Empower event detection with bi-directional neural language model

摘要

Event detection is an essential and challenging task in Information Extraction (IE). Recent advances in neural networks make it possible to build reliable models without complicated feature engineering. However, data scarcity hinders their further performance. Moreover, training data has been underused since majority of labels in datasets are not event triggers and contribute very little to the training process. In this paper, we propose a novel multi-task learning framework to extract more general patterns from raw data and make better use of the training data. Specifically, we present two paradigms to incorporate neural language model into event detection model on both word and character levels: (1) we use the features extracted by language model as an additional input to event detection model. (2) We use a hard parameter sharing approach between language model and event detection model. The extensive experiments demonstrate the benefits of the proposed multi-task learning framework for event detection. Compared to the previous methods, our method does not rely on any additional supervision but still beats the majority of them and achieves a competitive performance on the ACE 2005 benchmark.