TsEDM: A two-steps event detection model for Chinese event detection

Abstract

Event detection (ED) consists of two phases – trigger identification (TI) and trigger classification (TC). Traditional ED adopts a unified model to process the above two-stage tasks at once. We argue that there are certain differences in the contextual semantics required and the goals of these two phases in ED. In which, TI remains suffers from the word-trigger mismatch problems in languages without natural word delimiters such as Chinese. And the TC is facing challenging problems of trigger ambiguity and multiple triggers in a sentence. In this article, we propose a brand-new two-steps event detection model (TsEDM), which attempts to alleviate above-mentioned problems. Specifically, a novel ‘head-tail dual-pointer’ (HT-DP) labelling strategy is developed to obtain more candidate triggers to overcome the problems of continuous labelling, nested labelling and independent labelling in the first step (TI). Besides, an ‘entity–topic–candidate–trigger’ interaction graph (E2T-IG) is constructed in the second step (TC) to consider the interaction relationship between candidate triggers and core information inter or in all event sentences, which enhance the representation of each candidate trigger. Last but not least, a shake-gated and residual-based atrous convolution neural network (SGR-ACNN) is proposed as the common framework of these two steps, which dynamically integrates various representations as model inputs. Experiments on the ACE2005-CN show that TsEDM significantly outperforms state-of-the-art (SOTA) methods.