Commonsense Knowledge Enhanced Event Graph Representation Learning for Script Event Prediction

Abstract

Script event prediction plays an important role in many artificial intelligence applications. A key challenge in this task is accurately understanding the correlation between events and then inferring the subsequent events. Benefiting from exploring the rich connections among events, recently some event graph-based methods have achieved remarkable success. However, limited by the abstract representation of events, it is often difficult to derive the event relationships fully and precisely from the original event graphs. In this paper, we propose a novel framework, called Commonsense knowledge-enhanced event evolutionary graph (CEEG), to remedy this problem. CEEG constructs two core modules, event evolutionary graph and event commonsense graph, the former incorporates connections among the predicate events and object events to find the potential information among events, while the latter utilizes external commonsense knowledge to further enhance the understanding of event relationships. Pre-trained model RoBERTa and Graph Neural Network are also integrated into the framework to obtain more accuracy event nodes depiction and rich event graph information. Comprehensive experimental results on the benchmark dataset show the effectiveness of the proposed framework.