme Transformer,ST-Transformer)。利用词性与依赖关系注意力模块进行特殊句法编码。运用潜在狄利克雷分配模型(LatentDirichletAllocation,LDA)挖掘潜在主题并作为标识符引导特定主题唱词生成。采用三层异构注意力机制融合句法和主题信息提升生成能力。与表现较好的BART模型相比,在指标BLEU-2、BLEU-3、BLEU-4和Topic Consistency上分别提升19.14%、20.17%、20.78%和9.40%。结果表明ST-Transformer在句法和主题表达方面均展现出良好的性能,生成的唱词具备较高的质量。 |
| 关键词: Transformer模型 河州花儿 唱词生成 句法信息 主题信息 三层异构注意力 |
|
中图分类号: TP391.1
文献标识码: A
|
| 基金项目: 甘肃省重点研发计划项目(21YF5GA088) |
|
| Research on Lyric Generation Technology for “He zhou Hua’er” Based on the Transformer Model |
|
RAN Xiaorui, WANG Lianguo
|
(College of Information Science and Technology, Gansu Agricultural University, Lanzhou 730070, China)
18194218469@163.com; wanglg@gsau.edu.cn
|
| Abstract: To promote the inheritance and development of “Hezhou Hua’er”(a traditional Chinese folk song), this study proposes a syntax-and theme-enhanced lyric generation model (Syntactic-Theme-Transformer, ST-Transformer).First, par-t o-f speech and dependency-based attention modules are employed for specialized syntactic encoding. Second,the Latent Dirichlet Allocation (LDA) model is utilized to mine latent themes, which serve as identifiers to guide theme-specific lyric generation. Finally, a triple-layer heterogeneous attention mechanism integrates syntactic and thematic information in rea-l time to enhance generation capability. Compared with the high-performing BART model,ST-Transformer achieves improvements of 19.14% , 20.17% , 20.78% , and 9.40% on BLEU-2, BLEU-3, BLEU-4, and Topic Consistency metrics, respectively. The results demonstrate that ST-Transformer exhibits strong performance in both syntactic structure and thematic expression, generating lyrics of significantly higher quality |
| Keywords: transformer model Hezhou Hua’er lyric generation syntactic information thematic information triple-layer heterogeneous attention |
