Abstract
Event detection is an important task in natural language processing, which identifies event trigger words in a given sentence. Previous work use traditional RNN/CNN based text encoders, failing to remember long-range dependencies within the sentence. This paper proposes a Transformer-based event detection model, utilizing the self-attention technique of a Transformer encoder to capture long-range dependencies. However, existing labelled event detection datasets such as ACE2005 are in small scale, being insufficient to train a deep neural model. This paper expands the scale of training data with an event trigger-aware back-translation based data augmentation technique. In addition, a multiinstance learning strategy is adopted to alleviate noises in the generated data in training. Experiments on two event detection datasets in different languages verify the effectiveness of the proposed model (The code is released at https://github.com/gitofding/event-detection).