SESEMMI for LinkedMusic: Democratizing Access to Musical Archives via Large Language Models

Liam Pond, Sichen Meng, Linnea Kirby, Simon Ngassam, Sebastien Chow, Dylan Hillerbrand, Ichiro Fujinaga

Keywords: Large Language Models, Natural Language Query, SPARQL, Prompt Engineering, Music Metadata, Linked Data, Music Information Retrieval

TL;DR: This study systematically evaluates the ability for large language models to turn natural language queries into SPARQL queries for integrated music databases.

Abstract:

Currently, there are over one hundred music metadata databases online; comprehensively answering even simple questions often means querying dozens of them separately. This fragmentation makes large-scale, cross-cultural, or longitudinal research difficult and time-consuming. The LinkedMusic initiative aims to solve this problem by combining these databases in one place. The ingested data are stored in RDF format and can be queried using SPARQL, a querying language.

However, SPARQL’s complexity makes it prohibitively difficult for most users to use effectively. Our project, the Search Engine System for Enhancing Music Metadata Interoperability (SESEMMI), aims to overcome this barrier by providing a natural language interface for LinkedMusic. Using large language models, it translates the user’s plain-language queries into SPARQL queries that retrieve results from all integrated databases.

In this paper, we conduct the first systematic study of the ability of LLMs in translating Natural Language Queries (NLQ) to SPARQL in the domain of music metadata research. We evaluate five models on twenty music-domain NLQ-to-SPARQL pairs with manually prepared ground-truth outputs. Results indicate that Claude Sonnet 4 achieves the highest accuracy of 100.0% on single-database queries in both zero- and one-shot con-texts and 46.7% for complex zero-shot cross-database queries.

Submission Number: 10