基于句子类别信息的事件检测方法  

Event Detection Method Based on Sentence Type Information

作　　者：郭恒睿 王中卿[1] 朱巧明[1] 李培峰[1] GUO Hengrui;WANG Zhongqing;ZHU Qiaoming;LI Peifeng(School of Computer Science and Technology,Soochow University,Suzhou,Jiangsu 215006,China)

机构地区：[1]苏州大学计算机科学与技术学院,江苏苏州215006

出　　处：《中文信息学报》2024年第10期117-126,共10页Journal of Chinese Information Processing

基　　金：国家自然科学基金(61836007,62006167,61806137);江苏高校优势学科建设工程资助项目。

摘　　要：目前大多数事件检测研究都认为句子之间是相互独立的,忽略了句子之间、触发词之间的事件关联。基于此,该文提出了一种基于句子类别信息的事件检测方法。首先通过基于BERT和注意力机制的句子级事件检测模型获取句子类别信息,构建句子之间的事件关联。然后该文通过基于机器阅读理解的事件检测模型融合句子类别信息,并学习句子标签的文本信息。此外,加入对比学习模块,拉近同类别触发词之间的距离,学习触发词之间的事件关联。最终实验表明,该模型在ACE中文和英文数据集上相比基准模型F_(1)值分别获得了1.6%和1.1%的提升。Most of the current event detection researches treat the sentences as independent of each other,ignoring the event correlation between sentences and triggers.In contrast,this paper proposes an event detection method based on the sentence type information.Firstly,we obtain the sentence type information through a sentence-level event detection model based on BERT and attention mechanism,and construct event correlation between sentences.Then,we design an event detection model based on machine reading comprehension to fuse sentence type information and learn the textual information of the sentence label.In addition,a contrastive learning module is added to shorten the trigger distance of the same type and learn the event association between triggers.The experiments in ACE Chinese and English datasets show that the proposed model have improved F_(1) value by 1.6%and 1.1%,respectively,compared with the benchmark models.

关 键 词：事件检测 句子类别信息 对比学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]