基于TLDA和SVSM的音乐信息检索模型  

Tags Know You Better:A New Approach to Enhancing MIR System

作　　者：周利娟[1] 林鸿飞[1] 闫俊[1] 

机构地区：[1]大连理工大学计算机科学与技术学院,大连116023

出　　处：《计算机科学》2014年第2期174-178,共5页Computer Science

基　　金：国家自然科学基金(60973068;61272370);国家社科基金(08BTQ025);教育部博士点基金(20110041110034);辽宁省自然科学基金(201202031)资助

摘　　要：随着协同标注功能的普及,用户可以通过标注自己感兴趣的音乐实现个性化的分类管理,因此音乐共享系统中的社会化标签已成为互联网的重要资源。为了提高音乐检索系统的效率,综合考虑了社会化标签的特性及其对音乐检索模型的影响,利用了TLDA方法来进行标签聚类以获取更多的语义相关的标签,综合考虑了用户检索行为、歌词、音乐标签和音乐流行度来提高音乐信息检索系统的性能。实验表明,基于TLDA和SVSM的音乐检索模型相比于基于属性数据的音乐检索模型以及k-means标签聚类的模型,尤其是在音乐标签稀疏和非正规的情况下,能够在一定程度上提高音乐检索的性能。Music sharing systems with collaboratively tagging function have been important parts on the Internet. They make the system users to annotate and categorize their own interests and thoughts about the resources possible. In the paper,a novel and straightforward way was proposed to search music collections using metadata and descriptions （tags） of tracks, by jointly considering lyrics, tags and popularity of songs to enhance Music Information Retrieval （MIR） system. Furthermore,Tag Latent Diriehlet Allocation （TLDA） model was proposed in the paper to facilitate adjusted VSM by obtaining more semantically related tags. TLDA can better analyze collaboratively generated tags and understand the intent of user queries in a semantic way, acquiring more information than just keyword-matched tracks return list. By comparing the performance of the proposed approach with general tag clustering approach, a result was found that mu- sic information retrieval model proposed in the article performs better than conventional metadata-based music retrieval techniques and tags clustering, especially when tags for tracks are extremely sparse and informal.

关 键 词：音乐信息检索 音乐向量空间模型 标签聚类 标签推荐 TLDA模型 

分 类 号：TP319[自动化与计算机技术—计算机软件与理论]