[论文] Build2SPARQL: A Large-Scale Text-to-SPARQL Benchmark Dataset for Build...

论文概要

研究领域: ML
作者: Wooyoung Jung
发布时间: 2026-10-05
arXiv: 2610.00224

中文摘要

楼宇自动化系统越来越多地用 Brick、ASHRAE 223P 等本体表示为语义知识图谱，为 AI 应用提供机器可读基底。有前景的应用是将自然语言问题翻译为 SPARQL（text-to-SPARQL），让操作员通过语言智能体查询图谱，但进展受限于大规模基准稀缺。本文提出 Build2SPARQL，由知识图谱接地的流水线生成：SPARQL 查询完全由图遍历代码产生并验证，大模型只生成自然语言问题，使查询正确性与模型行为无关。流水线挖掘六类查询模式——线性链、分支、UNION、聚合、OPTIONAL、属性过滤，以五种词汇语域表述每个查询。应用于 201 个楼宇知识图谱（180 Brick、21 ASHRAE 223P），产出 6,136 个可执行 SPARQL 查询和 30,680 个问题。300 题双人验证：98.8% 语义保真、98.8% 自然度、84.0% 操作合理性。三个开源权重模型的检索增强评估将精确匹配从零样本 0.2–20% 提升到三样本检索 56–65%。

原文摘要

Building automation systems are increasingly represented as semantic knowledge graphs (KGs) using ontologies such as Brick and ASHRAE 223P, creating a machine-readable substrate for artificial-intelligence applications. One promising application is translating natural-language questions into SPARQL (text-to-SPARQL), which would let building operators query these graphs through language agents, but progress is limited by the scarcity of large natural-language/SPARQL benchmarks. This paper presents Build2SPARQL, a large-scale benchmark for building KGs generated by a KG-grounded pipeline: SPARQL queries are produced and validated entirely by graph-traversal code, while large language models generate only the natural-language questions, keeping query correctness independent of model behavior....

自动采集于 2026-10-05

#论文 #arXiv #ML #小凯