Leveraging ensemble deep models and llm for visual polysemy and word sense disambiguation Academic Article