How Low Can We Go? Quantization Effects on LLM SPARQL Generation

Matt Murtagh-White; P. J. Wall; Declan O’Sullivan

doi:10.1007/978-3-031-99554-5_18

How Low Can We Go? Quantization Effects on LLM SPARQL Generation

Matt Murtagh-White
, P. J. Wall
, Declan O’Sullivan

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › peer-review

Abstract

SPARQL is a powerful but complex language for querying knowledge graphs, motivating research into natural language-to-SPARQL generation using large language models (LLMs). While large, proprietary LLMs excel at this task, their resource requirements can limit practical deployment. This paper evaluates smaller, open-source LLMs (0.5B–9B parameters) with quantization methods (8-bit and 4-bit compression) to balance computational efficiency and query generation performance. Our findings demonstrate that 8-bit quantization can maintain or enhance performance in smaller models, whereas 4-bit quantization leads to notable degradation, especially for larger models. This highlights the potential of quantized, smaller LLMs for SPARQL generation in resource-constrained scenarios and provides insights for optimizing specialized NLP tasks.

Original language	English
Title of host publication	The Semantic Web
Subtitle of host publication	ESWC 2025 Satellite Events, Proceedings
Editors	Edward Curry, John McCrae, Valentina Presutti, Mehwish Alam, Pieter Colpaert, Josiane Xavier Parreira, Diego Collarana, Marta Sabou, Andreas Harth, Pasquale Lisena
Publisher	Springer Science and Business Media Deutschland GmbH
Pages	99-102
Number of pages	4
ISBN (Print)	9783031995538
DOIs	https://doi.org/10.1007/978-3-031-99554-5_18
Publication status	Published - 2026
Event	Satellite events held at the 22nd European Semantic Web Conference, ESWC 2025 - Portoroz, Slovenia Duration: 1 Jun 2025 → 5 Jun 2025

Publication series

Name	Lecture Notes in Computer Science
Volume	15832 LNCS
ISSN (Print)	0302-9743
ISSN (Electronic)	1611-3349

Conference

Conference	Satellite events held at the 22nd European Semantic Web Conference, ESWC 2025
Country/Territory	Slovenia
City	Portoroz
Period	1/06/25 → 5/06/25

UN SDGs

This output contributes to the following UN Sustainable Development Goals (SDGs)

SDG 8 Decent Work and Economic Growth
SDG 12 Responsible Consumption and Production

Keywords

Knowledge Graphs
Large Language Models
Quantization
Resource Efficiency
SPARQL

Access to Document

10.1007/978-3-031-99554-5_18

978-3-031-99554-5_18Final published version, 144 KBLicence: CC BY

Cite this

Murtagh-White, M., Wall, P. J., & O’Sullivan, D. (2026). How Low Can We Go? Quantization Effects on LLM SPARQL Generation. In E. Curry, J. McCrae, V. Presutti, M. Alam, P. Colpaert, J. Xavier Parreira, D. Collarana, M. Sabou, A. Harth, & P. Lisena (Eds.), The Semantic Web: ESWC 2025 Satellite Events, Proceedings (pp. 99-102). (Lecture Notes in Computer Science; Vol. 15832 LNCS). Springer Science and Business Media Deutschland GmbH. https://doi.org/10.1007/978-3-031-99554-5_18

Murtagh-White, Matt ; Wall, P. J. ; O’Sullivan, Declan. / How Low Can We Go? Quantization Effects on LLM SPARQL Generation. The Semantic Web: ESWC 2025 Satellite Events, Proceedings. editor / Edward Curry ; John McCrae ; Valentina Presutti ; Mehwish Alam ; Pieter Colpaert ; Josiane Xavier Parreira ; Diego Collarana ; Marta Sabou ; Andreas Harth ; Pasquale Lisena. Springer Science and Business Media Deutschland GmbH, 2026. pp. 99-102 (Lecture Notes in Computer Science).

@inproceedings{854eeffc8574490483dd82267e24ae0b,

title = "How Low Can We Go? Quantization Effects on LLM SPARQL Generation",

abstract = "SPARQL is a powerful but complex language for querying knowledge graphs, motivating research into natural language-to-SPARQL generation using large language models (LLMs). While large, proprietary LLMs excel at this task, their resource requirements can limit practical deployment. This paper evaluates smaller, open-source LLMs (0.5B–9B parameters) with quantization methods (8-bit and 4-bit compression) to balance computational efficiency and query generation performance. Our findings demonstrate that 8-bit quantization can maintain or enhance performance in smaller models, whereas 4-bit quantization leads to notable degradation, especially for larger models. This highlights the potential of quantized, smaller LLMs for SPARQL generation in resource-constrained scenarios and provides insights for optimizing specialized NLP tasks.",

keywords = "Knowledge Graphs, Large Language Models, Quantization, Resource Efficiency, SPARQL",

author = "Matt Murtagh-White and Wall, \{P. J.\} and Declan O{\textquoteright}Sullivan",

note = "Publisher Copyright: {\textcopyright} The Author(s), under exclusive license to Springer Nature Switzerland AG 2026.; Satellite events held at the 22nd European Semantic Web Conference, ESWC 2025 ; Conference date: 01-06-2025 Through 05-06-2025",

year = "2026",

doi = "10.1007/978-3-031-99554-5\_18",

language = "English",

isbn = "9783031995538",

series = "Lecture Notes in Computer Science",

publisher = "Springer Science and Business Media Deutschland GmbH",

pages = "99--102",

editor = "Edward Curry and John McCrae and Valentina Presutti and Mehwish Alam and Pieter Colpaert and \{Xavier Parreira\}, Josiane and Diego Collarana and Marta Sabou and Andreas Harth and Pasquale Lisena",

booktitle = "The Semantic Web",

address = "Germany",

}

Murtagh-White, M, Wall, PJ & O’Sullivan, D 2026, How Low Can We Go? Quantization Effects on LLM SPARQL Generation. in E Curry, J McCrae, V Presutti, M Alam, P Colpaert, J Xavier Parreira, D Collarana, M Sabou, A Harth & P Lisena (eds), The Semantic Web: ESWC 2025 Satellite Events, Proceedings. Lecture Notes in Computer Science, vol. 15832 LNCS, Springer Science and Business Media Deutschland GmbH, pp. 99-102, Satellite events held at the 22nd European Semantic Web Conference, ESWC 2025, Portoroz, Slovenia, 1/06/25. https://doi.org/10.1007/978-3-031-99554-5_18

How Low Can We Go? Quantization Effects on LLM SPARQL Generation. / Murtagh-White, Matt; Wall, P. J.; O’Sullivan, Declan.
The Semantic Web: ESWC 2025 Satellite Events, Proceedings. ed. / Edward Curry; John McCrae; Valentina Presutti; Mehwish Alam; Pieter Colpaert; Josiane Xavier Parreira; Diego Collarana; Marta Sabou; Andreas Harth; Pasquale Lisena. Springer Science and Business Media Deutschland GmbH, 2026. p. 99-102 (Lecture Notes in Computer Science; Vol. 15832 LNCS).

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › peer-review

TY - GEN

T1 - How Low Can We Go? Quantization Effects on LLM SPARQL Generation

AU - Murtagh-White, Matt

AU - Wall, P. J.

AU - O’Sullivan, Declan

N1 - Publisher Copyright: © The Author(s), under exclusive license to Springer Nature Switzerland AG 2026.

PY - 2026

Y1 - 2026

N2 - SPARQL is a powerful but complex language for querying knowledge graphs, motivating research into natural language-to-SPARQL generation using large language models (LLMs). While large, proprietary LLMs excel at this task, their resource requirements can limit practical deployment. This paper evaluates smaller, open-source LLMs (0.5B–9B parameters) with quantization methods (8-bit and 4-bit compression) to balance computational efficiency and query generation performance. Our findings demonstrate that 8-bit quantization can maintain or enhance performance in smaller models, whereas 4-bit quantization leads to notable degradation, especially for larger models. This highlights the potential of quantized, smaller LLMs for SPARQL generation in resource-constrained scenarios and provides insights for optimizing specialized NLP tasks.

AB - SPARQL is a powerful but complex language for querying knowledge graphs, motivating research into natural language-to-SPARQL generation using large language models (LLMs). While large, proprietary LLMs excel at this task, their resource requirements can limit practical deployment. This paper evaluates smaller, open-source LLMs (0.5B–9B parameters) with quantization methods (8-bit and 4-bit compression) to balance computational efficiency and query generation performance. Our findings demonstrate that 8-bit quantization can maintain or enhance performance in smaller models, whereas 4-bit quantization leads to notable degradation, especially for larger models. This highlights the potential of quantized, smaller LLMs for SPARQL generation in resource-constrained scenarios and provides insights for optimizing specialized NLP tasks.

KW - Knowledge Graphs

KW - Large Language Models

KW - Quantization

KW - Resource Efficiency

KW - SPARQL

UR - https://www.scopus.com/pages/publications/105020237438

U2 - 10.1007/978-3-031-99554-5_18

DO - 10.1007/978-3-031-99554-5_18

M3 - Conference contribution

AN - SCOPUS:105020237438

SN - 9783031995538

T3 - Lecture Notes in Computer Science

SP - 99

EP - 102

BT - The Semantic Web

A2 - Curry, Edward

A2 - McCrae, John

A2 - Presutti, Valentina

A2 - Alam, Mehwish

A2 - Colpaert, Pieter

A2 - Xavier Parreira, Josiane

A2 - Collarana, Diego

A2 - Sabou, Marta

A2 - Harth, Andreas

A2 - Lisena, Pasquale

PB - Springer Science and Business Media Deutschland GmbH

T2 - Satellite events held at the 22nd European Semantic Web Conference, ESWC 2025

Y2 - 1 June 2025 through 5 June 2025

ER -

Murtagh-White M, Wall PJ, O’Sullivan D. How Low Can We Go? Quantization Effects on LLM SPARQL Generation. In Curry E, McCrae J, Presutti V, Alam M, Colpaert P, Xavier Parreira J, Collarana D, Sabou M, Harth A, Lisena P, editors, The Semantic Web: ESWC 2025 Satellite Events, Proceedings. Springer Science and Business Media Deutschland GmbH. 2026. p. 99-102. (Lecture Notes in Computer Science). doi: 10.1007/978-3-031-99554-5_18

How Low Can We Go? Quantization Effects on LLM SPARQL Generation

Abstract

Publication series

Conference

UN SDGs

Keywords

Access to Document

Fingerprint

Cite this