Interview with Alex Yanishevsky: Forward-Thinking NMT & AI Language Solutions
Advances in neural machine translation (NMT) and artificial intelligence (AI) in language solutions are enabling more translations previously not possible due to cost, quality or time delays. According to a recent Gartner report, Market Guide for AI-Enabled Translation Services – “Many enterprises are not aware of the emerging AI-translation services and their capabilities and as a result, are not achieving the savings and efficiencies these might offer.”
In this special interview, Alex Yanishevsky, Senior Manager of AI, MT, and NLP Deployments at Welocalize shares his thoughts on the progress of NMT and AI in language solutions and how Welocalize are collaborating with ground-breaking MT technology providers to bring business benefits to global brands and their content strategies.
What’s the current industry climate with MT and NMT?
NMT has established itself as the predominant technology in language automation. It’s exciting as we’re seeing advancement at breakneck speeds around NMT and more widely, in natural language processing (NLP). Changes in the way we approach NMT (from RNN, recurrent neural networks, to transformer models) give even better quality and performance. The three major MT providers – Google, Microsoft, and Amazon, have engines built on these transformer models.
Sounds pretty cool if not slightly complicated for non-techies! From a real-life usage perspective, are brands embracing NMT to translate high volumes of content?
In general, MT has become to be so good that we’re at the point where global brands almost phrase that question in the negative – when should you not use MT?
Apart from specific cases that require transcreation, where you need a literary translation, MT performs exceptionally well. And NMT’s excellent performance has opened up the floodgates of running virtually everything through MT, either raw MT output (gisting) or adding a post-editing step to further improve final quality.
Part of the success story for NMT is that now it’s been operationalized by the large enterprises, we’re starting to see mid-level and smaller companies ask for MT. Like translation memory, people don’t question whether it is part of the workflow, they just expect it.
Welocalize is an early-access partner with the three big MT providers – Amazon, Google, and Microsoft. With Amazon, Welocalize has recently worked to test and pilot the new Active Custom Translation (ACT) methodology of Amazon Translate, which works to customize MT engines and give brands more accuracy and control over their MT output.
Why is MT engine customization so important?
Customization has a dramatic impact as it makes MT output more relevant than generic engines – you tailor an engine based on industry, product, and content type. If a brand is putting content through a customized engine for gisting purposes, the raw MT output will be better. With gisting, you only have one chance to put your best foot forward, so you want to get it right.
If we can get data from a client or mine data, we can build them a customized engine, tailored to their industry, brand, and corporate style. For post-edited MT, using a customized engine means the output is nearer to the desired quality, so linguists are more productive in post-editing. This saves time and costs for everyone involved in the supply chain.
Are there any sectors or verticals where you think engine customization is particularly beneficial?
In sectors where the source content is a bit more controlled or tightly written. For example, online technical support documentation, technical manuals for Life Sciences brands, finance content. Content that is more formulaic and less free flowing.
How does Welocalize collaborate with the big NMT engine providers such as Google AutoML, Microsoft Translate, and Amazon Translate?
Because we’re an early access partner, we get to see the products in beta form before they go to general availability – experiencing the product before it gets rolled out. We get to test it, to validate its usefulness, and report on any bugs. It’s our chance to get a glimpse into the future and more importantly, to affect change. This gives us the opportunity to influence the product road map.
How does this benefit Welocalize clients and global brands in general?
Because we’re using these ground-breaking engines daily in production, we can actually test what efficiencies are going to be gained by us and the whole production life cycle for our clients, whether we’re deploying a single or multi-engine strategy. Welocalize is engine agnostic – we’re neutral – we work with best-in-class technologies that will meet each individual client’s requirements. Testing and piloting the latest technologies isn’t an academic exercise; we use ROI calculators, auto-scoring algorithms, and human evaluation to predict performance when we start running millions of words through a program. We get to know these products inside out, their strengths and weaknesses.
Let’s talk AI. What role does AI play in NMT deployment?
We’re living in the golden age of AI, and every day we wake up with the idea that, wow, we can do more and more with AI. Machine learning algorithms and deep neural networks give us the ability for greater insights and predictability.
The reason Welocalize is positioned as an AI-enabled company is because AI drives decisions wherever there is value to clients. When we talk about NMT deployments to help enterprises translate higher volumes, AI-enabled linguistic processes are really at the core of all of them. When we access client data, we curate it to decide what data is relevant and not relevant to train and customize the MT engine – and that’s all done with AI. Any predictive analytics that allows us to forecast what the program looks like is also built on AI. In fact, AI plays a role in SmartLQA – rather than randomly choosing sentences to vet, we’re actually informed by our machine learning algorithms what we should be reviewing. Some of Welocalize’s SmartLQA initiatives are linked to our performance linguistics product that allows us to evaluate language on a website and build predictive models that forecast content performance for digital marketers. We’re also starting to pilot and use AI to detect sensitive content such as hate, offensive and non-inclusive speech, which is an extremely topical issue for large global brands.
AI is in places you wouldn’t imagine.
We have a lot of interesting MT-enabled solutions that we’re working on where AI is the backbone. All of them are using machine learning and will dramatically impact the multilingual content lifecycle for Welocalize clients.
Based in Boston, Alex Yanishevsky is Senior Manager for AI, MT, and NLP Deployments at Welocalize.
For more information on Welocalize MT & AI-Enabled Solutions, contact us here.
