What Would Happen Next? Predicting Consequences from An Event Causality Graph

Anonymous

Abstract: Predicting the consequences based on some past events has a huge potential in various Natural Language Processing applications. However, existing work faces two shortcomings: (1) Simple modeling scenarios, such as Script Event Prediction task, which predict subsequent events only based on an event chain; (2) Interpolation scenarios, such as Event Knowledge Graph Completion task, where the predicted event has already occurred in known events. In this paper, we propose a new task named Event Causality Graph Prediction, which forecast the consequence event based on an event causality graph constructed from a document describing complex event scenarios. To that end, we propose two corresponding datasets and an \textbf{G}raph \textbf{C}ontrastive \textbf{P}rompt \textbf{L}earning model(GCPL), which utilize the benefits of graph prompt learning and introduce the Dual Encoder to integrate node text and graph structure information. We conduct extensive experiments on two datasets and our GCPL achieves state-of-the-art performance among all competitors.

Paper Type: long

Research Area: Discourse and Pragmatics

Contribution Types: Model analysis & interpretability

Languages Studied: English