A Dual-Classifier Model for General Fine-Grained Event Detection Task

Yiming Hei, Qian Li, Caibo Zhou, Rui Sun, Jinglin Yang, Jiawei Sheng, Shu Guo, Lihong Wang

Abstract: This paper presents a winning solution for the CCKS-2021 general fine-grained event detection task whose goal is to identify event triggers and the corresponding event types from the MAssive eVENt detection dataset (MAVEN). In this task, we focus on two challenging problems in MAVEN: event identification and event confusion. The former problem is that it is hard to determine whether the current trigger word triggers an event. The latter problem means that some events are prone to category confusion. To solve the event identification issue, we propose a dual-classifier event detection model, which combines event identification and event classification to enhance the ability to judge the existence of events. In addition, to solve the problem of event confusion, we introduce adversarial training strategies to enhance the robustness of event category boundaries. The approach achieves an F1-score of 0.7058, ranking the first place in the competition.