Probing Representations for Document-level Event Extraction