Supervised Text Style Transfer Using Neural Machine Translation: Converting between Old and Modern Turkish as an Example

摘要

Languages evolve and change over time. Accordingly, texts that were written a hundred years ago may become incomprehensible, such as hundred-year-old Turkish texts. Additionally, making old written work accessible to today's generation requires qualified writers, who are responsible for the process of conversion. Unfortunately, that is costly in both time and resources. To work out this problem, we develop an automatic style conversion system. We formulate our problem as a machine translation problem and use the recently popularized Neural Machine Translation techniques. Furthermore, we introduce a data-driven approach to align source and target word vectors. Although we do not introduce new model components over the standard RNN encoder-decoder, the way we utilize monolingual data to pre-train our word vectors lead to significant improvements. Despite the simplicity of our approach, we outperform complex approaches. We achieve a BLEU score of 33.8 points, improving our baseline by 12 points.