ModernMT—neural machine translation as a Service


  • Administrator
  • Hero Member
  • *****
    • Posts: 811691
    • Gender:Male
  • point d’amour
ModernMT—neural machine translation as a Service

English ↔ Spanish (es)
English ↔ Italian (it)
English ↔ Portuguese (pt-PT)
English ↔ German (de)
English ↔ French (fr)
English ↔ Dutch (nl)
English ↔ Russian (ru)
English ↔ Arabic (ar)
English ↔ Chinese Simplified (zh-CN)
English ↔ Chinese Traditional (zh-TW)
English ↔ Portuguese Brazilian (pt-BR)
English ↔ Japanese (ja)
English ↔ Catalan (ca)
English → Bosnian (bs)
English → Croatian (hr)
English → Serbian Latin (sr)
English → Finnish (fi)
English → Greek (el)
English → Polish (pl)
English → Czech (cs)
English → Hungarian (hu)
English → Danish (da)
English → Norwegian Bokmål (nb)
English → Norwegian Nynorsk (nn)
English → Icelandic (is)
English → Swedish (sv)
English → Turkish (tr)
English → Indonesian (id)
English → Malay (ms)
English → Korean (ko)
English → Thai (th)

Again, that was two years ago. Now the research part of the project is finished, the tool has been separated into a commercial version and an open-source version (much like MateCat), and the commercial version is available for these language combinations: English <> ES, IT, PT-PT/BR, DE, FR, NL, RU, AR, ZH-CN/TW, JA, CA, and English > BS (Bosnian), HR (Croatian), SR (Serbian Latin), FI, EL, PL, CS, HU, DA, ND/NN (Norwegian Bokmål and Nynorsk), IS, SV, TR, ID, MS, KO, TH. (Forgive me for spelling some of the abbreviations out, but I'm assuming my own ignorance may also be shared by others -- and for BS/HR/HR, see this recent and interesting article.)

The reason why I'm revisiting MMT is twofold. First, they are now offering a product that is usable for individual translators, and second, reader Michael Beijer contacted me after testing it and being impressed by the product but shocked about the price (which is presently app. 4 euro/1,000 words for companies and 89 euro/month for up to 100,000 translated words for translators).
Jost Zetzsche, The 297th Tool Box Journal

I know we've always tried not to say this to our kids, but after reading the article I wrote in issue 297 of the Tool Box Journal about ModernMT, it's really hard not to at least think it. I described a refreshingly different and capable machine translation system (I'll give a short overview down below) with two major problems: the outrageously high pricing and the completely outdated privacy concept that used any customers' data for general training purposes. (Davide Caroselli, ModernMT's VP of product, whom I talked to for both that article and this, defended their privacy concept by saying: "You're right, we also find it difficult at first to explain the privacy issue to our customers.")

Well, I'm pleased to announce that they have completely overhauled both their pricing and their privacy considerations. There is no longer any cross-training, and any data that you upload to enhance your own machine translation will be strictly used by you only (see here). Kind of what you'd expect from a paid product these days. (By the way, the data is stored in their own data center in Italy, and they'll be adding a larger data center soon in the US to accommodate AirBnB, their flagship client.)

The pricing is in line with other tools. You pay by the number of characters as an LSP or translation buyer (between $8 and $50 per million characters, depending on whether you want to train the engine as you translate or queue documents for a batch translation) and a monthly fee if you are a freelance translator ($25).

ModernMT differs from its competitors (at least the ones off the shelf) because you adapt it to your own data and then continue to adapt it without actually spending time to train it. Its base engine is trained on the large data sets collected by Translated, the Italian LSP and tech developer of the massive MyMemory TM (Translated also owns a majority of the shares in ModernMT). And while the baseline engines are retrained once or twice a year, your data effectively sits in the middle and adapts the MT suggestions to your style and terminology. The system uses a technology called "instance-based adaptive NMT," which sends translation requests to a TM layer (consisting of even a relatively small TM as long as it's highly tuned). Once similar segments are found in that TM layer, the NMT engine's "hyperparameters" are adapted on-the-fly to generate a more suitable suggestion. (The concept is based on this paper by the Fondazio Bruno Kessler.)

I asked Davide how many freelance translators are using the tool with the old payment plan, and not surprisingly it's only a little handful. I've talked to one who praised the translation quality as far superior to other engines, a result that Intento has also concluded in one of their reports (see here -- note that they have also worked for ModernMT so there might be a conflict of interest). Either way, if you are inclined to work with adaptive MT, this might be worth a try. Presently you can use it directly within SDL Studio Trados with this app, and of course in MateCat (owned by translated) or via Intento (see elsewhere in this Journal).
Jost Zetzsche, The 308th Tool Box Journal

« Last Edit: 12 Jan, 2020, 01:02:14 by spiros »


Search Tools