Music relevance model training method, music search method, and device

Abstract

Embodiments of the present application provide a music relevance model training method, a music search method and equipment, which sequentially use a first sample set, a second sample set and a third sample set to train a music relevance model, wherein the first music samples in the first sample set are generated according to user interaction data on a music document corresponding to a search term, the relevance labels of the second music samples in the second sample set are annotated by a large language model, and the relevance labels of the third music samples in the third sample set are manually annotated. The second sample set is obtained by annotating the relevance of the original music samples by the large language model, which is more efficient than manual annotation, so that the training samples can be quickly and efficiently obtained. The relevance is annotated by the large language model, which increases the number of training samples, thereby improving the performance of the music relevance model obtained by training.

Images