Research Of Music Automatic Annotation Algorithm Based On Lyrics

With the rapid development of information technology,a massive amount of digital music appears on the internet.How to organize digital music data has become a challenge to industry and academic research.Music annotation,which describes the music information by generating tags for music,provides a way for quick retrieval,efficient management and music recommendation.There are three ways for labeling the music which are expert annotation,social annotation and automatic annotation based on machine learning algorithms.Though expert annotation can get high accuracy rate of labels,high cost and label abundance makes it unsatisfactory.Social annotation may solve these issues,but it has low accuracy performance.Automatic annotation based on machine learning algorithms is becoming a research topic.Especially,automatic music labeling based on lyrics attracts wide attention,considering the lyrics contain much useful semantic information.In this paper,we first propose a modified text classification method for the music annotation based on lyrics.We replace bag-of-words by bag-of-characters as the input features to eliminate noise which is caused by text preprocessing.Then we use n-gram bag-of-characters to mine more textual information from lyrics,and propose a feature extraction way by using joint bag-of-characters to solve the feature spareness problems of n-gram bag-of-characters.Next,we come up with two deep learning methods to mine deep semantic features automatically based on the modified text classification method above.One of the methods is to use Deep Belief Network(DBN)to mine deep features from the shallow textual features which are generated by the modified text classification method for the music annotation.The other way is to employ Convolutional Neural Network(CNN)to obtain the semantic features of lyrics,in which the network uses convolutional layers and pooling layers to reduce the number of network parameters.The experiments verify the superior performances of the two methods.Finally,we add the audio information from the music into the automatic music annotation framework.We then train the Convolutional Neural Network for lyrics and audio respectively and then connect the last layers of those two networks together to build the deep features to do the annotation tasks.We conduct the experiments to evaluate that the combination of lyrics and audio information is superior to using either one of them independently in music annotation.