Prompt-Based Event Coreference Resolution on Multi-dimensional Information and Auxiliary Tasks

Event Coreference Resolution (ECR) aims to group same event mentions into a cluster. Previous work on capturing multi-dimensional event information focused on rule matching or feature fusion, and this makes ability of model mainly relies in encoding instead of information chain. Moreover, previous work utilizing type and argument information often encoded or rigidly matched the respective information independently without establishing their connections. To address the above two issues, we propose a MLM (masked language model)-based method that progressively injects trigger explanations, types, arguments, realis, and external knowledge into the prompt mechanism, where event types and arguments are treated as event descriptions in the form of natural language. Furthermore, we introduce three prompt auxiliary tasks—realis predicting, trigger explanation matching and details matching, to assist the model in learning events from the perspective of trigger explanation and event details, which play the role of soft constraints. Experimental results on the KBP datasets show that our proposed method achieves the SOTA performance in both annotated event coreference resolution and end-to-end event coreference resolution.