Learning Context-Specific Word/Character Embeddings