Dependency Multi-Weight-view Graphs for Event Detection with Label Co-Occurrence

Yan Wang,Jian Wang,Hongfei Lin,Yijia Zhang,Zhihao Yang

DOI: https://doi.org/10.1016/j.ins.2022.05.029

IF: 8.1

2022-01-01

Information Sciences

Abstract:• A multi-weight-view graph is proposed to capture enriched semantic and structural information automatically. • A label co-occurrence graph is proposed to enhance dependency between different event labels. • DMWVG is an end-to-end model without external tools, knowledge, or feature engineering. • DMWVG model detects events without identifying trigger words. • DMWVG model obtains remarkable performance on MLEE, BioNLP 09, BioNLP 11, and ACE 2005 datasets. Event detection is an extremely important step for extracting events, and its goal is to detect the occurrences of events and categorize them. Prior works use either a fine-grained approach or a coarse-grained approach to detect events. However, both approaches encounter two problems: (1) Enriched feature information obtained from external tools probably introduces unnecessary errors. (2) Ignoring the phenomenon of event co-occurrence in a sentence degrades the performance of event detection. To address these issues, a novel framework called D ependency M ulti- W eight- V iew G raph (DMWVG) is proposed to detect events without trigger tagging and external feature engineering. Specifically, DMWVG leverages the tree transformer to automatically generate syntactic relations with different weights to mitigate the errors generated by external syntactic parsing tools. In addition, an event correlation matrix is constructed with the help of the event label co-occurrence to deliver essential information on triggers and to promote the performance of event detection. Finally, the experimental results indicate that our approach is competitive with existing SOTA methods on four datasets (75.77% F1-score on ACE 2005, 82.11% F1-score on MLEE, 75.61% F1-score on BioNLP 09, and 72.20% F1-score on BioNLP 11).