Creating a Parallel Corpora for Turkish-English Academic Translations
İlhami SEL, Hüseyin Üzen, Davut Hanbay · Computer Science · 2021
Parallel corpora are data sets created by representing sentences with the same meaning in different languages. One of the most important elements that determine the quality in machine translation systems is the parallel corpora created in large quantities and with high quality. Such data for the Turkish – English language pair are generally insufficient. In this study, a large amount of parallel corpora has been created that can be used for academic translations between Turkish and English languages. While creating this data set, the abstracts of the postgraduate theses were used. The best matches were obtained using sentence alignment algorithms such as Vecalign and Hunalign. As a result of the studies, 1M parallel sentence pairs were obtained. In addition, an Bi-LSTM-based translation system was created to measure the quality of the obtained data. The created model obtained 15.8 Bleu points with zero-shot learning method on the TED (Tr-En) test set.