Beyond Exact Match: Semantically Reassessing Event Extraction by Large Language Models