Enhancing Event Argument Extraction: A Span-level Representation Approach

In the domain of event extraction, the process of identifying and categorizing event arguments and their corresponding roles is a critical subtask. Current methodologies for event argument extraction are predicated on the classification of individual argument roles based on a single word, ignoring the form of argument phrases. However, words in a phrase are prone to be mistaken when the classification is based on word level. To address this issue, we propose a novel event argument extraction model that employs span-level representation to avoid false classification occurring on some words in the same phrase. We achieve this by using span-level information instead of word-level information. Specifically, we first use a bidirectional Gated Recurrent Unit to effectively encode input with context and predict the positions of all arguments in the sentence. Subsequently, an attention mechanism is applied to integrate multiple words under the same argument phrase, considering their different importance, and generate a high-quality span-level representation. Experiments conducted on the ACE2005 dataset demonstrates that the our model can effectively leverage span-level information from argument phrases.