Document-Level Event Role Filler Extraction using Multi-Granularity Contextualized Encoding