MSK-Net: Multi-source Knowledge Base Enhanced Networks for Script Event Prediction

Abstract

Script event prediction (SEP) aims to choose a correct subsequent event from a candidate list, according to a chain of ordered context events. It is easy for human but difficult for machine to perform such event reasoning. The reason is that human have relevant commonsense knowledge. If we supplement this knowledge from external knowledge bases, machine may be able to improve the reasoning ability. To this end, we introduce a novel approach, named MSK-Net, which consists of Question Encoder, Knowledge Searcher, Knowledge Encoder and Result Predictor. As far as we know, this is the first model utilizing multi-source knowledge to solve SEP problem. Specifically, first we use Question Encoder to encode the question, including candidate event to be judged and context events, focusing on intra-event contextualization and inter-event order information modelling. Then, we use Knowledge Searcher to retrieve relevant knowledge from multi-source knowledge bases (such as ASER and ATOMIC). Third, Knowledge Encoder is used to encode the knowledge retrieved in the second step. Last, Result Predictor gives the final prediction. Experiments on the widely-used multiple choice narrative cloze (MCNC) task demonstrate our approach achieves state-of-the-art performance compared to other methods. Also, it is worth noting that MSK-Net without external knowledge is still very competitive.