Research on Music Emotion Recognition Based on CBSA Fusion Model

Music emotion recognition is a key technology in music information retrieval. Its core challenge lies in how to extract interpretable emotion representation knowledge from heterogeneous audio features. At present, there are generally two deficiencies in knowledge management in the research: On the one hand, there is a lack of knowledge verification of feature-sentiment correlation, resulting in redundant features interfering with the knowledge acquisition of the model. On the other hand, traditional LSTM is difficult to establish knowledge dependencies across fragments. In response to the bottlenecks of the above-mentioned knowledge discovery, this study proposes a fusion knowledge modelling framework based on CNN-BILSTM-Self-Attention (CBSA). Its innovation in knowledge management is reflected in: At the knowledge representation layer, the primary emotional semantic unit is constructed through the local perception domain of CNN. At the knowledge evolution layer, BiLSTM is utilised to capture the transition patterns of emotional states in music sequences. In the knowledge refinement layer, a self-attention mechanism is adopted to quantify the weights of global emotional features and generate an interpretable knowledge graph of attention distribution. The experimental results show that this fusion model can improve the accuracy and efficiency of music emotion recognition. The contribution of this study lies not only in establishing a hierarchical knowledge extraction paradigm for musical emotional features and provide quantifiable emotional knowledge representation tools for the music recommendation system. It also verified the feature selection value of the attention mechanism in music knowledge modelling. The research scope mainly focus on the recognition of musical emotions. Future work will explore knowledge management tools for the discovery of cross-cultural emotional knowledge.