Improved conditional random fields model with multi-trigger embedding for Chinese event extraction

Abstract

Event extraction is a challenging task in natural language understanding, which aims to recognize event type, subtype and roles of relevant entities from unstructured text. Most current approaches address event extraction with highly local models that extract event type and arguments independently. However, this multi-step method cannot make full use of the reciprocal dependency relationship between event trigger and arguments, especially for nested event structure. E.g. the trigger of a Life/Injure event is embedded inside the argument, and a trigger is an event anchor as well as a modifier of argument. Meanwhile, In the same label space, the example proportion of triggers to arguments is scare, and there exists the issue of unbalanced data. Therefore, this kind of trigger is apt to be labeled as the event argument. In order to let event type recognition and event argument recognition guide each other, and resolve the problem of unbalanced data, we consider event extraction to be a sequence labeling problem and build a novel improved conditional random fields joint labeling model with multi-trigger embedding. Experimental results on ACE 2005 Chinese corpus show that this method improves the performance of event extraction.