Central-Guided Convolutional Dual Attention for Document-Level Event Argument Extraction

Abstract

Document-level event argument extraction is an important task in the field of natural language processing. The current mainstream document-level event argument extraction methods ignore two key points: the central information of the document and the cross-sentence dependency. To overcome these limitations, we propose a new document-level event argument extraction model VCECDA (Vector Content Expansion and Convolution-based Dual Attention), which consists of two innovative modules: the Vector Content Expansion (VCE) and the Convolution-based Dual Attention (CDA). The VCE module extracts the central information of the document and expands its content, combining deep semantic and structured information to effectively extract global context features. The CDA module optimizes the expression ability of features from two aspects: channel category relationship and spatial position relationship, effectively capturing local contextual information and global dependency of features in the text. We conduct a large number of experiments on RAMS and WikiEvents datasets. The experimental results show that VCECDA achieves optimal performance on WikiEvents dataset compared with the baseline, and also performs competitively on RAMS dataset.