Learning Contextualized Music Semantics from Tags via a Siamese Network

Automatic annotation of music with tags is a promising methodology for the acquisition of semantics that facilitates music information retrieval and understanding. One of the biggest challenges for this methodology is modeling concept semantics in context. Moreover, the out of vocabulary (OOV) problem exacerbates its difficulty and has yet to be addressed so far. In this paper, we propose a novel Siamese network to fight off the challenge. By means of tag features and a probabilistic topic model, our Siamese network captures contextualized music semantics from tags via unsupervised learning, which leads to a contextualized music semantic space and a potential solution to the OOV. We have conducted simulations on two public tag collections, CAL500 and MagTag5K, and compared our approach to a number of the state-of-the-art methods. Comparative results suggest that our approach outperforms the state-of-the-art methods in terms of semantic priming measures.