对文本中诸如实体与关系、事件及其论元等要素及其特定关系的联合抽取是自然语言处理的一项关键任务.现有研究大多采用统一编码或参数共享的方式隐性处理任务间的交互,缺乏对任务之间特定关系的显式建模,从而限制模型充分利用任务间的关联信息并影响任务间的有效协同.为此,提出了一种基于任务协作表示增强的要素及关系联合抽取模型(Task-Collaboration Representation Enhanced model for joint extraction of elements and relation-ships,TCRE).该模型旨在从多个阶段处理任务间的特定关系,帮助子任务进行更细致的调节和优化,促进整体性能的提升.在三个关系抽取和一个事件抽取数据集上进行实验,TCRE在实体识别和关系提取任务上平均性能分别提高0.57％和0.77％,在触发词识别和论元角色分类任务上分别提高0.7％和1.4％.此外,TCRE还显示出在缓解"跷跷板现象"方面的作用.
Abstract
Jointly extracting elements like entities and their relationships,as well as events and their arguments,is a crucial natural language processing task. Current methods,primarily based on unified coding or parameter sharing,fail to explicitly model inter-task relationships. This limitation restricts the use of inter-task correlations and hinders effective col-laboration. To address this,we propose a task-collaboration representation enhanced model for joint extraction of elements and relationships (TCRE). TCRE strategically captures and leverages specific inter-task relationship representations across multiple stages,facilitating precise tuning and optimization of subtasks,thereby enhancing overall model performance. In evaluations on three relation extraction and one event extraction datasets,TCRE demonstrated performance improvements of 0.57％ in entity recognition,0.77％ in relation extraction,0.7％ in trigger word recognition,and 1.4％ in argument role classification. Additionally,TCRE effectively mitigates the "seesaw phenomenon".
维普
万方数据