基于融合关系信息编码的法律文书实体关系抽取方法

Relation Enhanced Embedding Based Entities Relation Extraction from Legal Documents

关系重叠是实体关系抽取任务中的一个难题,具体表现为文本中的一个实体存在于多个关系之中。法律文书作为一类存在大量关系重叠的文本,在对其进行关系抽取时要识别这些重叠的关系,而现有的关系抽取方法存在错误传播或识别率不高的问题。针对这种情况,该文提出了联合标注法,构建了联合抽取模型。模型在编码器中获取文本中的关系信息,使用拼接法或权值法来对关系信息和文本信息进行融合,使编码带有关系信息,采用共享的解码器对重叠关系进行识别。在法律文书数据集上进行实验,结果表明,与现有前沿关系抽取方法相比,该关系抽取方法能够更好地识别重叠关系。

Overlapping relation,i.e.one entity with multiple relations in a text,is a challenging issue in the task of entities relation extraction.Legal documents,as a type of text with a large number of overlapping relations,typical-ly demonstrate the error propagation or low recognition rate owing to this issue for current relation extraction meth-ods.This paper presents a joint extraction model based on a joint annotation method.Our model obtains the relation information in the text encoder,and embeds the relation information with the text representation via concatenating or weighting.Then a shared decoder is employed to identify the overlapping relations.Experimental results show that,compared with the existing advanced relation extraction methods,this proposed method can better identify o-verlapping relation.