LLM-Driven Multimodal Opinion Expression Identification