Punctuation Correction in Uzbek: A Lightweight Machine Learning Approach
Year: 2026
Type: conference-paper
Abstract: This paper examines the problem of punctuation checking in Uzbek texts. Punctuation affects understanding of meaning, thought boundaries, and readability. Errors are common in real-world Uzbek texts because people write quickly, mix styles, use colloquial forms, and don’t always follow rules. Furthermore, many texts are published without an editor, especially online. Therefore, automatic punctuation checking can be useful both for regular users and for organizations that prepare and publish materials. This paper formulates the problem in a simple way: it considers the position between words and determines which character should appear in that position. The model selects one option from a small set of classes, such as “no character,” “comma,” “period,” “question mark,” and “exclamation mark.” This format makes the problem understandable and allows for easy comparison of different methods. The paper also compares three simple machine learning approaches: a linear model based on text features, a sequential CRF model, and a CatBoost model based on tabular features. All approaches use a unified data preparation logic and similar features to ensure fair comparison. Special attention is paid to testing on texts of various subject areas, as writing style and punctuation quality vary greatly by genre. We test models on news, official texts, educational and scientific materials, social media posts, and dialogues from literary texts. The results show that the sequential model, on average, better maintains consistency and is less likely to place “random” punctuation marks next to each other. The linear model produces similar results and performs well on “cleaner” genres. CatBoost shows consistent results, but often underperforms on “dirtier” texts with many non-standard forms and free-flowing styles.
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7167229294