25 Jan /18

Investing in Neural Machine Translation Capabilities

Investing in Neural Machine Translation Capabilities - EVS Translations
Investing in Neural Machine Translation Capabilities – EVS Translations

With the ongoing buzz surrounding neural machine translation, it can be very tempting for companies who require translation services on a regular basis to build their own in-house neural machine translation (NMT) capabilities. Logically, it makes more sense to build out your own capabilities than to pay someone else for the same service, until, of course, it doesn’t.

Before any company considers a move like this, it is important to take into account 3 particular aspects: resources; the learning curve; and the benefits of specialisation.

Naturally, the question for any business is, “How much will we need to invest before we see any return?” Setting up an NMT operation requires a substantial upfront investment of resources. In addition to the physical needs, such as space and graphics microprocessors, NMT will also require a significant investment in time, data, and training the artificial intelligence before producing any results.

And just because it is producing results, this doesn’t mean that it is producing the correct results according to your business. As with anything else and especially with NMT, there is a learning curve, and although it utilizes artificial intelligence, it can only utilize it according to its programming. And while there is a growing range of software tools to build and manage neural networks, the same does not apply for tools for building and verifying the accuracy of language data. NMT needs big amounts of data, and we really mean big, with training and terminology coming only from publicly available documents, like for example the European Union’s highly multilingual parallel corpora, it will take significant time for any in-house NMT to adapt to a particular company’s usage and terminology. Furthermore, the training applied to the NMT vocabulary, structure, and rules of one, single target language means that each time a company needs to add a new language, it will require new, specialized training and naturally, the amount and accuracy of corpora of publicly available training data will vary per language. Furthermore, for example, agglutinative languages are quite hard to be processed for NMT engines, compared to Romance and Germanic languages.

Finally, and most importantly, comes the aspect of specialisation. Though NMT is a great advancement from statistical machine translation, it is not perfect and will neither eliminate all the shortcomings of the previous machine translation models, nor produce perfect translations or solve all your company’s translation needs. Even if your company invests the necessary resources to build an “ideal NMT,” the translated output will still require the review of a post-editor and proofreader in order to assure formatting, style, and accuracy, along with your brand- and industry-specific terminology and style guides.

As developers are still working on solutions how to further improve the accuracy of NMT, your best bet is to keep entrusting your translation needs to a reputable language service provider, who, ideally, has already invested in NMT and has taken the guesswork out of the learning curve of AI.

We, at EVS Translations, for example, automated all our corporate processes and workflows into one state-of-the-art data management system, and our in-house translation technology department continued ahead with investments in neural machine translation solutions.

Translating over 50 000 sentences daily into tens of languages for over 25 years now, we do have the Big Data in-house.

And furthermore, servicing leading companies and bodies from all industry sectors, our Big Data, structured in terminology specific clusters, is specialized and provenly accurate.

We do have the technological and linguistic tools to regularly train our NMT capabilities, along with the expert personnel to precisely meet all your language needs.