XTM cloud was the first to implement NFA with Systran, for improving the MT by on-the-fly training of the inference model with a relevant fuzzy matched target. Some friends have asked me to explain the NFA, Neural Fuzzy Adaptation. I can show how easy it is to use it, by example. Translate from English: "I like nfa, because it is simple and beautiful." I intentionally kept nfa in lowercase, as I want to show one thing that it can also do: fix the capitalization Another thing is help in using gender. In English it's just the NFA. But in French and other languages, is it le or la NFA? In German, is it der, die or das ? (masculine, feminine or neuter) The sentence doesn't have much content to help determine such little details. But NFA can add to it. When going without NFA, the logical default is to pick gender = masculine, so "le" NFA in French. It will likely also keep the acronym as is. Sample translation without NFA: output: "J'aime le nfa, parce qu'il est simple et beau" The gender is clearly masculine on this French result, as expected: il est beau. If it were feminine, it would change to: elle est belle. (she is beautiful). The API call for something like this might look similar to this: https://$serverurl/translation/text/translate?key=$apikey&target=fr&input=I like nfa, because it is simple and beautiful Other options can be added such as the source language (auto-detected in the above example) and many more, including back-translation to see if anything got lost in translation. Now, when you're adding NFA, you simply provide a fuzzy matched target from another sentence that's relevant in your translation project, as found in your translation memory, one that is validated and therefore properly spelled and capitalized. Such example of another sentence and its validated translation might be: English: We use NFA since last year. French: Nous utilisons la NFA depuis l'an dernier. Simply add the extra target to your API call: ....&fuzzy=Nous utilisons la NFA depuis l'an dernier. Note: The new server (v10) also can do this for you now, if your TM resides in the server. The resulting call: https://$serverurl/.....?....&input=I like nfa, because it is simple and beautiful&fuzzy=Nous utilisons la NFA depuis l'an dernier. And the new translation immediately fixes the capitalization of nfa to NFA, and sees the gender as "la" NFA, feminine, and consequentially changes the adjectives. output: "J'aime la NFA, parce qu'elle est simple et belle." Comparison: before NFA: "J'aime le nfa, parce qu'il est simple et beau" after NFA: "J'aime la NFA, parce qu'elle est simple et belle." Clearly a better translation on several levels (capitalization and gender): nfa --> NFA il --> elle beau --> belle There are roughly 10-12 tokens in that output, and with 3 words improved by NFA, you might argue that this improved the translation by 25%. Not a shaby ROI if you ask me. You might enjoy some of this at XTM Live
What are the XTM Live Sponsors looking forward to the most at this year's event? LanguageLine Solutions Language Inspired Argos Multilingual SYSTRAN by ChapsVision Vistatec ContentQuo Find out below and get your ticket here to join them! 👉 https://lnkd.in/g5gQ_GVZ 👇
