Revealing the Capabilities of Large Language Models on Event Generalization

Event generalization is crucial in understanding the logical information with specific events. Current success of Large Language Models (LLMs) on other event-centric NLP tasks has indicated their potential on event generalization. However, due to the lack of open-sourced dataset, the capabilities of LLMs still remain unexplored. In this work, we develop the first open-sourced dataset CEG for event generalization, which is meticulously annotated to cover a diverse range of events. We then leverage CEG to systematically evaluate the zero-shot and few-shot in-context learning abilities of both state-of-the-art open-sourced and closed-sourced LLMs on event generalization tasks. The results show that LLMs demonstrate robust event generalization capabilities, which can be further enhanced as the number of few-shot examples increases.