基于用户评论的自动化音乐分类方法

发布时间：2018-05-12 23:01

本文选题：音乐分类 + 分词模型　；参考：《中国科学技术大学》2017年硕士论文

【摘要】：音乐分类作为音乐信息检索(Music Information Retrieval,MIR)领域的一个重要分支,常用于音乐检索和音乐推荐。现有的音乐分类方法从音乐的流派、感情、乐器、艺术家和标注五个角度进行分类。但是这些分类方法都过于局限,它们将音乐的类别限制在了一个固定的范围内,导致用户无法根据音乐的细节信息进行音乐检索。针对音乐分类类别固定、搜索内容过于局限的问题,本文提出了一种基于用户评论的自动化音乐分类方法。此方法不再局限于已有的音乐类别,可以得到更为多样化的分类结果,为用户提供更个性化的检索体验。该方法的出发点为用户对音乐的评论描述更为深入,这些详细的描述对音乐的分类有重要的参考价值。本文的主要工作如下:1)首先利用线性链条件随机场(linear Conditional Random Field,linear CRF)模型识别专业名词。然后使用N元词串提取和紧密度分析方法,利用种子生成的思想得到适合音乐语料分词的字典。此混合方法能获取较为准确和丰富的字典,降低基于统计的分词模型对标注语料的需求。2)使用linearCRF和上述的音乐字典进行分词。接着使用基于词汇紧密度分析的分合测试评估分词结果。接着使用MMSEG(Max Matching Segmentation)模型进行分词修正,使得修正后的分词结果拥有较高的准确率。3)对比多个关键词提取算法,选择TFIDF(Term Frequency-Inverse Document Frequency)算法并优化,削弱了词频在提取过程中的影响,提高了候选标签的准确性。然后再从全局角度对候选标签进行过滤,得到音乐的关联标签。4)建立音乐的多标签概率分类模型,对音乐进行分类。5)尝试对音乐标签按相似程度进行聚类,减小对音乐分类模型的影响。实验结果表明,该音乐分类方法准确率较高,可以无监督地获取音乐多个维度的标签,为个性化的音乐检索提供了保障。
[Abstract]:As an important branch of music Information retrieval, music classification is often used in music retrieval and music recommendation. The existing music classification methods are classified from five aspects: genre, emotion, musical instrument, artist and label. However, these classification methods are too limited, they limit the category of music to a fixed range, so users can not search the music according to the details of the music. In order to solve the problem of fixed categories of music classification and too limited search content, an automatic music classification method based on user comments is proposed in this paper. This method is no longer limited to the existing music categories and can obtain more diversified classification results and provide users with more personalized retrieval experience. The starting point of this method is that the user's comments on music are more in-depth, and these detailed descriptions have important reference value for the classification of music. The main work of this paper is as follows: 1) first, we use linear Conditional Random Conditional Random nonlinear CRF model to identify professional nouns. Then, using the method of N element string extraction and compactness analysis, a dictionary suitable for music corpus segmentation is obtained by using the idea of seed generation. This hybrid method can obtain more accurate and rich dictionaries and reduce the need of tagging corpus based on statistical participle model. (2) linearCRF and the music dictionary mentioned above are used to segment words. Then the word segmentation results were evaluated by compositional test based on lexical compactness analysis. Then we use the MMSEG(Max Matching Segmentation) model to modify the word segmentation, which makes the modified segmentation result have higher accuracy. 3) comparing with many keyword extraction algorithms, selecting and optimizing the TFIDF(Term Frequency-Inverse Document frequency algorithm, which weakens the influence of word frequency in the extraction process. The accuracy of candidate labels is improved. Then the candidate labels are filtered from the global perspective, and the associated labels of music. 4) the multi-label probability classification model of music is established, and the music is classified. 5) try to cluster the music labels according to the similarity degree. Reduce the influence on the music classification model. The experimental results show that this music classification method has a high accuracy and can obtain labels of multiple dimensions of music without supervision, which provides a guarantee for individualized music retrieval.
【学位授予单位】：中国科学技术大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.3

【参考文献】