Context-aware Event Type Identification Based on Context Fusion and Joint Learning

Automatic Event Type Identification from a text document is a particularly challenging task in event extraction. Consideration of the event triggers' context has been shown effective in this task. However, existing methods suffer from insufficient consideration of the semantic information in the event trigger context and the dependencies between event triggers. To fill this gap, we propose a novel joint learning model called CAED for event type identification. Given a text document, first, CAED carries out event context fusion by sending the sentence embeddings to BiLSTMs for capturing the weighted average of its hidden states and aggregates them using the self-attention mechanism to obtain the document-level context embedding. Then, CAED introduces a dynamic memory vector to record the occurrences of event triggers and their dependencies in each sentence. Finally, CAED concatenates the documentlevel context embedding, the dynamic memory vector, and the pre-trained word embedding into a joint word-level embedding. CAED carries out joint learning by sending the joint word-level embeddings to a LSTM based event type classifier that iteratively uses the dynamic memory vector to achieve Context-aware Event Type Identification. Experiment results on the CEC benchmark dataset and a case study demonstrate the superiority of CAED over six state-of-the-art event type identification models.