基于提及图和显式路径的文档级关系抽取方法

Document-level relation extraction method based on mention graph and explicit path

为有效汇聚实体间信息,利用实体间路径信息判断语义关系,提出一种基于提及图和显式路径的文档级关系抽取方法。利用基于提及对类型的图构建方法,通过融合提及间天然存在的结构化信息,结合图注意力网络,实现实体间信息的汇聚;利用基于显式路径的关系推理方法,包括显式路径构建方法和启发式的路径特征融合方法两个子方法,通过显式方式构建关系推理路径,将推理路径分为句内推理路径、句间推理路径、直接推理路径,实现区分句内推理和句间推理,差异化融合路径特征,提高关系路径推理能力,增强关系抽取的准确度。在3个公开数据集上的对比试验表明,本方法在F_(1)和Ign F_(1)指标上较目前主流方法存在优越性,验证了基于提及图和显式路径的文档级关系抽取方法能够更有效地支持文档级关系抽取任务。

To effectively aggregate information between entities and infer semantic relationships using entity-path information,a document-level relation extraction method based on mention graphs and explicit paths was proposed.A graph construction method based on entity mention types was employed,and the inherent structured information between mentions was integrated by combining graph attention networks to achieve the aggregation of information between entities.The relation inference method based on explicit paths was utilized,which included two sub-methods:explicit path construction method and heuristic path feature fusion method.These methods were employed to construct relation reasoning paths explicitly.The reasoning paths were categorized into intra-sentence reasoning paths,inter-sentence reasoning paths,and direct reasoning paths,aiming to differentiate between intra-sentence and inter-sentence reasoning.Differential fusion of path features was employed to enhance relationship path reasoning capabilities and improve the accuracy of relationship extraction.Comparative experiments conducted on three public datasets demonstrate the superiority of this method in terms of F_(1) and Ign F_(1) metrics compared to current mainstream methods.A document-level relation extraction method based on mention graphs and explicit paths could more effectively support the task of document-level relation extraction.