MusiCRS
收藏资源简介:
MusiCRS是一个音频为中心的对话式推荐系统基准,它将来自Reddit的真实用户对话与相应的音频曲目链接起来。该数据集包含477个高质量对话,涵盖古典、嘻哈、电子、金属、流行、独立和爵士等多样音乐类型,涉及3589个独特的音乐实体,并通过YouTube链接进行音频关联。MusiCRS支持在三种输入模态配置下进行评估:仅音频、仅查询和音频+查询(多模态),从而可以系统地比较音频大语言模型、检索模型和传统方法。实验结果表明,当前系统严重依赖文本信号,难以进行细微的音频推理,揭示了跨模态知识集成中的基本局限性。为了促进研究进展,我们发布了MusiCRS数据集、评估代码和全面的基准测试。
MusiCRS is an audio-centric conversational recommendation system benchmark that pairs real user conversations sourced from Reddit with corresponding audio track links. This dataset contains 477 high-quality conversations spanning diverse music genres including classical, hip-hop, electronic, metal, pop, indie, and jazz, involving 3589 unique music entities, with audio associations provided via YouTube links. MusiCRS supports evaluation under three input modality configurations: audio-only, query-only, and audio+query (multimodal), enabling systematic comparison of audio large language models, retrieval models, and traditional methods. Experimental results show that current systems heavily rely on textual signals and struggle with fine-grained audio reasoning, revealing fundamental limitations in cross-modal knowledge integration. To facilitate research progress, we have released the MusiCRS dataset, evaluation code, and comprehensive benchmark tests.




