Structured Information Matters: Incorporating Abstract Meaning Representation into LLMs for Improved Open-Domain Dialogue Evaluation