Music Recommendation Model Based on Semantic Representation of Song Lyrics Using BERT

Abstract

The rapid growth of digital music platforms has resulted in an information overload problem, making it difficult for users to discover songs that match their preferences. This study proposes a content-based music recommendation model through semantic analysis of song lyrics using a Natural Language Processing approach with Bidirectional Encoder Representations from Transformers. The research stages include Indonesian song lyric data collection, data cleaning, text preprocessing, contextual lyric embedding generation, and lyric similarity computation using cosine similarity. Model performance is evaluated using Mean Squared Error and accuracy. Experimental results show that the proposed model achieves an accuracy of 83.69% with a Mean Squared Error value of 1.4066, indicating that lyric representations generated by Bidirectional Encoder Representations from Transformers effectively capture semantic meaning and quantitatively improve the relevance of music recommendations. Therefore, the proposed approach enhances the accuracy and personalization of content-based music recommendation systems.