Enhance Incomplete Utterance Restoration by Joint Learning Token Extraction and Text Generation