Problems and prospects of using language testing in determining native speakers of the Russian language in solving crime prevention problems

Abstract

The study addresses the issue of reliably differentiating between native speakers of Russian (NSR) and individuals who speak Russian as a foreign language at levels C1-C2 (RFL) in the context of crime prevention and operational investigative activities of law enforcement agencies. The relevance is driven by the increasing number of linguistic examinations and the limitations of traditional expert methods, which are prone to subjectivity and unsuitable for mass and rapid analysis. The goal is to develop and test an integrated technological model for NSR/RFL classification that incorporates lexical-grammatical, error-based, and stylometric features. The materials comprise a balanced corpus of 7,8 million word usages, divided into NSR (4,2 million; 500 informants) and RFL (3,6 million; 450 informants with TORFL levels $3/4)$ subcorpora, including spontaneous speech, essays, and correspondence texts; all data were anonymized and normalized; oral speech was transcribed with annotations for pauses, intonation, and tempo. Methods: calculation of TTR, sentence length, syntactic complexity index, and part-of-speech distributions; taxonomy of more than 50 error types; stylometry of function word, n-gram, and punctuation frequencies; creation of multidimensional feature vectors (TF-IDF, BoW); training of SVM, gradient boosting, and LSTM models with 10-fold cross-validation. The results revealed systematic differences: NSR exhibited lower TTR but a higher compound sentence index, a greater proportion of function words, and longer sentences; in the RFL group, the frequencies of systematic errors (verb aspect, prepositional-case government, inversion, particles) were 5-13 times higher. Stylometric features (profiles of the 100 most frequent words and 3-grams) demonstrated the greatest predictive power, ensuring the maximum intergroup distance. The SVM + gradient boosting ensemble achieved high performance: accuracy 0.947, precision 0.962, recall 0.928, F1 0.945 on the holdout sample. The discussion emphasizes that reliable identification is impossible based on individual markers and requires the synergy of heterogeneous levels of analysis reflecting automated speech habits. The model's practical value lies in supporting decision-making during rapid text and audio analysis, reducing the risk of false positives, and improving reproducibility. Limitations are related to the corpus domain and the need to account for prosody; prospects include data expansion, inclusion of phonetic features, application of deeper architectures, and the development of legal and ethical frameworks for its use. Practical implementation involves integrating the model into departmental platforms for screening digital traces while adhering to the principles of proportionality and personal data protection.