Event Detection via Derangement Reading Comprehension

Anonymous

Abstract: Event detection (ED), aiming to detect events from texts and categorize them, is vital to understanding the actual happenings in real life. 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 mechanism on a machine Reading Comprehension (DRC) framework to tackle the above challenges. More specially, we treat the input text as {\em Context} and concatenate it with all event types that are deemed as {\em Answers} with an omitted default question. Thus, by appending input text and event types simultaneously, we can facilitate the power of self-attention in pre-trained language models, e.g., BERT, to absorb the semantic relation among them. Moreover, we design a simple yet effective {\em derangement} mechanism to relieve the imbalanced training. By introducing such perturbation mainly on major events, we can prohibit major events from excessive learning or implicitly under-sample the instances of the major events. This yields a more balanced training to resolve the imbalanced learning issue. The empirical results show that: (1) our proposed framework attains state-of-the-art performance over previous competitive models, and (2) by-product, our model can signify the connection of triggers and arguments to events for further analysis.

Paper Type: long