Review-Driven Multi-Label Music Style Classification by Exploiting Style Correlations

Proceedings of NAACL-HLT 2019, pages 2884–2891 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 2884 Review-Driven Multi-Label Music Style Classification by Exploiting Style Correlations Guangxiang Zhao1∗, Jingjing Xu2∗, Qi Zeng2, Xuancheng Ren2, Xu Sun1,2 1Center for Data Science, Beijing Institute of Big Data Research, Peking University 2MOE Key Lab of Computational Linguistics, School of EECS, Peking University {zhaoguangxiang,jingjingxu,pkuzengqi,renxc,xusun}@pku.edu.cn Abstract This paper explores a new natural language processing task, review-driven multi-label mu- sic style classification. This task requires sys- tems to identify multiple styles of music based on its reviews on websites. The biggest chal- lenge lies in the complicated relations of mu- sic styles. To tackle this problem, we pro- pose a novel deep learning approach to auto- matically learn and exploit style correlations. Experiment results show that our approach achieves large improvements over baselines on the proposed dataset. Furthermore, the visu- alized analysis shows that our approach per- forms well in capturing style correlations.1 1 Introduction As music style (e.g., Jazz, Pop, and Rock) is one of the most frequently used labels for music, music style classification is an important task for appli- cations of music recommendation, music informa- tion retrieval, etc. Several methods have been pro- posed for automatic music style classification (Qin and Ma, 2005; Zhou et al., 2006; Wang et al., 2009; Choi et al., 2017). Most of them mainly fo- cus on using audio information to identify styles. Motivated by the fact that a pieces of music could has different styles, several studies (Wang et al., 2009; Oramas et al., 2017) also aim at multi-label music style classification. Although these methods make promising progress, they are limited in two aspects. First, not all audio data is available in real-world ap- plications because of copyright restrictions, which limits the generalization ability. Second, some of them are based on a strong assumption that a piece of music should be assigned with only one style. Different from these studies, we focus on using ∗Equal Contribution 1The code and the dataset are available at https:// github.com/lancopku/RMSC easily obtained reviews in conjunction with multi- label music style classification. The motivation comes from the fact that lots of user reviews con- tain rich style-related information, which can be used for music style classification. The major challenge of this task lies in the com- plicated correlations of music styles. For exam- ple, Soul Music2 contains elements of R&B and Jazz. These three labels can be used alone or in combination. Traditional multi-label classification methods may mistake the true label [Soul Mu- sic, R&B, Jazz] for the false label [R&B, Jazz]. If well learned, style relations are useful knowl- edge for improving the performance, e.g., increas- ing the probability of Soul Music if we find that it is heavily linked with two high probability labels: R&B and Jazz. Therefore, to better exploit style correlations, we propose a novel deep learning ap- proach with two parts: a label-graph based neural network, and a soft training mechanism with cor- relation …

Content truncated to 500 words