A Practical Approach to Chinese Emergency Event Extraction using BiLSTM-CRF

Abstract

For enhancing situation awareness in the Internet Era, it is of great significance to automatically extract emergency events from online news reports. However, existing approaches to this task are mainly rule- or pattern-based, and hence suffer from poor generalization. Though recent researches on event extraction (EE) have explored deep neural models (DNM) and achieve impressing results, they are mostly tested using the benchmark dataset ACE2005, which contains no emergency events, and not well suited for practical applications as take as input the entity labels of candidate arguments provided by that dataset. Also, these methods ignore the inter-dependency between triggers and arguments, which could be useful for EE. Hence, this article proposes a DNM for Chinese emergency EE, which uses BiLSTM (bidirectional Long Short Term Memory), pretrains domain-specific word embeddings, and adopts CRF (Conditional Random Field) to capture the interplay between triggers and arguments. Extensive experiments, conducted on the Chinese Emergency Corpus, prove the effectiveness of the proposed model, and show that it outperforms other state-of-the-art methods substantially.