Research Highlights

Statistics-based Text Translation Technologies

Within project ISTRION, we have been developing phrase-based statistical machine translation through supervision. As we incorporate human validation, our systems are able to rapidly improve performance with time and obtain better translation quality than MOSES (a state-of-the-art SMT system) when trained on the same parallel text collection. Our translation quality is 14 BLEU points higher, mean value, than MOSES, for 16 directions among Portuguese, English, French, Spanish and German. Our results were spun off by a recently founded start-up named ISTRION BOX.