SENTIMENT ANALYSIS OF MUSIC REVIEWS USING DEEP LEARNING:A BIDIRECTIONAL LSTM APPROACH

Abstract

This study looks at sentiment analysis of music reviews in depth   using deep learning methods, specifically a bidirectional Long Short-Term Memory (BiLSTM) neural network architecture. The study looks at the expanding demand for automated sentiment categorization in the music industry, where understanding what people think about reviews can have a big effect on recommendation systems, marketing strategies, and content curation.

This study uses a dataset of 78,162 music album reviews that rate albums on a scale from 0 to 5. It uses a complex preprocessing pipeline that includes cleaning the data, balancing the classes by oversampling, and using advanced tokenization methods. The suggested BiLSTM model has an embedding layer with a vocabulary size of 30,000 and 128 dimensions, GlobalMaxPooling1D for feature extraction, and dense layers with ReLU and softmax activations for the final classification.

The findings of the trial show that the system works quite well, with a test accuracy of 91% across all rating classes. The model has good precision and recalls values for all sentiment categories, which means it can classify things quite well. The training took place over five epochs and used categorical Cross entropy loss and the Adam optimizer. After oversampling, the balanced dataset had about 39,045 samples per class.

This study adds to the body of knowledge by showing that bidirectional LSTM architectures function well for music-specific sentiment analysis and by giving a complete foundation for using comparable systems in entertainment industry applications. The results show that there is a lot of promise for real-world use in music streaming services and recommendation systems.