融合知识的多目标词联合框架语义分析模型

Knowledge-integrated Joint Model for Multi-target Frame Semantic Parsing

框架语义分析任务是自然语言处理领域的一项基础性任务。先前的研究工作大多针对单目标词进行模型设计,无法一次性完成多个目标词的框架语义结构提取。该文提出一个面向多目标的框架语义分析模型,实现对多目标词的联合预测。该模型对框架语义分析的各项子任务进行交互性建模,实现子任务间的双向交互。此外,该文利用关系图网络对框架关系信息进行编码,将其作为框架语义学知识融入模型中。实验表明,该文模型在不借助额外语料的情况下相比之前模型都有不同程度的提高。消融实验证明了该文模型设计的有效性。此外,该文分析了模型目前存在的局限性以及未来的改进方向。

Frame semantic parsing is a fundamental task in natural language processing.Most of the previous researches are focused on the design of the single-target model.This paper designs a frame semantic parsing model for multiple targets which jointly predicts the results of multiple targets.We model and achieve the bidirectional interaction among the subtasks of frame semantic parsing.Moreover,Relational Graph Convolution Network(R-GCN)is utilized to encode the frame relation information,which is a way to exploit frame semantic knowledge into our model.The experiments show that our model maintains good performance without extra training corpus.