Research Of Emotion Tag Based Music Retrieval Algorithm

With the booming of digital music on the Internet, more and more people search music on internet, so people pay more and more attention on the research of music infromation retrieval. Traditional MIR (music information retrieval) system which is based on the query of song's name, singer and lyric cannot meet people's various needs. Content based music retrieval technology usually consider music physical characteristics such as rhythm, tone, pitch and intensity, it is too specialized to be accepted. As there is a close connection between music and emotion, the idea of emotion based music retrieval is proposed. Emotion based MIR system allows users to search music via imputing emotional words, which satisfies the individual needs.As the prevalence of Web 2.0, web users are bestowed more freedom of motion to define music tags according to their own understanding of music, which denote users'distinguishing comprehension of music. Music is to express emotions and interpreted by tags, so the tags annotated by users inevitably involve emotion tags that reflect joviality or sorrow. In this paper, we study the novel problem of integrating emotion tags into music retrieval. Two novel iterative algorithms are proposed to improve the performance on MIR:(1) T_SimRank algorithm, which is to compute latent semantic similarity between tags and music based on a Bipartite graph of tags and music. (2) T_PageRank algorithm, which is to measure the popularity of music using tags. In the experimention, taking the advantage of learning to rank, we combine many methods to get the final ranking results.In addition, as the new music is't labeled by any tags, it could cause the emergence of "cold start" problems. This paper analyzes the lyrics, structure, and sing speed to mine the "hidden" emotional genes, after adding some "virtual" emotions tags, the new music could be returned to user by emotion tag based MIR system.