2018 has brought new domain-adaptive NMT solutions from virtually every company in this competitive market. Previously, one would either go with stock NMT or invest in a custom-trained NMT model (starting from $50K/year plus hardware costs). Now, it's possible to adjust a stock model to a custom domain using very little data (starting from 10K data points) and money (hundreds USD), with almost one-button experience. We present a thorough evaluation of available cloud solutions: their translation quality, learning curve, and the total cost of ownership.
After getting a PhD in 2008, Konstantin Savenkov led research and development efforts for online content services, then worked as CTO at Zvooq and as a chief operating officer at Bookmate. In 2016, he contributed his experience in artificial intelligence (AI), tech and operations to found Intento, Inc., where they build tools to source, evaluate and use machine translation and other cognitive AI services.