DPMN:面向重叠关系抽取问题的多任务学习网络  

DPMN:Multi-Task Learning Network for Problem of Overlapping Relation Extraction

作　　者：李雅杰 唐国根 李平[1] LI Yajie;TANG Guogen;LI Ping(College of Computer Science,Southwest Petroleum University,Chengdu 610500,China)

机构地区：[1]西南石油大学计算机科学学院,成都610500

出　　处：《计算机工程与应用》2024年第20期160-167,共8页Computer Engineering and Applications

基　　金：国家自然科学基金(62276099);四川省自然科学基金面上项目(2023NSFSC0501)。

摘　　要：作为自然语言处理的基本组件之一,关系抽取旨在从给定的非结构化文本中抽取关系事实。针对实际应用中句子级的实体信息缺失,往往会出现单个句子包含多个重叠关系三元组的场景,关系三元组会产生多种交叉重叠,使得关系抽取任务具有较大的挑战性。早期的研究采用流水线方法来处理,不仅忽略实体识别和关系预测的关联性,而且容易受到误差传播问题的影响。提出了一种基于依赖解析的多任务学习网络(DPMN),以依赖解析的方式更精确地识别实体跨度、丰富关系语义,具有多任务学习策略,以增强各个子任务之间的交互。与基线模型相比,DPMN具有更好的关系三元组抽取性能,一定程度上缓解了关系重叠问题。As one of the basic components of natural language processing,relation extraction aims to extract relation facts from a given unstructured text.In practical applications,there is a lack of entity information at the sentence level,and there are often scenarios where a single sentence contains multiple overlapping relation triplets.Relation triplets can generate multiple cross overlaps,making relation extraction tasks more challenging.Early research uses pipeline method to process,which not only ignores the relevance of entity recognition and relationship prediction,but also is vulnerable to propagation of uncertainty.This paper proposes a multi-task learning network(DPMN)based on dependency parsing,which can identify entity span more accurately by dependency parsing,enrich relation semantics,and have multi-task learning strategies to enhance the interaction between various subtasks.Compared with the baseline model,DPMN has better performance in relation triplet extraction,which alleviates the problem of overlapping relations to some extent.

关 键 词：关系抽取 多任务学习 依赖解析 重叠关系 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]