How do LLMs' Preferences Affect Event Argument Extraction? CAT: Addressing Preference Traps in Unsupervised EAE

Large Language Models (LLMs) have significantly improved the performance of unsupervised Event Argument Extraction (EAE) tasks. However, the prevalent preferences in LLMs severely hinder their effectiveness in EAE, leading to what we term preference traps, namely, the prior knowledge trap, the sycophancy hallucination trap, and the output contradiction trap. Existing methods often fall into these traps due to low-quality prior knowledge, ambiguous instructions, and contradictory outputs. To address this issue, we propose Choose-After-Think (CAT), an unsupervised EAE framework based on LLMs. CAT divides the task into two stages: identification of event information (think stage) and selection of arguments from a candidate argument set for template filling (choose stage). Both stages employ countermeasures to address these preference traps, while the choose stage's completely constrained outputs satisfy EAE's structured-output requirements. Experimental results demonstrate that CAT (based on the local 7B model, zero-shot setting) matches the performance of the best DeepSeek-R1 APIaccessible model, with a significantly lower time cost. 1