Document-level Event Extraction Based on Fusion of Understanding and Generation
CAO Kaichen
GAO Dongsheng
Abstract：Document-level event extraction aims at extracting key events and elements from long documents based on the in-coming documents to which they belong,which is the key to automatically acquiring knowledge from the domain text corpus.Most existing document-level event extraction systems are based on text understanding and extraction and are prone to the problem of multi-event confusion,while the latest generation-based event extraction systems are prone to phantom errors.To address these problems,a document-level event extraction model that fuses comprehension and generation is proposed.The model is based on an encoder-decoder architecture that utilizes a bidirectional encoder for document understanding and fragment prediction-based event element extraction,and the model also utilizes a unidirectional decoder for event element generation with the aid of an input-specif-ic cueing framework.Contrast loss learning is proposed to allow the extraction-based encoder part and the generation-based decod-er part to learn from each other,thus mutually improving the performance level of event element extraction and eventually achieving a unified output of comprehension and generation through a fused output approach.Experiments show that the approach has a large performance improvement over traditional methods on both typical Chinese and English datasets.
Keywords：document-level event extractionencoder-decoder architecturefusion of understanding and generationcon-trast loss learning
Publication Date：2023-10-20
Online Publishing Date：2025-08-15(First online date of this platform, not the publication date of the document)
Pages：6( 47-52 )