Controllability for English-Ukrainian Machine Translation by Using Style Transfer Techniques

Abstract

While straightforward machine translation got significant improvements in the last 10 years with the arrival of encoder-decoder neural networks and transformers architecture, controllable machine translation still remains a difficult task, which requires lots of research. Existing methods like tagging provide very limited control over model results or they require to support multiple models at once, like domain fine-tuning approach.In this paper, we propose a method to control translation results style by transferring features from a set of texts with target structure and wording. Our solution consists of new modifications for the encoder-decoder networks, where we can add feature descriptors to each token embedding to decode input text into the translation with the proposed domain. In conducted experiments with English-Ukrainian translation and a set of 4 domains our proposed model gives more options to influence the result than some existing approaches to solve the controllability model.