基于条件随机场的音乐共同语义标注

Conditional random fields model for collective annotation of music

音乐的语义标注旨在使用词语或标签自动将一段音乐标注为一个语义标注集。通常,人们将多标注学习转换为独立二进制分类问题解决,再给每一个语义标注单独建模。为了得到更好的分类结果,应考虑标注之间的依赖关系。文章中尝试共同的音乐语义标注,对单标注和具有高相关性的成对标注同时建立模型。使用多标注条件随机场(CRF)模型直接参数化多标注分类中的共现标注。用到两种CRF模型,一种是使用无条件标注相关的共同多标注分类(CML)模型,另一种是使用有条件标注相关的考虑特征的共同多标注分类器(CMLF)模型。实验表明,将这两种模型用到CAL10K数据集上,平均精确度、宏F1分数和微F1分数比用高斯混合模型(GMM)给单个标注建模要高。

Music semantic annotation aims to automatically annotate a music signal with a set of semantic labels using words or tags.Usually,multi-label learning is done by transforming into multiple independent binary classification problems and model each semantic label individually.In order to produce better classification result,label correlations should be taken into account.In this paper,we try to use collective music semantic annotation,which not only builds a model for each semantic label,but also builds models for the pairs of labels that have significant correlations.We use multi-label conditional random fields（CRF）model to parametric the co-occurrence of multi-label classification.Two CRF models are proposed here,one is called collective multi-label classifier（CML）while utilized unconditional label correlation;while the other is called collective multi-label with features classifier（CMLF）which utilizes conditional label correlation.Experiments show that using these two models have higher average accurancy,macro-averaged F1 score and micro-averaged F1 score than using Guassian mixture model on the issue of semantic annotation while using CAL10 Kdatasets.