Event Causality Is Key to Computational Story Understanding

Anonymous

TL;DR: We propose a method for story event causality extraction with LLMs and demonstrate that the automatically extracted event causality facilitates computational story understanding.

Abstract: Cognitive science research and symbolic AI techniques suggest that event causality provides vital information for story understanding. However, machine learning systems for story understanding rarely employ event causality, partially due to the lack of methods that reliably identify open-world causal event relations. Leveraging recent progress in large language models (LLMs), we present the first method for event causality identification that leads to material improvements in computational story understanding. We verify the validity of identified causal relations against human-annotated datasets and find that our in-context-learning approach performs on par with or better than supervised models. Further, we test the identified causal relations in two downstream tasks, story quality evaluation and story video-text alignment. On the first task, the causal relations lead to a 3.6-16.6% relative improvement on correlation with human ratings. On the second task, we attain 4.1-10.9% increase on clip accuracy and 4.2-13.5% increase on sentence IoU. The findings indicate substantial untapped potential for event causality in computational story understanding.

Paper Type: long

Research Area: Discourse and Pragmatics

Contribution Types: NLP engineering experiment

Languages Studied: English