NMT-translation – basic models, quality assessment

Abstract

This article is devoted to the topic of evaluating the quality of NMT-translation models, as well as methods of evaluating the quality of the obtained translations. The research we have conducted in the field of comparison of NMT-based translations of special texts is of high relevance in today’s translator work environment, where modern technologies are used to facilitate work processes and improve the quality of services. The purpose of this study is to provide a comprehensive review of current neural machine translation (NMT) models and methods for assessing the quality of translated texts produced using them, with a focus on the application of these models to the translation of legal and legislative literature. The study aims to identify the advantages and limitations of existing NMT technologies, as well as to identify areas for their further improvement. In order to achieve this goal, the following tasks were formulated: to review and analyze the main NMT models, including Seq2Seq, Transformer, BERT and their variations, in order to identify their peculiarities and applicability to the translation of different types of texts; to study and evaluate the quality of translation made with the help of modern NMT systems; to carry out a comparative analysis of the translation of legal texts using the methods of contextual analysis, structural-semantic and comparative-comparative analysis; to determine the limitations of the existing NMT models; to identify the advantages and limitations of the existing NMT models for the translation of legal texts. The paper analyses modern NMT models, including Seq2Seq, Transformer, BERT and their variations, investigating the features of each model and their applicability to different language pairs. Special attention is paid to the evaluation of NMT translation quality using BLEU, METEOR, TER metrics as well as human evaluation of translation quality. The texts of translations of specialized literature in various fields and publicistic texts on relevant topics made with the help of NMT systems were studied using contextual analysis methods, as well as structural-semantic and comparative-semantic methods. The research conducted allows us to conclude that modern NMT models are highly effective and capable of generating high-quality translations in different languages. However, despite the success achieved, some limitations have been identified that require further research and improvements. The study provides a better understanding of current NMT translation models, their applicability and translation quality. The results and conclusions of the study can be useful for both the academic community and practitioners in the field of (machine) translation and linguistics. Furthermore, the results and conclusions of this paper emphasize the importance of further research to improve the quality of automatic translation and to extend its field of application.

Keywords

• nmt translation
• neural networks
• seq2seq
• transformer
• quality assessment
• metrics