Hybrid Generative and Commonsense Knowledge for Script Event Prediction
Abstract: Script event prediction aims to predict subsequent events given contextual events, which requires inferring correlations between contexts and candidate events. Current research focuses on improving script event prediction using external knowledge and pre-trained language models, but faces the problems of sparse event-level correlation knowledge and separation of word-level correlation knowledge. In this paper, we propose a novel model CoGen-Predictor based on hybrid generative and commonsense knowledge that combines explicit event-level and implicit word-level correlation knowledge for prediction. CoGen-Predictor constructs event-level correlations through a commonsense knowledge base and updates the event representations using graph neural networks, then learns word-level contextual event correlations through a generative approach. Experimental results on the multi-choice narrative cloze (MCNC) task demonstrate the effectiveness of the model.
Paper Type: Long
Research Area: Special Theme (conference specific)
Research Area Keywords: graph-based methods, knowledge-augmented methods, generative models
Contribution Types: Model analysis & interpretability
Languages Studied: English
Submission Number: 4666
Loading