Сравнение методов автоматической разметки речевых формул в русскоязычном интернет-дискурсе: пилотное исследование

С. 420–429.

Попова Т. И., Масленикова А. С.

This study focuses on developing and comparing methods for automatic annotation of speech formulas in a corpus of Russian internet comments. Speech formulas are a class of multiword expressions that convey emotional reactions in dialogue. The research material consisted of a corpus of 10,000 comments (157,261 tokens) collected from five Telegram channels. Dictionary-based formal search using 437 units achieved 21% precision. To improve accuracy, three methods were developed: Random Forest classification (56% precision), dependency parsing-based syntactic filtering (73.3% precision, 8.7% recall), and punctuation-based filtering (76.4% precision, 74.0% recall). Analysis revealed that syntactic parsers systematically misclassify interjective units: 68.5% of true speech formulas received the advmod label instead of the correct ROOT. The punctuation-based method showed the best results, improving precision by 3.64 times over the baseline formal search. The research demonstrates that for linguistic phenomena with clear formal markers, simple rule-based methods can outperform machine learning, especially with limited annotated data.

Keywords: машинное обучение, интернет-дискурс, речевые формулы, устойчивые неоднословные единицы, синтаксический парсинг

In book

Компьютерная лингвистика и интеллектуальные технологии: По материалам ежегодной международной конференции «Диалог». Выпуск 24.

Issue 24. , M.: Max press, 2026.