January 21, Dublin, Ireland: KantanMT is pleased to announce that it has launched a major new release of its market leading KantanMT platform. This release follows an extensive study of Transformer Neural Networks (TNN); research that clearly demonstrated TNNs produce a notably higher translation quality output than Recurrent Neural Networks (RNN). RNNs is the principal technology employed by most MT providers today.
“Working with one of the world’s largest eCommerce companies, extensive comparisons were carried out to determine what type of Neural Network architecture leads to the best possible translation outputs”, said Tony O’Dowd, Chief Architect at KantanMT.com.
“Comparisons using RNN (Recurrent Neural Networks), CNN (Convolutional Neural Networks) and TNN (Transformer Neural Networks) were carried out on the KantanMT platform and translation outputs were compared to Google and Bing Translate. The comparisons demonstrated that KantanMT’s Transformer models were remarkably better in translation fluency and adequacy than all the other major players in the market,” O’Dowd added.
“We decided to switch over to contextually more accurate and faster Transformer Networks”, said Marek Mazur, Director of Development at KantanMT.com, “as we believe these offer the best all-round quality and decoding speeds for our clients”.
The detailed report, which compared MT-produced translation quality for the language arc English => Italian, and undertaken by one of the world’s largest eCommerce companies in conjunction with KantanMT.com, can be found at https://kantanmtblog.com/2019/01/14/joint-study-confirms-kantan-tnn-delivers-remarkable-quality-scores/
It is free to access.
(Since the blog was first published, comparative analysis has also been carried out for English=>German, English=>Spanish and English=>French language combinations and in all cases Kantan TNNs outperformed CNNs, RNNs, Google and Bing Translate.)