How Moravia Approaches Neural MT

How Moravia Approaches Neural MT

Despite some early success and rapid advancements, Neural Machine Translation (NMT) is still in its early days when it comes to deployment in a localization program.

NMT is an approach to machine translation in which a large neural network is trained by deep learning techniques. A radical departure from phrase-based statistical translation approaches, it predicts the likelihood of a sequence of words.

And it’s hot. The localization industry has been ablaze with the topic since the seminal Google neural MT paper that was published in 2016.

Amazon, Facebook, Microsoft, Google and IBM have all been researching and/or developing NMT for internal or public use (or both). Recently, Microsoft released a customizable neural MT engine, and Google did the same just a little while later. Some of these big NMT providers are even moving away from statistical-based MT in favor of neural—potentially leaving clients a little lost if they do not have a transition plan.

Moravia believes NMT is incredibly exciting and has a lot to offer global businesses, but we’re as pragmatic as we are optimistic. It’s not as simple as swapping-out your statistical engine for a neural one, or implementing a neural engine without considering all your options. It’s time to slow down a bit. Why?

Some pros, some cons

Industry analysts don’t think that NMT is a silver bullet. For example, Slator’s report summarizes the current advantages of NMT well: it’s more fluent, it makes better translation choices, it can automatically switch between languages in one body of content, and it reduces post-editing effort.

On the other hand, there may be accuracy errors, it doesn’t capably handle ambiguous or creative language (irony, metaphors, etc.), it struggles with longer sentences, and terminology may be inaccurate. (One of our NMT experts believes that the next advancement will be fixing the terminology consistency issues.)

As for post-editing, the effort may actually compound because, unlike statistical machine translation that often had glaring accuracy issues, editors will have to sift through the otherwise fluent NMT output to find the inaccuracies.

Moravia agrees with Slator’s bullish perspective on NMT: “By now, the generic praise heaped upon the new technology is becoming repetitive: it outperforms statistical machine translation (SMT), it is a genuine breakthrough in AI tech, and it is fast-paced in terms of research and deployment. The industry is well past discussing the emergence of NMT. Clearly, neural is the new black. Now the main concern is to see if you look good in black.”

(Read the detailed Slator report on the present and future of neural MT here.)

NMT at Moravia

Our clients are faced with massive volumes, blinding speeds, continuous releases, smaller batch sizes, and multiple communication channels—and they need a solution.

Neural MT may be a part of that solution—but maybe not. NMT is evolving nearly daily, and Moravia thinks the results are promising. It’s probable (we’re projecting) that neural MT will be part of every maturing globalization program in the near future, but it may not be sufficient. It won’t ever stand alone as the entire solution.

Websites abound that suggest that NMT is ready for prime time for everyone, that it’s a standalone service, or that it’s the right solution for a business who wants to translate more content faster.

But at Moravia, we help businesses look at any and all tools that help overcome challenges. It’s about finding a best fit solution, comprised of any tools that help resolve a business problem, satisfy requirements, and provide ROI. And what’s more, even if you choose the best tools available, they cannot be deployed in isolation; they need to be operationalized, integrated with each other, and built into a workflow that considers project management, quality requirements, rounds of reviews, and many other project components.

We believe MT, whether statistical or neural, should be part of an ecosystem of tools that meets a business’s goals. Our tech experts know the MT landscape—including all industry-standard tools—and our clients are confident that we will take a balanced view when it comes to implementing MT technology. And besides, if you already have an MT program in place, you will need a transition plan and some expertise in order to switch to a more favorable engine—whatever that new engine may be.

When it comes to NMT, we’re researching, piloting, and getting our hands dirty. We focus our technological expertise where it will make the biggest impact for each client. And we’re keeping our finger on the pulse of the needs of our clients. We deploy NMT when it drives ROI and makes good business sense.

Our advice to you? Watch, wait, learn, and involve an MT expert to help you asses the use of MT, including neural, in the context of your localization program.

 

Comments