Improving Topic Relevance Model by Mix-structured Summarization and LLM-based Data Augmentation