【作者】 周利娟；

【导师】 林鸿飞；

【作者基本信息】 大连理工大学 ， 计算机应用技术， 2011， 硕士

【摘要】 随着科技的发展,电子数据的存储能力越来越强,使得海量电子多媒体数据的存储和管理成为可能。在线音乐数据数量的爆炸式增长给音乐收听用户带来了很大的选择空间,但同时也给在线音乐收听服务系统带来了巨大挑战,如何识别用户内心最真实的需求和如何推荐给用户最感兴趣的歌曲成了目前很多音乐检索和推荐领域研究人员面临的巨大问题。在音乐信息检索系统中,很多用户无法确切地指定自己想听哪一首歌曲,而是提供给音乐分享和收听系统一些不含有任何和音乐相关的描述性信息(本文中定义为“非描述性音乐查询”),目前流行的基于音乐属性关键字匹配的音乐检索算法就不能满足用户的需求。如何理解用户提供的非描述性音乐查询和如何将音乐数据和非描述性音乐查询进行相似度计算就成了目前流行的音乐检索系统中需要解决的问题。情感计算是目前人机交互领域中重要的研究内容,其主要目标是解决如何让机器识别人的情感从而让机器更好地与人交互为人服务的问题。本文提出在音乐检索系统中存在非描述性查询无法准确处理的问题,并基于音乐是情感的表达的观点,提出用情感语义分析方式解决音乐检索系统中非描述性查询的处理问题。本文基于文本情感计算的研究,借助于文本情感分类和识别的方法,对非描述性查询和音乐进行高维情感语义空间建模,并在情感语义空间中计算查询和音乐之间的情感语义相似度,并给出相似度排序。本文关注点在于音乐的情感建模和分类以及在基于情感分析的音乐检索模型研究。具体过程如下：首先,根据WordNet-Affect定义音乐情感语义空间模型,将音乐情感分成七大类,这种分类方法考虑到了音乐情感表达的内在特征,并扩充了文本情感的分类。其次,本文从国际知名音乐分享网站last.fm自动抽取海量音乐相关数据,构建音乐检索数据集(DUTMIR-Dataset),并人工标注了情感信息用于机器学习音乐情感的分类。并根据音乐的属性数据和社会化数据等属性构建不同的特征集合,验证不同的特征选择对于音乐情感语义空间表示的作用和对检索性能的影响。再次,为了进行音乐的情感分类,考虑到了音乐语料集短小、精炼、含蓄的特点,本文利用LDA模型和聚类方法扩充和均衡了音乐数据的属性集,解决了音乐数据的稀疏性的问题,并通过实验验证多种音乐情感分类器在音乐情感分类方面的性能。最后,为了综合显示本文提出的基于音乐情感识别的音乐检索模型的性能,本文开发出对应的应用原型系统并测试运行。更多还原

【Abstract】 As the rapid development of science and technology, the fast increase of digital storage capacity makes it possible that enormous multimedia data can be stored and managed automatically. The exploration of online shared music provides more choices to users as well more challenges to music sharing web service systems. The problem of understanding the real intent of system users and recommending them with most relevant music items has been increasingly challenging. In Music Information Retrieval (MIR) System, many system users submit queries that contain no descriptive information about music. This kind of query with no desciptive information are defined as "Non-Descriptive Query" in this paper, and usually can not be tackled well in common music search and download websites. It is urgent to find a solution to understand the implicit emotional inquiry of systems users and compute the similarity between music and non-descriptive queries.Affective Computing is a significant research area in Human Computer Interaction. The main purpose is to recognize the emotion of computer users and serve them emotionally. As music is sentimentally expressive, the problem of processing Non-Descriptive Queries can be addressed via detecting implicit emotion. Our study on new model of Music Information Retrieval (MIR) is based on text emotion detection and recognition techniques. Queries and music are represented in a high dimensional emotion space and the similarity is computed according to their relevance in the high dimensional emotion space.The focus of this paper is on modeling music in emotion space, including:First, we define music emotion space according to WordNet-Affect. The categories are extended to 7 types considering the intrinsic feature of music.Second, we download a large dataset from last.fm and build DUTMIR-Dataset with manual annotation, which applied in our machine learning theories based music emotion classification and produce 3 types of different feature sets to evaluate their influence on MIR.Third, as music data is short, concise and implicit, we utilize LDA model to attach recommended tags to music, which conquers the sparsity and imbalance of the music dataset. Besides, different classifiers are tested to get the best one for MIR system.Last but not least, in order to show our model in an obvious way, we design and develop a prototype system.更多还原