Curva Fin Bloque
POSTS 22 MAY, 2012

Pangeanic machine translation at TMS Inspiration Days, Krakow


Just a few days ago, I was an invited speaker on behalf of Pangeanic at the TMS Inspiration Days. This was the first opportunity to present the new developments in machine translation at PangeaMT, providing a sneak preview of the full features of the SaaS Power version to be released in July 2012. TMS Inspiration Days is a small yet very friendly and socially active conference. Networking and industry catching-up are excellent because the conference is well organized and not too big nor too small for it. Industry solutions are focused and well presented. The conference began with a keynote by Ben Sargent, from CSA, who stressed the importance of technology application in the translation industry. “Every LSP has a system, a TMS system, even if they do not know it. Years ago, it was FedEx and DHL instead of email. It is the proof that LSPs have been integrating systems and applications for a long time and are keen technology adapters.” Ben stressed that there are 3 ways the LSPs look at technology: they develop it for their use and for their clients (developers of a system), they are buyers of systems (they do not want the headache of starting from scratch and prefer to buy ready-build solutions) and finally there are those who prefer the mix&match approach (buying some good solutions outside and building interfaces and what they know works best for their business). The trend is towards unification of tools rather than having an array of TMS tools and working their interface. The challenges are greater for LSPs and that is why LSPs need portals and online access. Six or seven years ago LSPs were translating in 5-10 languages, where nowadays is not uncommon to have projects from 10-20 languages, with all human interaction that comes with it. TMS tries to reduce human contact to availability/unavailability and checking quality of purchased good. There was a well-prepared statement “user friendliness of TMS is….not that important” which caused general laughter. Really nowadays people look for features, so make sure all the important features and functionalities are there. Then, look at the beauty of the system. For example, how many clicks are required to finish a job? From 4 to 10, can you multiply that by the number of languages. The number of clicks per project is paramount as a measurement of staff efficiency. Elegance, usability are very important. Bob Donalson commented that the more clicks the more chances to make a mistake. Keep TMS simple so everything can be done with as few clicks as possible. I was fortunate enough to be the second presenter and introduce the benefits of machine translation in our fields, challenges for adoption and, in general, how Pangeanic took the risk and invested in the technology to provide self-sustainable, self-updating systems fit for LSP and corporate use rather than rely on external solutions. A copy of the presentation can be found here. Here’s some of the most relevant links:

Tweets (see #tmsinspiration for a summary of the venue)

  • All the attendees to @manuelhrrnz presentation are already familiar with #MT What does that tell you? Yes, it’s coming #tmsinspiration
  • @manuelhrrnz: A skilled post-editor becomes increasingly good at spotting right or wrong n-grams, not just individual words #tmsinspiration
  • @BobD_Austin I guess the audience was not so familiar with #MT after all #tmsinspiration localization
  • @manuelhrrnz: Re-training your #MT system with anything over 2,000 words would have an impact in your results #tmsinspiration
  • @BobD_Austin: RT “Statistical #MT w/out rules-based core doesn’t work for LSPs” <- I do not agree w/”core”.-> True. Add-in? #tmsinspiration”
  • @BobD_Austin RT @localizing: “Statistical #MT without rules-based core doesn’t work for LSPs” #tmsinspiration <- I do not agree w/”core” #tmsinspiration Loc For those who didn’t know already: Statistical #MT without a rules-based core in it doesn’t work for LSPs #tmsinspiration #tmsinspiration
  • @manuelhrrnz data cleaning is most important and least appreciated stfp. “Your TMs are _not_ clean! #tmsinspiration
  • @manuelhrrnz Example: reordering English source into “natural” Japanese word order prior to MT training & utilization. loc Working on the quality of source content is paramount to get good #MT results. -> Not only MT, it’s key to every translation #tmsinspiration loc Great lesson to learn about #mt DIY products: They’re not out-of-the-box sw. Actually each implementation is different #tmsinspiration #tmsinspiration
  • @manuelhrrnz PangeaMT depends on LSP user knowledge of MT – good solution for @b2sargent ‘s Assiduous Assemblers? andrzj nedoma/gustavolucardi PangeaMT increased translation output in automotive field with MT from 400 to 900 words per hour! #tmsinspiration
Next time you think languages, think Pangeanic
Translation Services, Translation Technologies, Machine Translation