linkedmusic-queries

There are various methods to query our data lake, e.g., Using SPARQL against Open Link Virtuoso's graphs. However, For music users, what if they don't know how to handle SPARQL coding, the accessibility of our data lake would be weakened. Therefore, the initial goal in this repository is to explore "How to convert the users' natural language queries to SPARQL query, with the efficacy of Large Language Models". Above all, it's about ->

Achieving NLQ2SPARQL through Prompt Engineering on Large Language Models

0. The initial experiment

(2024June-2024Oct)

For any individual question, we custimize context as minimun as possible. The context can be anything, including ontology rendered in OWL language. We conducted experiment with result in the "LLM2SPARQL_TheSession_WikiData_MusicBrainz" folder, domonstrated on LinkedMusic 2024 conference

1. Use ontology snippet as context for prompte engineering

(2024Dec-2025Mar)

Junjun Cao has been utlizing a Chinese Traditional Music KnowledgeBase (abbreviated as CTM) as a prior experiment because this knowledge base has a comprehensive ontology, which can serve as ready-made context for prompting ChatGPT to convert NLQ(natural language query) into SPARQL

for issues marked with "CTM", it's specifically meant for Chinese Traditional Music knowledgeBase

This experiment has solved some problems/issues:

This experiment addresses the challenge of oversized ontologies hindering LLMs in generating accurate SPARQL queries. Our ->

"ontology-subgraph-extraction approach"

--focuses the LLM on relevant local ontology snippets. For detailed information, please refer to the paper abstract at https://github.com/DDMAL/linkedmusic-queries/blob/main/ChineseTraditionalMusicKnowledgeBase/Knowledge%20Base%20for%20ESEA%20(East-and-Southeast-Asian)%20Traditional%20Music%20and%20its%20NLQ2SPARQL%20Intelligent%20Question-Answering%20System%20Development.pdf

For more details, please contact Junjun Cao (alienmusedh@gmail.com) for the entire paper on this subject

1.1 For other RDF databases, how to extract the ontology automatically?

There are at least 2 methods:

1.1.1 Via VOID (Vocabulary of Interlinked Datasets)

• Please locate to a cloned repository to DDMAL: https://github.com/DDMAL/void-generator which has been refined specifically, or refer to the original: https://github.com/JervenBolleman/void-generator

• void-generator is associated with the paper "LLM-based SPARQL Query Generation from Natural Language over Federated Knowledge Graphs"(Vincent, 2024) which proposed to:

generate "shapes" (a genuine schema of an RDF database) for any given RDF graph

• That said, we can prompt LLMs to automatically create a "norminal" ontology directly based on VOID information. Refer to https://github.com/DDMAL/void-generator/tree/main

1.1.2 From Schema of (RDB or serveral CSVs) to Ontology

For RDB2RDF in the internal process of Open Link Virtuoso, you can map a schema of a relational database (which usually contains sevearl CSVs whthin) to an ontology. Please refer to recorded tutorials in https://github.com/DDMAL/linkedmusic-datalake/tree/main/doc/rdb2rdf

However, for ontology customization in the process, refer to R2RML language that is introduced in "2. ISQL of Terminal + R2RML" section of the file "Documentation_CSV2RDF_Virtuoso2024Nov" located at https://github.com/DDMAL/linkedmusic-datalake/tree/main/doc/CSV2RDFInVirtuoso

1.2 How to carry out NLQ2SPARQL in federal queries-query across different databases?

--especially when the different databases are all reconciled with Wikidata. This is to be explored

2. There are other potential approaches especially when it's difficult to obtain the ontology or in other special cases

2.1 Please refer to #27

2.2 Junjun Cao has also conducted a primary literature review on NLQ2SPARQL which can be read on his paper to be published, titled "Knowledge Base for ESEA (East-and-Southeast-Asian) Traditional Music and its Ontology-subgraph-driven NLQ2SPARQL Intelligent Question-Answering System Research"