Area Audiovisual Technologies

PhD project: Natural Language Processing methods for Music Information Retrieval

Supervisor: Xavier Serra, Horacio Saggion

MdM subproject: Music meets Natural Language Processing

Contact for application:  Aurelio Ruiz ([email protected])


Today, we are witnessing an unprecedented information explosion thanks to the dramatic technological advancement brought by the Information Age. This technological (r)evolution has set the foundations for the release and publication of huge amounts of data onto online repositories such as web pages, forums, wikis and social media. Art and culture have benefited dramatically from this context, which allows potentially anyone with an available Internet connection to access, produce, publish, comment or interact with any form of media.

In this context, the music domain has attracted interest from diverse fields such as Computer Science, Music, Musicology, AI, Maths or Linguistics. In our project Music Meets Natural Language Processing we aim at representing what humans say about music, discovering and extracting knowledge from available textual sources, instead of audio and symbolic sources, as it has been studied earlier. This knowledge can be exploited to create large-scale, high quality software libraries and datasets in the music domain. For this, we will combine complementary expertise from the PI and team members in both Music, Musicology and Music Information Retrieval (MIR), along with Natural Language Processing (NLP) and Web Search & Data Mining (WSDM).

Specifically, we aim at bringing together NLP, WSDM and MIR in order to create richer repositories and software, as well as ad-hoc methods relevant to the music domain. We acknowledge the fact that Music is an artistic and scientific area which requires its own approaches and tools to cope with those cases in which generic NLP methods may fall short. One of the most exciting avenues in MIR is the exploitation of freely occurring and noisy data, thus parting ways from approaches which solely rely on structured information present in Knowledge Bases like DBpedia or Freebase. We need to account for the idiosyncratic challenges it poses (e.g. songs, albums and bands having the same name) as well as its potential applications (e.g. music recommendation or semantic search). Our focus is on knowledge extraction and exploitation from unstructured text sources. More specifically our work is focused in the following areas, Named Entity Linking (EL), Relation Extraction, Opinion Mining, entity-based Retrieval and Recommendation, and Music Information Retrieval (MIR).