A neural model for joint event detection and prediction

Event prediction aims to predict the future possible event given a sequence of previously happened events. Event prediction is important since it can benefit the government, agencies and companies for avoiding damages by taking proactive actions. A further related task is event detection, which is to classify each event to predefined types, helping users quickly find relevant information. Event prediction is related to event detection, since salient information of events is universal between the tasks. In this paper, we propose a novel neural model for joint event detection and prediction, which classifies the events to predefined types as well as predicts the next probable event by generating a sequence of words describing it. In addition, we propose a hierarchical attention mechanism to enable the model to capture important information at both word level and event level for next event prediction. Empirical experiments on a real-world dataset reveal that our joint model with hierarchical attention achieves substantial improvements on event prediction, advancing state-of-the-art models. With joint learning, our model also improves the performance on event detection.