北京大学学报（自然科学版） ›› 2026, Vol. 62 ›› Issue (2): 266-274.DOI: 10.13209/j.0479-8023.2025.097

上一篇    下一篇

融合CINO-LoRA和Self-condition的DiffuSum藏文文本自动摘要

王蓉1,2, 才智杰1,2,†   

  1. 1. 青海师范大学计算机学院, 西宁 810016 2. 藏语智能全国重点实验室, 西宁 810008
  • 收稿日期:2025-02-18 修回日期:2025-09-06 出版日期:2026-03-20 发布日期:2026-03-20
  • 基金资助:
    国家自然科学基金(616966031, 6186646462)资助

Automatic Summarization of Tibetan Texts Using DiffuSum with CINO-LoRA and Self-condition Integration

WANG Rong1,2, CAI Zhijie1,2,†   

  1. 1. College of Computer Science and Technology, Qinghai Normal University, Xining 810016 2. The State Key Laboratory of Tibetan Intelligence, Xining 810008
  • Received:2025-02-18 Revised:2025-09-06 Online:2026-03-20 Published:2026-03-20

PDF

摘要：

为进一步提升藏文文本自动摘要的性能, 针对DiffuSum模型在藏文摘要任务中因句子表征能力不足、参数规模过大导致的上下文建模受限以及训练成本高等问题, 提出一种融合CINO-LoRA与自调节(Self-condition)的藏文文本自动摘要模型TiDiffuSum。该模型在句子编码器中引入CINO-LoRA机制, 以增强藏文语义表征并显著减少训练参数量; 在扩散生成模块中集成Self-condition策略, 加强对上下文语义的理解与利用。实验结果表明, TiDiffuSum在藏文摘要数据集TSUM上能够将参数量有效压缩至基线模型的0.45%, 且ROUGE-1, ROUGE-2和ROUGE-L指标分别提升1.07, 0.78和1.08, 显著优于基线模型。

关键词: 藏文, 文本自动摘要, DiffuSum模型, 句子表征

Abstract:

To further improve the performance of Tibetan text automatic summarization, the paper proposes a Tibetan text summarization model TiDiffuSum, which integrates CINO-LoRA and Self-condition into the DiffuSum to address issues of insufficient sentence representation, large parameter scale limiting contextual modeling, and high training costs in the Tibetan task. TiDiffuSum model introduces CINO-LoRA mechanism into the sentence encoder to enhance Tibetan semantic representation and significantly reduce the number of training parameters. Additionally, it incorporates Self-condition strategy in the diffusion generation module to strengthen the comprehension and utilization of contextual semantics. Experimental results indicate that TiDiffuSum can effectively reduce the parameter count to 0.45% of the baseline model on the Tibetan summarization dataset (TSUM), and achieves improvements of 1.07, 0.78, and 1.08 in ROUGE-1, ROUGE-2, and ROUGE-L scores, significantly outperforming baseline models.

Key words: Tibetan, text automatic summarization, Diffusum model, sentence representation

版权所有 © 《北京大学学报（自然科学版）》编辑部
技术支持：北京玛格泰克科技发展有限公司