基于预训练语言模型的电子乐谱情感分类研究

Research on Sentiment Classification of Digital Sheet Music Based on Pre-trained Language Model

乐谱是人们进行音乐学习与交流的重要媒介。以MusicXML为代表的电子乐谱格式可准确记录音乐序列和歌词等信息,现已在音乐行业内得到广泛应用,对其进行情感分类可在音乐推荐、基于情感调控的音乐生成等下游任务中起到一定的辅助作用。符号音乐与自然语言文本具有类似的序列特征,故自然语言处理的相关方法也可应用至符号音乐的研究中。本文构建了MusicXML乐谱数据集和带有情感标签的英文歌词数据集,并提出了基于精调后的大规模预训练语言模型BERT的电子乐谱情感分类模型。实验表明,该模型在本文构建的数据集上取得了良好的效果,性能相比于基线方法有较为显著的提升。

Sheet music is an important medium for music learning and communication.The digital sheet music format represented by MusicXML can accurately record information such as music sequences and lyrics,and has been widely used in the music industry,the sentiment classification of which plays a certain role in a series of downstream tasks such as music recommendation and sentiment-controlled music generation.Symbolic music and natural language text share similar sequence characteristics,thus the methods of natural language processing can also be applied to symbolic music.This paper constructs a MusicXML sheet music dataset and a sentiment labeled English lyrics dataset,and proposes a digital sheet music sentiment classification model based on fine-tuned BERT,a large-scale pre-trained language model.Experiments show that the model has achieved good results on the dataset constructed in this paper,and there is a significant improvement of the performance compared with baseline methods.