Corpus-Based Translation Automation of Adaptable Corpus Translation Module

Thesis deals with the translation module development which automates corpus-based translation. The translation module is a part of an adaptable corpus tool and is implemented as a separate microservice. This translation module provides for the linguist and translator the ability to translate texts with conveying their style. The translation module is domain-specific oriented which allows to convey text style better than public cloud translation services. In this research religious and historical texts were analyzed. Neural machine translation method was justified and used. Sequence to sequence Transformer model as a neural network model was chosen. All stages of text processing by the Transformer model which based on the Multi-Head Attention mechanism were analyzed. Software libraries and toolkits for the Sequence to sequence Transformer model were analyzed and chosen. Based on chosen software libraries implemented own Transformer model implementation. Developed model comprises text preprocessing and neural network model implementation. Cost-efficient computer system which comprises hardware and software components for neural network model training was configured. Based on heuristic approach by carrying out computational experiments neural network model hyper-parameters were chosen and justified. Loss function, learning rate, perplexity and BLEU as a key model training criterion were analyzed and applied. Training and test samples of text data sets were prepared. Training and test data sets comprise language pairs of Ukrainian text fragments and their English equivalents. Configured neural network model was trained and tested. Automatic assessment approach of trained model which based on semantic closeness was suggested and tested.