Table-Based Two-Stage Relation Classification Method for Trigger-Free Document-Level Event Extraction

Table-Based Two-Stage Relation Classification Method for Trigger-Free Document-Level Event Extraction

연구 분야: Software Development

학회: International Conference on Neural Information Processing

Document-level event extraction without triggers is a significant challenge compared to traditional event extraction. Documents often contain multiple event records with several arguments and their roles of an event are generally dispersed throughout sentences, making accurate extraction of all event records challenging. Moreover, the absence of annotated event triggers complicates the detection of event records, further increasing the difficulty of the task. Previous approaches typically employ a pipeline approach, sequentially performing entity extraction, event detection, and argument classification, which leads to error propagation due to performance limitations in the subtasks. Thus, we propose a Table-Based Two-Stage Relation Classification Method for Trigger-Free Document-Level Event Extraction (T3DEE) in this paper, where the first stage is focused on argument recognition through word-pair relation classification and the second stage is dedicated to generating event records through argument-pair relation classification. Our method T3DEE does not require prior entity extraction or event count/type detection in the document, as it directly identifies arguments and generates events through a two-stage process, thereby reducing overall task complexity. We carry out trials utilizing a broadly applied document-level dataset, yielding outcomes that showcase the proficiency of our approach, achieving new SOTA performance.