High quality clean parallel corpora is a must for creating statistical machine translation or neural machine translation systems. Although high quality parallel corpora is largely available for official languages of the European Union, the United Nations and other organization, it is hard to encounter open parallel corpora for languages such as Turkish, which, in turn, leads to lower quality Machine Translation for these languages. In this study, we use automatic and semi-automatic procedures to collect and prepare parallel corpora in cardiology domain. We crawl a journal website and obtain 6500 Turkish abstracts and their English translations by using HTTrack. By aligning these abstracts and converting them into a translation memory in a computer-aided translation tool environment, we make it possible to use the corpora for machine translation training as well as term extraction. We argue that new tools integrating and streamlining the web crawling, alignment and cleaning steps are needed in order to support the preparation of parallel corpora for low-resource languages.
@InProceedings{DOGRU18.5, author = {Gökhan Dogru ,Adrià Martín-Mor and Anna Aguilar-Amat}, title = {Parallel Corpora Preparation for Machine Translation of Low-Resource Languages: Turkish to English Cardiology Corpora}, booktitle = {Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018)}, year = {2018}, month = {may}, date = {7-12}, location = {Miyazaki, Japan}, editor = {Maite Melero and Martin Krallinger and
Aitor Gonzalez-Agirre}, publisher = {European Language Resources Association (ELRA)}, address = {Paris, France}, isbn = {979-10-95546-03-0}, language = {english} }
Summary of the paper