【作者】 刘怡； 高玥；

【Author】 Yi Liu, Yue Gao School of Information, Renmin University, Beijing, China

【机构】 中国人民大学信息学院；

【摘要】 情感是音乐的语义之一,按照情感自动对歌曲进行分类,能够为基于情感的音乐检索检索提供基础的数据支持,具有重要的实际意义和学术意义,也因之成为音乐信息检索领域的热点研究课题。本文利用大型音乐网站中的评论和标注构建情感模型及数据集,提出以歌曲为单位进行分类的方法,通过大量实验分析考察了片段长度、分类器和特征选择算法对情感分类的影响;并对Active Selection特征选择算法,提出最大差值探测法选取初始特征集,大大提高了分类准确率。实验结果表明,15s长片段最利于歌曲的情感分类,SVM与利用最大差值探测法选取初始特征集的Active Selection算法相结合的方法可以获得最好的分类效果。在对取自著名大型音乐网站Last.fm的400首歌的10,365个15s的音乐片段进行6类分类时,在没有人工标注的情况下,达到54.82%的分类准确率。更多还原

【Abstract】 Nowadays, there are more and more music databases in our daily life and people prefer to use music retrieval system to get songs. Mood can express inherent emotional meaning of a song,and have great influence on people, so music retrieval based on mood become more and more important. Music mood classification can provide basic data support for searching, therefore that has significantly actual meaning and academic meaning and thus it is becoming an important research topic. There are some problems existing in the present work on music mood classification. The first problem is how to establish the classification model. Most of mood classification models are derived from psychologies and derivative models, but whether it’s appropriate applying traditional psychological model to mood classification or not is still a question should be solved. The second problem is the establishment of data set. The frequently used methods are almost based on manual labeling, but those are time exhausting and the dataset on that basis is always small and composed of clips. The third problem is what kind of impact the parameters (such as clip length, classifier, etc.) selected in mood classification will have. The last but not least factor to which we should pay attention will be the features used in classification, usually including timbre, intensity, rhythm, etc. Due to the existence of invalid and redundant features, the classification performances of feature combinations are not necessarily optimal. Aiming at the above-mentioned problems, this thesis has done the following research: Firstly, mood model and dataset have been established by taking advantages of comments and labels from major music websites. Considering the different ways to express emotion, mood model and datasets per song have been constructed respectively for Chinese classic music and western’s pop songs and comparisons have been made between mood model and traditional psychological model to evaluate the differences and applicability of both. Then classificationmethod based on unit of songs is proposed. The impact which clip lengths have on mood classification is analyzed to find out the most proper length of clips for classification and to express different emotions. Experiment shows that using 15s long clips for classification can get the highest accuracy. Furthermore, comparisons have been made between two classifiers, Support vector machine (SVM) and Gauss mixture model (GMM), which are frequently used in previous work. SVM performs best in the experiment. Finally, four feature selection algorithms, ReliefF, Fisher Rule, Step Forward Selection and Active Selection, are used to eliminate invalid and redundant features. Comparisons have been made among these algorithms, and for Active Selection algorithm, largest spread detection method is proposed to extract initial feature set to improve the feature selection efficiency.更多还原