Securing Multilingual Legal Information Infrastructures: A Systematic Mapping Review and Threat-and-Failure-Mode Taxonomy for Terminological Integrity

Authors:
Preprint
Publisher website
Google Scholar
Abstract
Multilingual legal information infrastructures connect official legislation repositories, terminology databases, translation memories, ontologies, search services, and artificial-intelligence systems. Conventional confidentiality-integrity-availability controls do not fully address a distinctive failure: a technically unchanged record can convey the wrong legal concept because its source, version, concept identity, cross-language scope, semantic relation, or approval history is defective. This article presents a PRISMA-informed systematic mapping review and standards synthesis. Ten query clusters yielded 100 retained scholarly occurrences and 41 official-document occurrences; after deduplication and screening, 52 substantive scholarly sources and 30 current standards, specifications, legal instruments, and guidance documents were included, with four methodology sources analysed separately. Of the scholarly corpus, 53.8% was empirical or technical, 38.5% conceptual or doctrinal, and 7.7% review-based; only 25.0% directly examined multilingual legal systems and 7.7% explicitly addressed low-resource languages. Evidence concentrated on semantic corruption, poisoning of data and retrieval, and governance/privacy, while script-specific attacks remained thinly studied. The review develops MLTIT, a seven-family taxonomy containing 33 threat and failure modes; MLTIAF, a fifteen-control defence-in-depth architecture; and a seven-risk privacy layer mapped to NIST, ISO, OWASP, MITRE ATLAS, Unicode, and legal-document standards. Analytical application to four observed Kazakh-Russian terminology discrepancies and one negative control in the SOZZ online dictionary illustrates the model’s discriminating function without treating the records as a prevalence sample or evidence of malicious attack. The central design rule is that AI may detect, rank, and explain risk, but may not autonomously alter authoritative legal content.

This publication has 37 references indexed in Scilit: