软件导刊2025，Vol.24Issue(10)：65-72,8.DOI:10.11907/rjdk.241782
基于元学习实例权重法的跨领域深度文本匹配
An Instance Weighting Method Based on Meta-Learning for Cross-Domain Deep Text Matching
摘要
Abstract
In low resource deep text matching scenarios,transfer learning utilizes resource rich source domain data to alleviate the data fam-ine problem in the target domain;When there is a distribution difference between the source domain data and the target domain data,the un-reasonable use of the source domain data will result in negative transfer phenomenon.Therefore,a cross domain deep text matching method based on meta learning instance weighting method is proposed.Firstly,by minimizing the loss of a small amount of annotated data in the low re-source target domain,appropriate importance weights are adaptively assigned to the data in the source domain to better fit the data distribution in the target domain;Secondly,in order to alleviate the catastrophic forgetting problem that may occur in the pre trained language model dur-ing the transfer process,meta learning is only performed at the classifier layer of the pre trained language model.Experiments have shown that on the natural language inference dataset SciTail,RE2 and ALBERT have improved F1 values by 2.0%and 1.5%,respectively,compared to the suboptimal baseline method;On the CQADupStack dataset for paraphrasing detection,RE2 and ALBERT showed a maximum average im-provement of 4.1%and 1.7%in F1 values compared to the suboptimal baseline method,respectively.关键词
低资源/文本匹配/元学习/迁移学习/实例权重法Key words
low-resource/text matching/meta-learning/transfer learning/instance weighting algorithm分类
计算机与自动化引用本文复制引用
李铂鑫,黄琪,鲁骁,张霄,王斌..基于元学习实例权重法的跨领域深度文本匹配[J].软件导刊,2025,24(10):65-72,8.基金项目
国家自然科学基金项目(62466028) （62466028）
