Work

The Saudi Legal Corpus for AI: An Auditable, Official-Source-Grounded, Article-Level Corpus of Saudi Arabian Legislation

Year: 2026
Type: preprint
Abstract: Legal natural language processing has advanced rapidly for English and several European languages, but Arabic --- and Saudi Arabian law in particular --- remains critically under-resourced: to our knowledge, no openly documented, machine-readable, article-level corpus of Saudi legislation has previously been described. We present the \corpusname, a corpus of 290 legislative instruments (201 statutory laws, 64 statutory regulations, 24 implementing regulations, and one set of procedural rules) covering the Kingdom of Saudi Arabia's principal legislation, structured into 15{,}689 article-level records ($\sim$1.2 million Arabic tokens) in a unified, retrieval-ready index spanning twelve legal domains. The corpus rests on three design commitments that distinguish it from web-scraped legal collections: official-source grounding, in which every track records its issuing authority and official source and every record carries its own provenance label (251 distinct labels) under a four-tier source-verification taxonomy; auditability, enforced by 376 read-only, idempotent validators and fully deterministic derived layers; and the governing-text principle, under which the official Arabic text governs and all other language layers are explicitly non-authoritative. Beyond the text, the corpus ships analytical layers that are, to our knowledge, the first of their kind for Saudi law: a statutory citation graph (3{,}607 references, 585 of them between distinct instruments), a hand-classified supersession graph (102 edges), a glossary of 1{,}920 statutorily defined terms with 3{,}318 definitions, an embedding-ready chunking layer (16{,}304 chunks), and a manually confirmed retrieval benchmark of 519 gold queries. On that benchmark the corpus's metadata-based lexical searcher reaches 93.3\% top-1 accuracy against 90.4\% for a standard BM25 baseline, with the gap concentrated on definitional queries (90.9\% versus 74.5\%), quantifying the value of the engineered retrieval metadata while leaving informative headroom for neural Arabic legal retrieval. We describe the construction methodology in full, report statistics that are reproducible from the archived release, and discuss intended uses, limitations, and the legal and ethical boundaries of releasing structured national legislation for artificial-intelligence research.
Language: English
Subfield: Law
Open Access status: green
OpenAlex ID: W7204887766
Cites: 13
Cited by:
Related to: 10