Research Integrity in Low-Resource Languages: Lessons from a Multilingual Plagiarism Detector for Central Asia

Abstract

Since Resolution No. 836 of the Cabinet of Ministers (13 December 2024), every higher-education institution in Uzbekistan must check graduation projects and master's dissertations with an anti-plagiarism system that works in Uzbek, Karakalpak, Russian and English and detects paraphrase and AI-generated text. Most commercial detectors were not designed for these languages: Uzbek is written in two scripts and uses several apostrophe-like letters, and Karakalpak adds characters that generic text pipelines damage. This paper presents Similex (similex.org), a plagiarism detector built for these conditions. Similex maps Latin and Cyrillic Uzbek onto one character skeleton and fingerprints text by winnowing character n-grams (k = 32, w = 16). On a frozen in-house benchmark of 250 cases, it reached 84.7% recall, 95.4% precision and an F1 of 89.7%. A change in verdict logic raised recall on suffix-modified copies from 20.5% to 100%, while an attack sweep of 1,880 checks exposed a detection floor of about 500 characters. By 17 October 2026 the index held about 41.9 million documents. The main lesson is that, in low-resource languages, orthography and verdict design cause more failures than the fingerprinting algorithm itself. Resolution 836 sets no accuracy thresholds; national test sets and independent evaluation are needed.