Research On Lyric-based Music Sentiment Classification

Music is significant in people’s ordinary life which, as a special manner of emotion expression and communication, provides rich information. Moreover, recent years have seen widespread existence of varying kinds of music (audio, video, and lyric, etc.) on the Web. How to use computer to process music on the Web has become an interesting research topic in the field of information processing. In this thesis, we focus on the task of lyric-based sentiment classification. Especially we define this task as a binary-classification problemWe first study several basic problems of sentiment classification in the framework of supervised learning:1) we examine the effect of Chinese word segmentation on sentiment classification accuracy and show the necessity of building domain-specific segmented for sentiment classification;2) we compare several representative classifiers including a supervised topic model which is of interest in machine learning field;3) we compare different methods for computing feature values including simple word counts and TF-IDF.Second, we then study the problem of using unlabeled data for sentiment classification. In particular, we focus on partially-supervised and semi-supervised learning frameworks. Partially-supervised learning is regarded as a one-class classification problem whose training data contains human annotations for only one class. In respect to semi-supervised learning, we pay special attention to an unsupervised topic model to mine topic information under documents which is used as new features for sentiment classification.Finally, we study two approaches to expanding sentiment lexicon for sentiment classification. In particular, we propose to use bootstrapping to obtain sentimental words from unlabeled data, and propose to use a Maximum Entropy model to obtain sentiment words from human-labeled data.