Localization Leaders: Meet Amazon’s Watson Srivathsan

Atara Braun

August 5 , 2021 · 9 min

Meet Watson Srivathsan, Product Manager at Amazon AI. We spoke to him about the groundbreaking work he’s doing at Amazon to improve their NMT and make it on par with professional human translation.

Can you tell us a bit about your background and role as Product Manager at Amazon?

I grew up in Chennai, India. I had the opportunity to spend my formative years in Dubai, where I made friends with kids of expats from different countries and got my first taste of life outside of India. I earned my Bachelor of Technology from the Indian Institute of Technology in Madras and my Ph.D. in Mechanical Engineering from the University of Minnesota. After working as a scientist in the field of Mechanical Engineering for 8 years, I wanted to switch to Product Management (PM). I pursued my MBA from the University of Chicago Booth School of Business.

The first product I worked on as a PM was Windows 10. After spending 6 years at Microsoft, I joined Amazon as the Product Manager for Amazon Translate in 2019. At Amazon, my goal is to create the most customer-obsessed machine translation solution.

Can you share a bit about what your work as Product Manager on Amazon’s neural machine translation (NMT) engine entails?

I am responsible for setting the vision, outlining the strategy, and delivering features and services for Amazon Translate to achieve customer satisfaction. Our customers are businesses, enterprises, and localization service providers who need fluent, accurate, fast, and cost-effective machine translation to connect partners and stakeholders across languages. We listen to our customers and work backward from their needs to bring the full benefits of artificial intelligence (AI) technology to machine translation. To understand how AI benefits machine translation, we need to understand why professional translators perform better translation than machines. Professional translators understand the context better than traditional machines. They don’t just translate the word or the sentence, they understand the sentiment and the circumstance around the sentence to provide an ideal translation. AI has shown great results so far in providing fluent translation by taking the context of the sentence into account, and in the future, we expect AI to assist professional translators to perform their jobs more efficiently.

You mentioned (previously) how NMT has made significant progress over the last few years in closing the gap between human and machine translation. To what is that success attributed?

Amazon Translate’s NMT uses deep learning models to deliver more accurate and more natural-sounding translations than traditional statistical and rule-based algorithms. This technology goes beyond simple word-level interpretations by understanding the context of the words it’s translating. Let’s say you want to translate a simple word—“crane”—in a sentence. Amazon Translate will understand whether you meant a species of bird or if you meant a crane used for construction to disambiguate sentences like, “The crane flew over the construction crane,” where the words “flew” and “construction” provide context clues for the word “crane”.

NMT becomes even more important when you want to translate an idea across language families. A language family is a group of languages related through descent from a common ancestral language. Examples of language families are Indo-European, Sino-Tibetan, Afroasiatic, etc. Across language families, the sentence structure, the available spectrum of word choices, and the way you write usually vary a lot. The best way to translate between two languages that belong to different families is to understand the meaning in the source language and recreate the same idea in the target language.

The old methods of statistical and rule-based translations sometimes don’t provide a fluent output. I am going to give you an example that involves English and my native language, Tamil. Say I want to translate “Get that book from Watson” from English to Tamil. Amazon Translate’s NMT output is “வாட்சனிடமிருந்து அந்த புத்தகத்தை பெறுங்கள்”. Focus on the first word வாட்சனிடமிருந்து. In English, the noun (Watson) and the preposition (from) are two separate words. Whereas, in Tamil, the preposition changes the ending of the noun. The noun “Watson” (வாட்சன்) gets modified as (“வாட்சனிடமிருந்து”) “from Watson”, a single word, to be translated correctly in Tamil. This is possible with NMT as it reads the whole English sentence to translate to Tamil.

While NMT is great at inferring the context by reading the whole sentence, it should also be “intelligent” at identifying a word salad. If I try to translate “shout cat umbrella” from English to another language, it should identify the input as a list of distinct words and perform translation word for word. NMT has provided very promising results so far.

AI has shown great results so far in providing fluent translation by taking the context of the sentence into account, and in the future, we expect AI to assist professional translators to perform their jobs more efficiently.

Seems like great progress. But why now?

The AI revolution is happening now because the computing power available per adjusted dollar has increased by more than one million times over the last 25 years. Second, cloud technology has brought down this cost further. In the past, companies that wanted to do machine translation had to purchase the computing hardware up front, which is a costly investment. Today, if you use the cloud, this upfront cost is zero and you only pay for the compute resource that you consume. Cloud technology allows you to trade capital expense for variable expense, paying for IT as you consume it. This has reduced the cost barrier to entry to perform large and complex calculations. This lower cost to entry has enabled many machine translation scientists to try new experiments, thereby stimulating innovations to improve the quality of machine translation.

You also mentioned how you see the current juncture in time as an inflection point for NMT engines – why is that? And what next?

We are in the early stages of NMT evolution, an inflection point. The cost of computing is coming down. Cloud-based technology has brought the upfront cost down to zero. We are investing in a lot of experiments to make translation more accurate, fluent, contextual, faster, and more efficient. There was a point in our past where the high standard was complete manual translation and translation experts were unsure of how computers could contribute to translation. Today, NMT is state of the art. Localization leaders and large enterprises are switching from older translation technology to NMT to scale their operations efficiently.

Translation is an intensely creative process of conveying an idea in one language to another. So, in the past, many professional translators preferred to translate on a blank slate. They didn’t believe that computers could aid in any creative work. Today things have changed; AI and NMT have made enormous progress in providing contextual and fluent machine translation. For materials that don’t require artistic input—such as news, technical documentation, product description, regulatory filings—professional translators now start from NMT output, review the translation, and make corrections to the output as necessary.

Even for materials that need creative input—such as marketing materials—NMT has proven to aid and enable the creative process of the professional translator in their pursuit to convey the message powerfully in a new language. Think of it like a stencil in the hands of an artist. The translators are able to increase their productivity and save time. Moreover, today, you can even customize your machine translation to provide different outputs that suit your domains such as healthcare, manufacturing, and legal translation. NMT and machine translation-post editing (MTPE) is becoming the norm. This is possible only now because MT quality powered by AI is more accurate than ever before – and we are just getting started.

Think of it [NMT] like a stencil in the hands of an artist. The translators are able to increase their productivity and save time.

While machine translation today is remarkable in its capabilities, it’s still not as good as human translation. What’s holding us back?

One of the many reasons why machines are still behind humans in translation is that people are always evolving in the way we speak. We are always inventing and discovering new things, which get added to our lexicon. In the case of English, we don’t speak English the same way it was spoken during Elizabethan times. Furthermore, one language can evolve in different ways in different geographies, and when the two variants meet again, people adapt and alter the way they express words and concepts to accommodate variants. We also come up with new ways to express ideas; say for example, how we use emojis. Thus, as people lead this change, professional translators are faster than AI in keeping up with this change, leaving machines trailing behind. At Amazon Translate, we are doing a lot of work to close the gap.

What are some of the ways in which machine translation has changed the localization industry?

Machine translation, especially NMT, which is more accurate, fluent, and contextual, has helped a lot in reducing language barriers. Today, if I have to read from a new website or write to a person who doesn’t share my language, I can do that using machine translation. In fact, our customers are using our machine translation for localization, information analytics, and inter-language communication. Especially in localization, Amazon Translate introduced a new product called Active Custom Translation (ACT) for localization service providers and enterprises. This uses the translator’s feedback in the post-editing process in MTPE to continuously enhance the machine-translation output.

If you are curious to get started with ACT, all you need to do is sign up for an AWS account, and follow this blog to get started. You can even try the service risk-free with the free tier. With ACT, you don’t have a need to build a custom translation model; the customization happens during run-time. There is absolutely zero cost to build, modify, and update your custom translation. Before ACT, most customers had to spend money to build a custom translation model before they could start translating. With ACT, you only pay by the number of characters you translate. You don’t incur cost if you didn’t happen to perform any translation.

Say you are providing translation services and you choose MTPE to increase translator productivity, and your translator did not like the way a particular segment got translated. With ACT, the translator can suggest a different translation and you can use it to customize the next translation job in the same domain. Today, many translation service providers use custom translation to enhance NMT output. Building, maintaining, and updating the custom translation model is a complex, costly, and time-consuming process. As I mentioned earlier, ACT doesn’t require model creation. ACT frees you from worrying about model building, maintenance, and updating. This also liberates your translators from making the same corrections to the machine translation over and over, as their feedback from the previous translation jobs will be used to enhance the next. In fact, BLEND is one of our early adopters for ACT.

Other than localization, machine translation has made an impact in interlanguage communication and information analytics. For example, many chat applications use Amazon Translate to connect folks who use different languages in communication. Large firms can also use it to listen to their customers or learn about an industry from diverse sources that publish information in different languages, and use the content to make informed decisions.

Professional translators are faster than AI in keeping up with this [language changes], leaving machines trailing behind. At Amazon Translate, we are doing a lot of work to close the gap.

What separates Amazon’s translation engine from those of competitors?

Amazon Translate is a relative newcomer to the machine translation party. In about 3 years we have grown significantly. Today we support 71 different languages, which covers more than 99.99% of enterprise translations that happen online – and we’re just getting started. Our customers have control over the data residency of machine translation, which means that, as a customer, you can specify in which AWS region you want to run your machine translation workload. This is necessary, especially for our enterprise customers who have to comply with various local and regional laws. Apart from enabling Active Custom Translation, ACT, we were one of the first to introduce document translation, where you can translate an archive with thousands of Text, Word documents, Excel, PPT, or HTML files in one go.

Are there any new features you can share with us that you’re working on?

Please tune in to our AWS What’s new post! We have a whole lot of exciting Amazon Translate launches lined up.