A Structural Ontology of the Law: BFO-Aligned OWL Rendering (v1.0)

This dataset is a Basic Formal Ontology (BFO 2020) aligned formal rendering of the ontological commitments expressed in David R. Koepsell's forthcoming monograph A Structural Ontology of the Law (Palgrave Macmillan, 2026). The ontology was constructed using a dialogue-driven pipeline in which a large language model proposed typed class and relation assertions for each atomic claim extracted from the source text, and each proposal was validated by the HermiT OWL-DL reasoner against the current working graph before commitment.The resulting ontology contains 6,881 classes and 1,678 individuals, derived from 6,445 committed claims. All classes are typed against BFO 2020; the property layer includes 16 standard BFO/RO relations with their formal characteristics (transitivity, inverse pairs, domain and range constraints) and nine top-level disjointness axioms. HermiT confirms the final ontology is consistent under these axioms.The pipeline that produced this ontology is released separately under Apache 2.0. dkoepsell/bfo-agent: bfo-agent - extracts a BFO ontology from a text Evaluation of the pipeline on questions targeting this ontology's content showed a 0.933 ± 0.058 refusal rate on out-of-graph questions, compared to 0.100 ± 0.100 for an unmodified LLM baseline (n=3 runs per condition), demonstrating that persistent ontological memory substantially reduces LLM confabulation on domain-specific questions. The complete evaluation reports are included in this deposit.Included files:

SOoL_final.owl: the ontology itself in OWL/XML formatbfo_relations.ttl: BFO/RO property and disjointness axioms (loaded alongside the main ontology to enable reasoner inference)legal_seed.ttl: initial legal-domain class scaffold (if applicable)frozen_results/: evaluation reports establishing the confabulation-resistance result*_summary.json: aggregated variance results across independent evaluation runs

Methodological caveats: This is a pilot artifact derived from a single monograph by a single author. The reconstruction reflects what the pipeline extracted and the reviewer (the author) approved; a small number of post-extraction near-duplicate classes were merged in a deduplication pass. The ontology is intended as a reference artifact for readers of the monograph and for researchers interested in BFO-aligned formal renderings of philosophical positions, not as a comprehensive ontology of law in general.Suggested citation: Koepsell, D. R. (2026). A Structural Ontology of the Law: BFO-Aligned OWL Rendering (v1.0) [Data set]. Zenodo. [DOI will be generated on publish]

本数据集为适配基础形式本体论2020版（Basic Formal Ontology 2020, BFO 2020）的形式化表示，对应大卫·R·凯普塞尔（David R. Koepsell）即将出版的专著《法律的结构本体论》（A Structural Ontology of the Law, Palgrave Macmillan, 2026）中所阐述的本体论承诺。

本本体采用对话驱动的流水线构建：首先由大语言模型（Large Language Model, LLM）针对从源文本中提取的每一条原子断言，生成带类型的类与关系断言；随后在提交前，使用HermiT OWL-DL推理机基于当前工作图谱对所有断言提案进行校验。

最终生成的本体共包含6881个类与1678个个体，源自6445条经确认的断言。所有类均基于BFO 2020进行类型标注；属性层包含16条标准BFO/RO关系及其形式化特征（传递性、逆关系对、定义域与值域约束），以及9条顶层不相交公理。经HermiT推理机校验，最终本体在上述公理下具备一致性。

构建该本体的流水线已以Apache 2.0协议单独开源：dkoepsell/bfo-agent：bfo-agent——从文本中提取BFO本体的工具。

针对本本体内容的流水线评估结果显示：针对图谱外问题的拒答率为0.933±0.058，而未修改的大语言模型基线仅为0.100±0.100（每组条件均进行3次运行），这表明持久化本体记忆可显著降低大语言模型在领域特定问题上的幻觉生成现象。本数据集附带完整的评估报告。

附带文件：
SOoL_final.owl：采用OWL/XML格式存储的本体文件
bfo_relations.ttl：BFO/RO属性与不相交公理文件（与主体本体一同加载以支持推理机推理）
legal_seed.ttl：初始法律领域类模板（如适用）
frozen_results/：用于验证本体抗幻觉生成性能的评估报告目录
*_summary.json：跨独立评估运行的聚合方差结果文件

方法学说明：本数据集为单作者单本专著衍生的试点产物。本体重构结果反映了流水线提取内容以及审核者（即原作者）认可的内容；我们在去重步骤中合并了少量提取后出现的近似重复类。本本体旨在为该专著读者以及关注基于BFO的哲学立场形式化表示的研究者提供参考样例，而非通用法律领域的全面本体。

推荐引用格式：Koepsell, D. R. (2026). A Structural Ontology of the Law: BFO-Aligned OWL Rendering (v1.0) [Data set]. Zenodo. 发布时将生成DOI编号。