Music Artist Tag Propagation with Wikipedia abstracts

Abstract

In this paper we tackle the problem of automatically assigning tags to music artists in the Web 2.0 radio Last.fm. We present a proof-of-concept method that, using a reference list of Last.fm user-defined tags, searches Wikipedia abstracts of music artists (only those written in English language) for new tag candidates. Tag candidates are ranked using an heuristic weighting function. We evaluate the top ranked tag suggestion for over 27,000 artists by (i) performing automatic evaluation using diachronic Last.fm data, and (ii) by performing manual evaluation on a sample of artists.

Our method shows promising results regarding the accurate propagation of artist tags: the top ranked suggestion is relevant for more than 50% of the artists. More specifically, the method shows good performance for artists with no previous user-defined tags, confirming that it can be worthwhile to investigate further in the context of the "cold start problem" typical of social tagging system. After analysing and discussing errors, we present several directions for future improvement of our method.