Global information enhancement and subgraph-level weakly contrastive learning for lightweight weakly supervised document-level event extraction

Abstract

Document-level event extraction (DEE) is a branch of information extraction task that is valuable but difficult at present. To effectively implement DEE, previous works with better performance are often accompanied by larger parameters and computation. In this paper, we propose a novel framework based on global information enhancement and subgraph-level weakly contrastive learning, which bring better performance at lower extra computing costs. We verify the effectiveness of our method in a lightweight DEE model. The experimental results indicate that our proposed method outperforms other approaches in terms of the F1 score for both overall event extraction evaluation and single event extraction. In the inference phase, our model uses only 25% of the GPU memory required by the optimal model and maintains the advantage of 6.4 times faster inference speed.