Event Detection via Derangement Question Answering

Anonymous

Abstract: Event detection (ED), aiming to detect events from texts and categorize them, is vital to understanding the messages. Recently, ED without triggers has been proposed and gained benefits since it relieves the tedious effort of data labeling. However, it still suffers from several formidable challenges: multi-label, insufficient clues, and imbalanced event types. We, therefore, propose a novel Derangement Question-Answering (DQA) framework on top of BERT to tackle the above challenges. More specially, we treat the input text as a {\em question} and directly concatenate it with all event types, who are deemed as {\em answers}. Thus, by utilizing the original information, we can facilitate the power of self-attention in BERT to absorb the semantic relation between the original input text and the event types. Moreover, we design a simple yet effective {\em derangement} mechanism to relieve the issue of imbalanced event types. By including such perturbation, we can train a more robust model to promote the semantic information in the major events while recording the position of the minor events than the vanilla QA framework. The empirical results show that: (1) our proposed DQA framework attains state-of-the-art performance over previous competitive models. (2) Our model can automatically link the triggers with the event types while signifying the corresponding arguments.