TECHNOLOGY FOR GRAMMATICAL ERRORS CORRECTION IN UKRAINIAN TEXT CONTENT BASED ON MACHINE LEARNING METHODS

Abstract

Context. Most research in grammatical and stylistic error correction focuses on error correction in English-language textual content. Thanks to the availability of large data sets, a significant increase in the accuracy of English grammar correction has been achieved. Unfortunately, there are few studies on other languages. Systems for the English language are constantly developing and currently actively use machine learning methods: classification (sequence tagging) and machine translation. A large amount of parallel or manually labelled data is required to build a high-quality machine learning model for correcting grammatical/stylistic errors in the texts of those morphologically complex languages. Manual data annotation requires a lot of effort by professional linguists, which makes the creation of text corpora, especially in morphologically rich languages, mainly Ukrainian, a time- and resource-consuming process.
 Objective of the study is to develop a technology for correcting errors in Ukrainian-language texts based on machine learning methods using a small set of annotated parallel data.
 Method. For this study, machine learning algorithms were selected when developing a system for correcting errors in Ukrainianlanguage texts using an optimal pipeline, including pre-processing and selecting text content and generating features in small annotated data corpora. The neural network’s use with a new architecture, a review of state-of-the-art methods, and a comparison of different pipeline stages will make it possible to determine such a combination of them, allowing a high-quality error correction model in Ukrainian-language texts.
 Results. A machine learning model for error correction in Ukrainian-language texts has been developed. A universal scheme for creating an error correction system for different languages is proposed. According to the results, the neural network can correct simple sentences written in Ukrainian. However, creating a full-fledged system will require spell-checking using dictionaries and checking rules, both simple and based on the result of parsing dependencies or other features. The pre-trained neural translation model mT5 has the best performance among the three models. To save computing resources, it is also possible to use a pre-trained BERT-type neural network as an encoder and a decoder. Such a neural network has half the number of parameters as other pretrained machine translation models and shows satisfactory results in correcting grammatical and stylistic errors.
 Conclusions. The created model shows excellent classification results on test data. The calculated machine translation quality metrics allow only a partial comparison of the models since most of the words and phrases in the original and corrected sentences are the same. The best value for both BLEU (0.908) and METEOR (0.956) is obtained for mT5, which is consistent with the case study in which the most accurate error corrections without changing the initial value of the sentence are obtained for such a neural network. The M2M100 has a higher BLEU score (0.847) than the “Ukrainian Roberta” Encoder-Decoder (0.697). However, subjectively evaluating the results of the correction of examples, the M2M100 does a much worse job than the other two models. For METEOR, M2M100 (0.925) also has a higher score than the “Ukrainian Roberta” Encoder-Decoder (0.876).