论证(复杂分析)
自动汇总
计算机科学
约束(计算机辅助设计)
事件(粒子物理)
树(集合论)
人工智能
数学
组合数学
物理
量子力学
生物化学
化学
几何学
作者
Ji Zhou,Kai Shuang,Qiwei Wang,Xuyang Yao
标识
DOI:10.1016/j.ipm.2023.103559
摘要
There are two key challenges remaining for the document-level event argument extraction tasks: long-range dependency and same-role argument assignment. The existing methods could not effectively handle the above two challenges at the same time, resulting in argument misidentification and over- or under-extraction, reducing the precision and recall of event argument extraction. In this paper, we propose a document-level event argument extraction model with argument constraint enhancement (EACE), which constructs the argument constraint tree using the hierarchical constraints between arguments to address the above two challenges simultaneously. Specifically, EACE first constructs an argument constraint decoder and uses abstractive summarization to establish the long-range hierarchical constraint relationships between arguments and to obtain the trunk structure of the argument constraint tree, which improves argument identification precision. Secondly, EACE calculates dynamic branch thresholds to expand the branch structure of the argument constraint tree and improve the recall of argument extraction. Extensive experiments on WIKIEVENTS and RAMS have shown that EACE outperforms the baseline models by 2.2% F1 and 0.2% F1, respectively. Moreover, it exceeds the baseline model (PAIE) by up to 17.1% F1 in the same-role argument assignment setting in WIKIEVENTS.
科研通智能强力驱动
Strongly Powered by AbleSci AI