Empowering document-level relation extraction with efficient evidence extraction and inference-stage fusion