摘要
随着Web 2.0的发展,音乐自动语义标注成为音乐检索系统的关键技术。但是,目前主流的语义模型都是对音频的内容特征进行处理,并且对每个标签独立建模,忽略了标签间的关联产生的音乐上下文特征。将最大熵模型应用于音乐自动语义标注中,对音乐上下文特征进行建模处理,可以通过约束条件的多少调节模型对已知数据的拟合程度和对未知数据的适应度,并自然地解决统计模型中参数平滑的问题。实验表明,最大熵模型具有较高的预测准确率,同时,在建模过程中引入音乐相似度对特征信息函数进行选择,可以提高系统性能。
With the development of Web 2. 0, music auto-tagging has become a key technology of music retrieval system. However, in typical music auto-tagging and retrieval systems, all tag level models are trained based on music content of audio features independently, ignoring the music context features between tags. This article applies the maximum entropy model to music auto-tagging system, in order to process the music context features, adjust the fitness of both known data and unknown data, and naturally smooth the parameters in the statistical model by changing the number of constraint conditions. Experimental results show that the maximum entropy model has relatively high prediction accuracy. In addition, the applying of tag similarity in feature information function selection can improve the prediction performance.
出处
《电子测量技术》
2014年第12期32-35,40,共5页
Electronic Measurement Technology
关键词
音乐自动语义标注
最大熵模型
特征信息函数选择
music auto-tagging
maximum entropy model
feature information function selection
