融合多任务学习和实体遮掩的关系三元组抽取模型  

Relational Triplet Extraction Model Fused with Multi Ttask Learning and Entity Masking

作　　者：薛志豪 李永强[1] 赵永智 冯远静[1] XUE Zhihao;LI Yongqiang;ZHAO Yongzhi;FENG Yuanjing(School of Information Engineering,Zhejiang University of Technology,Hangzhou 310023,China)

机构地区：[1]浙江工业大学信息工程学院,杭州310023

出　　处：《小型微型计算机系统》2024年第4期845-852,共8页Journal of Chinese Computer Systems

基　　金：国家自然科学基金面上项目(62073294)资助.

摘　　要：针对流水线式三元组抽取模型中,命名实体识别任务里实体识别和实体分类两个子任务之间存在干扰,关系抽取任务中实体提及词对关系分类也存在干扰的问题,提出一种融合多任务学习和实体遮掩的关系三元组抽取模型.在命名实体识别任务中,先利用预训练模型对输入句子进行特征编码表示,然后利用首尾指针标注进行实体识别,最后利用注意力机制融入实体类型信息进行实体分类.在关系抽取任务中,提出了一种实体遮掩的方法,先利用实体类型信息替换实体提及词,并在其前后插入实体标记,之后利用预训练模型对输入句子进行特征编码表示,最后利用头尾实体的特征表示进行关系分类.在SCIERC和SKE两个数据集上进行大量实验,实验结果表明,所提模型相较于基于实体标记方法的PURE模型整体性能提升了2.5和1.5个百分点.充分验证了在三元组抽取任务中,分解命名实体识别任务以及在关系抽取中用实体类型信息替换实体提及词的有效性.To solve the problem of interference between entity recognition and entity classification sub tasks in named entity recognition task and interference between entity reference words and relationship extraction task in relationship extraction task in pipeline relational triplet extraction model,we proposed a relational triplet extraction model that integrates multi task learning and entity masking.In the task of named entity recognition,the input sentences were coded using the pre training model.Then,the first and last pointer labels were used for entity recognition.Finally,the attention mechanism was used to integrate entity type information to classify entities.In relation extraction task,we proposed an entity masking method.Firstly,we used entity type information to replace entity reference words,and inserted entity markers before and after them.Then we used the pre training model to encode the input sentences.Finally,we used pointer labels of head and tail entities to classify relationships.A large number of experiments had been conducted on SCIERC and SKE data sets.The experimental results showed that the overall performance of the proposed model was 2.5 and 1.5 percentage points higher than the model of PURE which based on entity tagging method.It fully verified the effectiveness of decomposing entity recognition tasks and eliminating entity references in relation extraction for triplet extraction tasks.

关 键 词：信息抽取 关系三元组抽取 多任务学习 实体遮掩 注意力机制 深度学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]