歌詞情緒分析

last update : Oct.8.2019

Abstract :

This repo implement some lyrics sentiment classification method, including TFIDF-based, CNN, RNN, and BERT classifier.

Dataset :

NJU_MusicMood V1.0, contains 777 songs with lyrics, and each song comes with one label, representing its sentiment : angry, happy, relaxed, or sad. Please check data/NJU-MusicMood-v1.0.htm for more information.

Reference : Multimodal Music Mood Classification by Fusion of Audio and Lyrics.

files and usage :

• data/ : Original datasets.
• pkls/ : Processed data (in pickle format) to use in CNN and RNN. Check cnn_data_process.py
• To use pre-trained GoogleNews word2vec model, please download from here.

Method :

• new: BERT classifier(see bert_utli.py and train_bert.py), now training and tuning...

　 　 　 　 　 　 　 　 　 　 　

　

　 　 　 　 　

　

　 　 　 　 　

　

　 　 　 　 　

Method	Description	Train_Acc	Test_Acc
TFIDF(Baseline)	Regular bag-of-word TFIDF with SVM and Random Forest P.S. SVD DOESNOT help	SVM:0.8625 RF :0.8625	SVM:0.292 RF :0.284
tfidf based classifer	Ref: AUTOMATIC MOOD CLASSIFICATIONUSING TF*IDF BASED ON LYRICS	0.845	0.297
Convolutional Neural Network	Ref: Convolutional Neural Networks for Sentence Classification A Sensitivity Analysis of (and Practitioners’ Guide to) Convolutional Neural Networks for Sentence Classification	0.95-0.98	0.40-0.47
RNN(bidirectional LSTM)	Fucked up.	over 0.95	about 0.3