CascadePAIE: Reallocating Relevance for Event Roles and Event Text in Event Argument Extraction

Abstract

Event argument extraction is one of the subtasks in event extraction, and the current mainstream approaches define it as a span extraction task utilizing the concept of prompt learning. However, upon reproducing the current mainstream work, we identified two issues with event argument span extraction: 1) the predicted spans for arguments lack precision and may encompass positions near the ground-truth spans, and 2) different roles identify the same argument span. These issues somewhat impact the model’s performance. To address these challenges, we propose CascadePAIE. On the one hand, this method reallocates relevance for event text by mapping the relevance to event representations and the loss function, thereby enhancing event representations and penalizing erroneous predictions near the ground-truth spans. On the other hand, it reallocates attention for event roles, assigning lower attention to subsequent roles for spans already predicted by preceding roles, thereby surpassing the limitations of those roles. We conducted extensive experiments on sentence-level and document-level datasets. The results validate the effectiveness of our approach, with average improvements of 1.63%, 1.84%, and 2.28% on the three datasets in the base version, and even achieving improvements of 2.95%, 2.56% and 3.02% in the best experimental results.