基于改进级联二元标记框架的关系抽取方法

Relationship extraction method based on improvedcascaded binary tag framework

在实体关系抽取中,以往的关系抽取模型将实体间关系视作离散标签,导致抽取的特征信息存在误差,造成特征样本数量不均衡,使得多个关系三元组共享同一个实体,产生了三元组重叠的情况。在研究三元组抽取方法的过程中,针对上述问题,本文提出了一种基于注意力机制改进级联二元标记框架的关系抽取方法(AMCT)。本文在NYT和Web NLG数据集上开展了对比实验,实验结果表明新的抽取方法与一些通用的关系抽取方法比较,有效地提高了抽取的精确度和召回率。

In entity relationship extraction,the previous relationship extraction model regards the relationship between entities as discrete labels,resulting in errors in the extracted feature information,causing an imbalance in the number of feature samples,making multiple relational triplets share the same entity,resulting in triple overlap.In the process of studying the triplet extraction method,we propose an relation extraction method(AMCT)based on attention mechanism and improved cascaded binary tag framework.In this paper,comparative experiments are carried out on NYT and Web NLG data sets,and the experimental results show that compared with some general relational extraction methods,the new extraction method can effectively improve the extraction accuracy and recall rate.