Joint Prior Relation Enhancement and Non-autoregressive Decoding for Document-Level Event Extraction

Document-level event extraction (DEE) aims to extract all structured event information from a given document. This research currently faces three main challenges: (1) Event arguments often scatter across multiple sentences; (2) A document typically comprises multiple interrelated events; (3) Traditional autoregressive decoding extraction methods result in exceedingly slow extraction speeds. To address these challenges, this paper proposes a Document-Level Event Extraction method by Joint Prior Relation Enhancement and Non-Autoregressive Decoding (PRENA). Specifically, we propose a Prior Relationship Enhancement Network (PREN) to explicitly model multiple prior dependencies between entity mentions and sentences, effectively capturing global interactions between cross-sentence arguments and sentences, and modeling dependencies between multiple correlated events. In addition, we propose a non-autoregressive decoding algorithm based on an optimized pseudo triggers selection strategy, which efficiently extracts multiple event argument combinations in a joint parallel manner, significantly improving the efficiency of model extraction. Comparative experiments validate the outstanding performance of the proposed PRENA in enhancing both the accuracy and efficiency of event extraction.