PATB:一种面向联合实体和关系抽取的信息聚合器

PATB:an Information Booster for Joint Entity and Relationship Extraction

现有实体关系联合抽取方法中,主体抽取与客体和关系抽取任务的交互不足或方法单一,对关系三元组内部潜在的位置及上下文语义关系利用不足.为此,提出了一种融合实体位置及上下文注意力的信息聚合器(Position and Attention based Booster,PATB)用于级联式实体关系联合抽取.首先抽取主体,再融合主体位置更新主体的表示,融合主体与上下文的注意力更新文本的表示,将更新的主体及文本表示进一步用于客体及关系抽取.模型在公共数据集NYT和WebNLG上的F1值分别为90.9%、92.5%,较基线模型分别提升1.3%和0.7%;在3种不同关系模式的测试数据Normal、EPO及SEO中,NYT上的F1值分别为88.9%、93.2%和92.6%,均优于基线模型;在含1~5个三元组的对比实验中的F1值也均优于基线模型,表明融合位置及上下文语义的PATB不仅可提升三元组抽取性能,且能在有复杂重叠关系、多个三元组情况下保持稳定的提取性能.

In existing cascaded entity relationship extraction methods,the interaction between subject extraction and object and relationship extraction tasks is generally insufficient or simple,which makes the potential positional and contextual semantic relationships between three elements underutilized.The paper proposes a Position and Attention Booster(PATB)based on entity positions and contextual attention for cascading entity-relationship extraction.Firstly,the subject is extracted,then the subject position is fused to update the subject representation,the attention of subject and context is fused to update the text representation,and the updated subject and text representations are further used for object and relationship extraction.The proposed method achieves F1 score of 90.9%and 92.5%on public datasets NYT and WebNLG,respectively,which are 1.3%and 0.7%outperforming the baseline.For three overlapping patterns,including Normal,EPO,and SEO,F1 values of the proposed method are 88.9%,93.2%,and 92.6%,respectively,which are also better than baseline.Moreover,F1 score are also outperformed the baseline in comparison experiments for texts containing 1 to 5 triple elements.It can be demonstrated that PATB fusing positional and contextual semantics can not only improve the extraction performance of triplets,but also achieve stable extraction performance in case of complex overlapping relationships and multiple triplets.