A Bimodal-based Algorithm for Song Sentiment Classification

Addressing the limitation of existing emotion classification methods for songs that only consider unimodal features from either tunes or lyrics text, and restrict classification to segments rather than the entire song, this paper proposes a novel song emotion classification algorithm based on two modalities: musical tunes and lyrics text for the entire song. Employing a two-stage self-attention method, intra-modal information within the feature vectors of lyrics and tunes is separately extracted. Subsequently, inter-modal information, fused with both modalities, is extracted. The efficacy of the proposed algorithm is validated using a self-constructed dataset and several publicly available music emotion classification datasets. Experimental results demonstrate that the algorithm presented in this paper outperforms existing song emotion classification algorithms, yielding superior classification performance.