Analysis of typical errors of Russian-speaking learners in English speech using NLP technologies by the example of ecological vocabulary

Abstract

Systematic deviations in the use of English ecological vocabulary by Russian-speaking learners arise under cross-linguistic interference, yet the available typologies of such deviations remain fragmentary and rely mainly on the manual processing of small samples without quantitative verification. The study aims to identify and quantitatively describe the typical errors in the English speech of Russian-speaking learners on the material of ecological vocabulary by means of automated natural language processing. The empirical basis was an author's model learner corpus of 1,247 texts with a total volume of 218,643 tokens, annotated by error type using the spaCy tokenizer and dependency parser, the ERRANT error classifier and a BERT-based contextual detector, with subsequent manual verification of a gold-standard subsample. The error density reached 44.9 deviations per thousand tokens; the grammatical class accounted for 51.3% of the annotated spans and the lexical-semantic class for 36.0%. The highest resistance to interference was shown by high-frequency concrete terms, whereas abstract process nominations concentrated a disproportionately high share of word-choice and collocation errors. The accuracy of automatic annotation predictably decreased from formal categories towards lexical-semantic ones. The distribution of deviations in ecological vocabulary is determined not by the frequency of a term but by the degree of structural divergence between its English model and the Russian equivalent, which shifts the priority of corrective work from memorising terms to rebuilding collocational and derivational schemes.