基于适应性自训练的少样本关系抽取建模  

Modeling of Few-Shot Relation Extraction Based on Adaptive Self-Training

作　　者：陈洪辉[1] 郑建明 蔡飞[1] 韩毅[2] Chen Honghui;Zheng Jianming;Cai Fei;and Han Yi(College of System Engineering,National University of Defense Technology,Changsha 410073;College of Meteorology and Oceanography,National University of Defense Technology,Changsha 410073)

机构地区：[1]国防科技大学系统工程学院,长沙410073 [2]国防科技大学气象海洋学院,长沙410073

出　　处：《计算机研究与发展》2023年第7期1581-1591,共11页Journal of Computer Research and Development

基　　金：湖南省研究生科研创新项目(CX20190034,CX20210068)。

摘　　要：关系抽取(relation extraction,RE)是自然语言处理中的一项基础任务,可以支撑许多下游任务,例如对话生成和机器阅读理解等.在现实生活中,由于新关系类别不断涌现,人工标注的成本和速度无法满足传统基于有监督学习的关系抽取模型的训练要求.面对这种现实挑战,神经雪球提出一种自助采样的方法,通过对有限标注数据的信息迁移,不断为无标注数据打上标签,增加标注数据量,从而提升模型分类性能.然而,固定的阈值选择以及同等对待入选的无标注数据使得神经雪球模型容易受到噪声数据的影响.为了解决这2个缺陷,基于适应性自训练的关系抽取(adaptive self-training relation extraction,Ada-SRE)模型由此提出.具体地,Ada-SRE基于元学习的思想提出自适应阈值模块,能够为每个关系类别提供合适的阈值选择.另外,Ada-SRE还提出基于梯度反馈的赋权策略,为每个入选的示例提供相应的权重,避免噪声数据的干扰.实验结果表明,相比于神经雪球模型,Ada-SRE有更好的关系抽取能力.Relation extraction(RE)is a basic task in natural language processing,which supports plenty of downstream tasks,e.g.,dialogue generation,machine reading comprehension,etc.In real life,due to the continuously emerging new relation labels,the speed and cost of human annotation cannot catch up with the data quantity that the training of the traditional supervised RE models demands.Facing this practical challenge,the neural snowball model proposes a bootstrapping method that transfers the RE knowledge from limited labeled instances to iteratively annotate unlabeled data as to increase the amount of labeled data,thereby improving the classification performance of the model.However,the fixed threshold selection and the equally treated unlabeled data make the neural snowball model vulnerable to noise data.To solve these two defects,an adaptive self-training relation extraction(Ada-SRE)model is proposed.In specific,for the fixed-threshold issue,Ada-SRE proposes an adaptive threshold module by the meta learning of threshold,which can provide an appropriate threshold for each relation category.For the equallytreated issue,Ada-SRE designs a gradient-feedback strategy to weight each selected example,avoiding the interference of noise data.The experimental results show that compared with the neural snowball model,Ada-SRE has a better relation extraction ability.

关 键 词：自训练 关系抽取 梯度反馈 少样本学习 元学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]