基于联合层级问答的新闻人物言论抽取方法  

A joint hierarchical question-answering based approach forextracting remarks from news

作　　者：张能欢 冯爽[2] 周正宇 ZHANG Nenghuan;FENG Shuang;ZHOU Zhengyu(Academy of Broadcasting Science,National Radio and Television Administration,Beijing 100866,China;Communication University of China,Beijing 100024,China)

机构地区：[1]国家广播电视总局广播电视科学研究院,北京100866 [2]中国传媒大学,北京100024

出　　处：《中国传媒大学学报（自然科学版）》2024年第3期57-65,共9页Journal of Communication University of China：Science and Technology

基　　金：中国传媒大学校级项目(CUC23WH005);国家重点研发计划课题(2021YFF0901602)。

摘　　要：新闻人物言论的抽取对理解社会动态、分析公众意见、辅助决策制定至关重要。本文针对新闻场景定义了人物言论事件框架,提出了一种基于联合层级问答的新闻人物言论抽取方法。该方法改变原有的针对起始位置进行预测的指针标注方式和序列标注,采用联合层级标注,以提高模型在面对复杂人物名称时的识别能力,并利用多轮问答的特性,准确匹配人物名称和触发词,以解决复杂人物言论抽取中的多任务问题。为进一步提升模型训练效果,引入了额外位置向量和交替双向训练策略。实验结果表明,本文方法在(人物,触发词)抽取、言论抽取以及(人物、触发词、言论)三元组抽取任务中均取得了较好的结果。The extraction of remarks made by news figures is crucial for understanding social dynamics,analyzing public opinion,and assisting decision-making.First,we defined an event framework for remarks in news scenarios.Then we proposed a novel method for extracting remarks made by news figures based on joint hierarchical question-answering.Departing from conventional pointer annotation and sequence labeling that predict starting positions,our method adopts joint hierarchical annotation to improve the model's ability to identify complex figure names.It also utilizes the characteristics of multi-round question-answering to accurately match figure names and trigger words,addressing the multitasking problem in extracting remarks made by complex figures.To further enhance the effectiveness of model training,this paper introduces extra positional embedding and Bi-Directional training frame.Experimental results show that the proposed method achieves good results in extracting(figure,trigger word)pairs,extracting remarks,and extracting(figure,trigger word,remark)triplets.

关 键 词：人物言论 事件抽取 联合层级标注 多轮问答 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]