Reasoning subevent relation over heterogeneous event graph

Tingting Wu, Xiao Ding, Li Du, Bing Qin, Ting Liu

Abstract: Subevent relation identification (SRI) is a challenging natural language processing task of great value for knowledge acquisition and reasoning. Given an event pair, previous work mainly defines SRI as a classification task and usually relies on hand-crafted features extracted from a limited context. However, there are lots of evidence information in external knowledge bases which are not well explored by prior work but helpful for this task. To fill this gap, we propose to reason the subevent relation based on plentiful evidence information in addition to limited context features. Specifically, we notice that evidence information together with the input event pair can be organized into a heterogeneous event graph, which comprises various types of event relations. To facilitate this, we present a heterogeneous graph attention model with mix-hop reasoning mechanism to reason subevent relation over heterogeneous event graph. In particular, our model allows nodes to capture more extensive and intact event knowledge by mixing feature representations of neighbors at multiple distances to aggregate information from diverse types of neighbors. Moreover, to explicitly model the hierarchical relations between event pairs and to improve the model consistency, we devise a novel hierarchy loss function for the proposed model. Our model is evaluated on two annotated datasets with subevent hierarchies: SeRI and HiEve. Experimental results show that our approach can outperform state-of-the-art baseline methods. Further evaluation suggests that reasoning paths in the heterogeneous graph can regard as a reasonable explanation for the prediction result.