Chinese event recognition via ensemble model

Abstract

Event recognition is one of the most fundamental and critical field in information extraction. In this paper, Event recognition task can be divided into two sub-problems containing candidate event triggers identification and the classification of candidate event trigger words. Firstly, we use trigger vocabulary generated by trigger expansion to identify candidate event trigger, and then input sequences are generated according to the following three features: word embedding, POS (part of speech) and DP (dependency parsing). Finally multiclass classifier based on joint neural networks is introduced in the step of candidate trigger classification. The experiments in CEC (Chinese Emergency Corpus) have shown the superiority of our proposal model with a maximum F-measure of 80.55%.