Document-Level Event Causality Extraction via Automatic Prompt Engineering on Large Language Models

Document-level event causality extraction(DECE) aims at identifying and extracting event pairs with causal relations from a given document. Recently, Large language models (LLMs) applied with prompt engineering techniques demonstrate remarkable capabilities in comprehending and deducing complex logical relations within text. However, due to the lack of flexibility and adaptability in static prompts, LLMs struggle to fully capture implicit causal event pairs. To address these issues, this paper proposes a novel CausalAPE model. Specifically, the model employs automatic prompt engineering (APE) techniques to iteratively optimize prompts, thereby enhancing prompt quality to better extract causal event pairs. And it identifies previously unrecognized results or causes corresponding to already identified causes or results, which improves the recognition of one-to-many and m any-to-one causal relations. Extensive experiments were conducted on the EventStoryLine and Causal- TimeBank datasets to validate this approach. The results achieve 11.23% improvement in F1 compared to baseline methods and present the effectiveness of this framework.