Russian compliance specialists have created an automated system to help websites comply with Federal Law No. 168-FZ on protecting the Russian language ahead of expected enforcement actions by Rospotrebnadzor and the Federal Antimonopoly Service.
The law prohibits the use of foreign words when common Russian equivalents exist and bans both Latin script and transliterations written in Cyrillic. It requires identification of prohibited borrowings by comparing text against four official dictionaries approved by the Russian Academy of Sciences: orthographic, orthoepic, explanatory, and the dictionary of foreign words.
The developed microservice, named Normograph, implements a five-stage pipeline that significantly reduces the volume of text sent to the language model. First, HTML pages are cleaned of markup and tokenized. Second, primary filtering removes words present in normative dictionaries, user-defined white lists, and registered trademarks. Third, the system flags potential issues including words absent from dictionaries, known anglicisms, and Latin-script terms.
Only after these algorithmic filters does the LLM receive small relevant fragments. The model determines whether a word qualifies as an impermissible borrowing, considers context, and proposes accurate Russian synonyms that match in gender, number, and case. The pilot uses GigaChat API but the architecture allows easy substitution with other models such as YandexGPT.
An additional OCR module based on Yandex Cloud Vision extracts text from images collected by a crawler. The same multi-stage filter is applied to recognized text, with caching and size-based filtering to control costs. Logos and registered trademarks are automatically excluded from checks.
Testing showed that manual review of one page previously took about one hour, while the automated system reduces the process to between two and ten minutes. The solution is positioned as a corporate compliance tool for pharmaceutical, fintech, and large e-commerce platforms that must adapt hundreds of pages before regulatory inspections begin.
