​AI Machine Translation and Terminology Consistency

The importance of terminology

All our banking and legal clients care that translations of their documents use the correct terminology and glossary terms. After all this is understandable, it comes across incredibly unprofessional when a piece of text misspells key person or company names or other important keywords. Ensuring important terms are translated consistently throughout all documents is the job of translators and their editors.

Translation glossaries

These days translators possess and manage various glossary lists to make sure key terms are translated consistently based on the document type and audience. Many years ago, companies would paid experienced translators a premium because glossary management and consistent usage was for the most part performed manually in one's head through. experience. In recently years, software made it much easier for translators of all levels to stick consistent glossary usage via highlighting and suggesting glossary terms based on the original texts being translated. For example, if I am translating the sentence 'Microsoft is a great company.' into Chinese, the computer assisted translation tool (CAT tool) will likely see 'Microsoft' in the text and suggest the translation '微软'.

Sometimes muddled, it's important to note translation memory is not the same as translation glossary. Translation memory is a dictionary of sentences (a string of words) and their translations in different languages (see my other post). Translation glossaries are on the other hand, a list of specific translations for particular names, keywords, and abbreviations. They are different because the technology to best benefit from each of the two are different.

Neural machine translation is a different technology

In the past, translators used CAT tools that are statistical and/or rule based. All translations are created from scratch or based on a translation memory. This generally required more effort on the part of the translator: glossary terms are suggested but need to be verified individually by the translator while they translate each sentence. Recently, the technology has moved towards a much smarter type of (AI) technology called neural machine translation -software mimicking how human brains learn. With this new technology, translators no longer need to create translations from scratch. The machine automatically translates the text and translators instead spend their more time post-editing or reviewing machine translated text. One big problem with neural machine translation, however, is that in this new paradigm the process no longer incorporates human glossary selection and verification during the translation process.

This is a big issue

Some say neural machine translators are too smart for their own good. Neural machines translate texts according to the best translation predicted based on the source sentence provided. These machine don't take into account they may have translated a particular keyword (such as 'revenue' or 'expenses') one way in the past and a different way in another example. While in these cases the translation may still be grammatically correct, it is inconsistent and not what humans are accustomed to. Without say, translation of specific names are even a bigger problem for neural machine translation because the machine has no reference for words symbolizing entities created by humans.

Knowing the importance of this issue, some of the major neural machine translation providers (such Google Translate & Bing Translate) offer the option to include glossary terms along with the source sentence. At TranslateFX, we were excited to learn that someone had already solved our customer's problems but only later came to find that their results (and those of many other providers) fell flat delivering the quality results we and our clients expect.

We built an AI model to target this problem

At TranslateFX, our focus on the financial and legal translation industry allows us to focus our energy on a narrower set of problems compared to our competition. We spent a good half year developing a proprietary AI solution of our own (standing on the work of many others of course) allowing us to improve the consistency of our neural translation machines when provided with a glossary. Perhaps I'm too proud here, but the first version of our solution permits minimal reduction in the quality of our translation, an issue plaguing most neural machine translation engines in the market offering the option to include glossary terms.

If you want to check out the the results of our machine translations while keeping glossary terms consistent, please do reach out to me at erik[at]translatefx.com