Story Understanding with External Knowledge Based Attention

How we use commonsense knowledge to guide the learning of the model is critical for natural language understanding. In this paper, we introduce a keywords extraction weight based word-level attention and a sentiment analysis based sentence-level attention network so as to achieve a better understanding of the story. We train on the publicly available ROC story cloze task data. We mainly focus on two aspects of the narrative: topic correlation and sentiment polar. Since the train data set contains a lot of distractor, we train our model on the validation set and achieve the state of art accuracy 80.5%. We mainly provide a way to use external knowledge to guide training process.