Nested Causality Extraction on Traffic Accident Texts as Question Answering

Abstract

As an important type of relationship, causality plays a vital role in relation reasoning. Therefore, causality extraction from natural language texts is a crucial task, especially in the field of traffic. For example, we can quickly discover the causes of traffic accidents and the correlation between the cause events. Existing methods utilize machine learning models to extract simple causality, however, there is nested causality in the traffic accident sentences, which is important for us to reason the cause and effect. In order to extract the causality successfully, we simplify the complex nested causality structure to the pairwise causality structure. On this basis, we propose a method that contains two steps. First, we extract the cause events from the input sentence, second, combine the extracted cause with the incomplete question template to obtain a complete question sentence, then we adopt the way of question answering tasks to extract the effect events. Experiments on the traffic accident dataset show the effectiveness of our model. However, we observe that due to the small training set, there is still room for improvement in the extraction accuracy of nested causality.