Lexical and Compositional Stream Learning for Event Detection with Sememe Knowledge

As one of the most significant subtasks for event extraction, event detection(ED) aims to identify the trigger words in a sentence and classify them with correct event types. Most methods in previous work rely on various neural networks to extract trigger features automatically which still suffer a lot from word-trigger mismatch and disability of sparse triggers detecting, especially in Chinese corpus. In this paper, we propose a lexical and compositional stream learning approach to alleviate these two limitations in ED task with sememes in How Net as the external knowledge base. Concretely, we employ convolutional neural network (CNN) to learn lexical representation and compositional representation separately, and we consolidate event sememe information into structural features where the event sememe embeddings provide sememe trigger clues in sentence-level and word sememe-type tertiary structure enriches the compositional features. Then we fuse both of them into a hybrid representation to achieve trigger identification and event type classification. Experiments conducted on ACE2005 dataset show our model outperforms the state-of-the-art method especially for event type classifier.