基于小样本对抗学习的Transformer检索框架OA
Transformer Retrieval Framework Based on Few-shot Adversarial Learning
针对密集检索方法在面对查询变体时性能下降的问题,提出一种基于小样本对抗学习的 Transformer 检索框架.首先,引入小样本学习,在数据有限的情况下提升模型泛化能力,并利用对抗训练增强对查询变体的鲁棒性;其次,设计代理注意力机制,通过代理标记与查询和文档关键特征的交互,在保留全局上下文的同时捕捉隐含语义层次,从而改善语义映射效果.在 TREC-DL-2019 和 TREC-DL-2020 数据集上的实验结果表明,所提框架具有优秀的检索精度和鲁棒性,为复杂查询场景下的信息检索提供了新思路.
A Transformer retrieval framework based on few-shot adversarial learning was proposed to ad-dress the performance degradation of dense retrieval methods when dealing with query variants.Firstly,few-shot learning was introduced to improve the generalization ability of the model with limited data,and adversarial training was employed to enhance robustness against query variants.Secondly,a proxy atten-tion mechanism was designed,in which proxy tokens interacted with key features in queries and docu-ments,so that implicit semantic levels were captured while global context was preserved,thereby impro-ving semantic mapping performance.The experimental results on the TREC-DL-2019 and TREC-DL-2020 datasets showed that the proposed framework had excellent retrieval accuracy and robustness,providing a new perspective for information retrieval in complex query scenarios.
胡广成;胡新荣
武汉纺织大学 计算机与人工智能学院 湖北 武汉 430200武汉纺织大学 计算机与人工智能学院 湖北 武汉 430200
信息技术与安全科学
密集检索小样本对抗学习查询变体代理注意力机制
dense retrievalfew-shot adversarial learningquery variantproxy attention mechanism
《郑州大学学报(理学版)》 2026 (6)
61-66,6
纺织工业联合会应用基础研究项目(202209)
评论