Improving music auto-tagging with trigger-based context model

摘要

Music auto-tagging has been an active research topic as it learns the relationship between the content of audio tracks and semantic tags such that users can query by both tags and audio segments without being troubled by the cold start problem. In this paper, we propose a new trigger-based context model to refine the existing content model based auto-tagging systems. The trigger based context model improves accruacy of weakly labeled tags in “Genre”, “Solo” and “Usage” by 10.63%, 10% and 26.43% respectively, which are usually poorly modeled due to lack of data in the content model based systems. Experiment results indicate that a combination of the content and context models outperforms the content based only auto-tagging system and the baseline Turnbull's MixHier model by 0.74% and 2.64% in average precision rate respectively.