Automatic Music Classification Method Based On Users' Comments

Music classification is one of the most important part in MIR(Music Information Retrieval),which is often applied to music recommendation and music retrieval.Ex-isting music classification methods are based on genre,moods,instruments,artists and tags.However,these methods are so limited that users cannot search music with more details.An automatic music classification method based on users' comments is presented for few categories and limited search content.This method can diversify music clas-sification type and provide users with better search experience,instead of classifying music in a limited type.Because it can make best use of the advantage that users who are familiar with music are likely to comment with more details.These comments can provide significant reference for music classification.The main work in this paper is listed as follows:1)We take advantage of linear CRF model(linear Conditional Random Field)to get specialized vocabulary,and utilize N-gram word extraction and affinity analysis method to obtain a dictionary with seed generation which can be used to music corpus segmentation.Combining them makes music dictionary more accurate and much richer,and it also overcomes the shortcomings of statistic based model,which need a large scale of annotated corpus.2)We use linear CRF model to segment comments of each music,and then revise the segmentation result via split-merge testing and MMSEG(Max Matching Segmen-tation)model.These steps can make segmentation result more accurate.3)We extract candidate tags of each music by using optimized TFIDF(Term Fre-quency-Inverse Document Frequency)algorithm after comparing some keyword ex-traction methods.Optimized TFIDF can reduce the influence of words frequency and improve the accuracy of candidate tags.After that,we filter these tags in global view to delete tags which occur in few documents.4)We construct a probability classification model between music and tags to clas-sify music.5)We try to cluster similar music tags in order to minimize its effect on music classification.The experiment result shows that this method is highly effective,which can gen-erate varied music tags according to comments automatically.So that it can provide guarantees for users' personalized music retrieval.