DRAM: A Deep Reinforced Intra-attentive Model for Event Prediction

Abstract

We address the problem of event prediction which aims to predict next probable event given a sequence of previous historical events. Event prediction is meaningful and important for the government, agencies and companies to take proactive actions to avoid damages. By acquiring knowledge from large-scale news series which record sequences of real-world events, we are expected to learn from the past and see into the future. Most existing works focus on predicting known events from a given candidate set, instead of devoting to more realistic unknown event prediction. In this paper, we propose a novel deep reinforced intra-attentive model, named DRAM, for unknown event prediction, by automatically generating the text description of the next probable unknown event. Specifically, DRAM designs a novel hierarchical intra-attention mechanism to take care not only the previous events but also those words describing the events. In addition, DRAM combines standard supervised word prediction and reinforcement learning in model training, allowing it to directly optimize the non-differentiable BLEU score tracking human evaluation and generate higher quality of events. Extensive experiments on real-world datasets demonstrate that our model significantly outperforms state-of-the-art methods.