Blog chevron right Translation

Neural Machine Translation: The Rise of the Thinking Machines

Michael Gallagher
Michael Gallagher
Posted in Zoom Sep 6 · 8 Sep, 2022
Neural Machine Translation: The Rise of the Thinking Machines

There was a time when a machine translating one language to another in real-time was unthinkable. People would most likely laugh at you if you brought up the idea of a “thinking machine.” It’s absurd and is nothing more than the product of craziness. 

However, bring up the topic today, and you’ll be surprised with the responses. Sure, some folks will still find the notion of a machine translator far-fetched. 

And you’re in luck if you consider yourself one of them. This article will share what Neural Machine Translation (NMT) is and how it revolutionized how we do business globally. 

An Overview of Machine Translation

Machine translation uses artificial intelligence (AI) to translate one language to another, requiring no human input. It has modest beginnings, spurred by the need to understand different languages without relying on human interpreters.

In 1834, Charles Babbage imagined having a programmable machine translating different languages. A dozen decades later, IBM introduced the world’s first “electronic brain” that translated Russian into English in real-time. 

The 701 Translator processed Russian statements and phrases from participating Georgetown University Institute of Languages and Linguistics linguists within two to three seconds and printed them at 2.5 lines per second. That’s impressive for a 1950s computer. 

Although the 701 Translator is a dinosaur by 21st-century standards, it was revolutionary. 

Unfortunately, it never caught fire. Only a few people took the challenge to “train” the machines, requiring thousands of hours manually defining and programming a dizzying set of rules. It would take about half a century more before statisticians came into the picture and provided a more efficient way of translating languages. 

It would be best to understand the two machine translation types serving as predecessors to modern NMT.

·        Rule-based Machine Translation

You might want to consider this type of machine translation the grand-daddy of neural machine translation. The world no longer uses this technology because of its tedious and time-consuming processes.

Rule-based machine translation requires programmers to work with linguists specializing in a target language. For example, developers can only work with Russian language experts if they want to translate Russian into English and vice versa. Programmers cannot rely on Chinese or Japanese linguists, regardless of how impeccable their English is. 

Programmers and linguists examine and analyze grammar structures to establish “rules” for phraseology, word order, sentence structure, and other critical language elements for both the source and target languages. 

Linguists then provide developers with the “rules” to create a program for mapping the translation, word by word. The computer must analyze the source word based on the programmed “rules” and translate that into a target language using the “rules” for that language.

It’s like physically looking up the meaning of a word in a dictionary. The only difference is it’s a computer that searches. 

·        Statistical Machine Translation

Rule-based machine translation (RMT) might be inefficient, but it spurred developers into thinking outside the box. You can think of RMT as the necessary foundation for future neural machine transmission.

Statistical machine translation (SMT) is the crucial bridge between RMT and NMT. It utilized existing multilingual corpora and translations to identify and describe patterns. We can look at these RMT-based translations as data points in a massive research study. It’s like plotting each one on a chart to define statistical patterns.

Linguists and developers used statistical models to program a machine translator that predicts similarly constructed texts. Like RMT, statistical machine translation demanded immense resources to “train” the computer, requiring millions of words for a single domain. 

Nevertheless, the output was more impressive than RMT-based results, particularly in scientific and technical texts. Although SMT started with word translation, it evolved into phrases and syntax reflecting the word’s context. Here are the statistical machine translation “forms” you might want to know.

Word-based Translation – This SMT type translates texts word for word. Unfortunately, idioms, morphology, and compound words can put a translated text out of context. For example, translating the word “corner” to Spanish can produce “esquina” or “rincon.” Unfortunately, these words might have different implications depending on the context.

Phrase-based Translation – A more advanced SMT than word-based translation, phrase-based translation minimizes WBT restrictions by translating whole word sequences or phrases. The first Google Translate version, launched in 2006, is an example of a phrase-based machine translation algorithm. If you can remember, Google Translate wasn’t that accurate in its early years because PBT does not focus on linguistic phrases. Hence, you’ll get a weird translation. Regardless, PBT remains a popular SMT today.

Syntax-based Translation – This SMT has been around since the 1980s, although it didn’t gain ground for another decade. It focuses on translating syntactic units instead of phrases and single words.

Hierarchical Phrase-based Translation – Combining the strengths of syntax-based and phrase-based machine translation, HPBT focuses on synchronous context-free grammar rules. Chiang’s Hiero system is an example of hierarchical phrase-based translation.

These programs lorded the language-translation world at the beginning of the 21st century before giving way to a more advanced translation system more than a decade later.

Neural Machine Translation

The Rise of Neural Machine Translation

Although rules-based machine translation and statistical machine translation can translate multiple languages into more readable and understandable forms, they are inefficient. Programmers spend countless hours designing the code forming the digital translator’s backbone.

Google sought to improve the phrase-based machine translation system, Google Translate, by leveraging Recurrent Neural Networks (RNNs). 

The tech giant recognized the PBT’s limitations in accurately translating sentences from one language to another. The PBT system breaks down the text sequence into words and phrases, translating them independently. That’s why we have had questionable Google Translate results in the past.

Recurrent neural networks don’t dissect the word sequence to produce an accurate translation. These “digital brains” look at the entire input sentence as a translation unit – not words or phrases. It’s worth pointing out that RNNs paved the way for the rise of neural machine translation. 

In 2016, Google relaunched its Google Translate system as the Google Neural Machine Translation system. The new technology utilized highly advanced training techniques to improve translation speed and accuracy while minimizing the need for design and engineering choices.

It’s easy to recognize the difference between the 2000s Google Translate and the modern version. The translation accuracy is more spot-on. It’s not perfect, but the program is more credible than its predecessor.

So, how does neural machine translation (NMT) work?

It’s best to appreciate a neural network before we delve into the inner workings of NMT. You can think of a neural network as a digital representation of the human brain. 

Our brain has nerves consisting of a circular cell body, multiple short appendages called dendrites, and a single extra-long axon. If you look at a single nerve cell, it looks like a node.

Imagine several “nodes” transmitting and exchanging information with other “nodes.” This structure is similar to what you’ll see in a neural machine translation system. When you input a source sentence, the message goes through the network of nodes for processing and translation. The output? You’ll get a target-language sentence specific to the source sentence.

The advantage of neural machine translation over other machine translation systems is it can process extremely large datasets to produce highly-accurate target-language translation at blistering speeds and without human intervention.

The beauty of neural machine translation lies in its ability to translate a source-language sentence into a precise target-language sentence. It relies on a well-defined encoder (source) and decoder (target) networks. 

However, it’s worth pointing out that NMT’s power depends on the integrity of its neural network. Like the human brain, the neural network architecture defines how efficiently and accurately the NMT can process a sentence for translation. 

Neural Machine Translation Benefits

Neural Machine Translation

Neural machine translation is an efficient and accurate system for translating different languages. Like the human brain, NMT is also capable of learning and adapting. 

It can process vast data and adjust to new contexts. Unsurprisingly, many organizations today employ NMT in their translation requirements, empowering them to deliver accurate language-sensitive content quickly and flexibly. The following benefits await businesses that use NMT.

·        High Accuracy

Companies want 100% translation accuracy to make brand messaging more precise to their target customers. We don’t need to look for examples other than Google Translate. The contemporary Translate is more accurate than the Google Translate of the first decade of the 21st-century. We have neural machine translation to thank. 

However, it’s worth pointing out that an NMT system’s accuracy depends on the chosen engine, available training data, language pair, and type of text for translation. An NMT system becomes more accurate with more translations it performs on a specific language or domain. 

Hence, it’s unsurprising to see Google at the forefront of NMT accuracy because of the number of translations it performs on Google Translate. Other noteworthy NMT engines include Amazon, DeepL, and Microsoft.

·        Fast Learning

It took developers a few decades to produce a better alternative to the 701 Translator of the 1950s. The rule-based machine translation system improved the processing or “learning” speed. Unfortunately, it wasn’t enough. The statistical machine translation framework provided a faster way of training the program. 

However, NMT is much faster because it relies on fully automated processes to reduce the training time. Hence, companies using NMT can depend on a quick-learning system to adapt to rapidly changing times. 

·        Scalable

Businesses today have ever-changing needs that reflect the ever-changing times. Your company might have modest translation requirements today, but you can never be sure what tomorrow’s translation picture might be. The good news is that NMT allows you to upgrade or modify it to meet growing demand.

·        Cost-efficient

Did you know that professional translators charge an hourly rate of $30 to $70 depending on the subject matter, language combination, required turnaround time, and volume? It’s not a problem if you only need to translate a few hundred pages of written material. It becomes costly if you’re running thousands of documents in different languages. 

Using neural machine translation saves your company money. You can leverage the system’s highly accurate and scalable properties to optimize your budget. Of course, you can still employ human translators for post-editing to guarantee absolute translation precision. 

·        Flexible and Effortless Integration

One issue with SMTs is their complicated integration requirements. They only work with limited platforms, restricting your business. On the other hand, NMT provides effortless integration into SDKs and APIs, empowering you to utilize these libraries for various applications.

·        Customizable 

It’s painstaking to customize a statistical machine translation system. Meanwhile, NMT makes customization a breeze, empowering you to tailor-fit the program to suit organizational requirements. Companies can integrate brand-specific glossaries and industry-related terminology databases to update the NMT model. 

Neural Machine Translation Applications

Neural Machine Translation

As beneficial as neural machine translation is, you’d still want to know how you can apply this technology in your business. Let’s look at several ways.

·        High-volume translation with quick turnaround requirements

One of the most significant advantages of NMT over SMT is its precision and speed. It can process large volumes of data in seconds, allowing organizations to complete their tasks faster than other machine translation systems.

For example, natural disasters leave chaos, confusion, and despair. Relief organizations must work round the clock to deliver much-needed services. 

Unfortunately, some people might not understand the disseminated information because of language issues. NMT can help fast-track the translation of large volumes of data, allowing organizations to keep people in disaster areas updated on the latest developments. 

Companies can also leverage NMT in translating product reviews by non-English-speaking customers. For example, Mexican, Chinese, and Japanese customers might leave reviews on an Amazon or eBay product page. Unfortunately, other customers will not read these reviews because they’re in a language they don’t understand.

NMT can help by translating these customer reviews into English and posting them on the platform before potential customers decide to look for another brand. 

·        Highly repetitive content

Contemporary customers expect brands to provide supporting documents for every product they buy. Hence, the refrigerator or TV you buy always comes with an owner’s manual and user guide. Some brands also offer service manuals, troubleshooting guides, and other reference materials.

Unfortunately, manually translating these materials into each major language is time-consuming and laborious. Can you imagine translating hundreds of pages of reference materials from English to Mandarin, Hindi, Spanish, French, Arabic, Bengal, Russian, and other principal languages? 

Neural machine translation simplifies the translation process, empowering brands to speed up their production and delivery of offerings. The system also hastens the creation of reference materials and other repetitive content. 

·        Social sentiment analysis or user-generated content

People are engaging brands with greater tenacity and conviction than previous generations of consumers. They leave comments on product pages, official websites, social media platforms, and other internet places. Modern consumers are braver than ever to voice their opinions, concerns, and sentiments.

Brands want to know how customers feel about their offerings. And they can only do this by reviewing and analyzing user-generated content. Unfortunately, not everyone speaks the same language and it would be tedious to translate volumes of reviews manually. 

Neural machine translation offers a more efficient way of transforming language-restricted user-generated content into meaningful, measurable data. It provides brands a clearer picture of consumer sentiment, allowing them to fine-tune their offerings and revise marketing strategies to fit localized markets.

·        Online customer service

Global customers expect their favorite brands to communicate in a language they know. Although translation systems can provide language-specific reference materials, there are situations when customers would prefer having a customer service representative addressing their concerns.

Try to imagine an Arab customer complaining about a product they cannot get to function normally. A helpdesk with an integrated NMT can resolve the issue quickly and precisely. Customers will never second-guess the technician’s instructions because they understand the language.

Neural Machine Translation Outlook

The future looks bright for NMT. Experts say the continuing push towards a more robust artificial intelligence can only make NMT relevant to future organizations. Developers can continue working with linguists to refine existing NMT architecture, allowing for the more accurate translation of certain communication styles and nuanced expressions.

Although it’s safe to assume that NMT can replace human translators, there will be instances where humans are better than machines. Like other things in life, language is ever-evolving. The communication styles we use can become obsolete tomorrow, rendering NMT less effective in processing newer expressions. 

Hence, it’s sensible to think that NMT will thrive alongside human translators. The “thinking” language machine will provide scalability and flexibility to future organizations, while human translators will focus on creativity, nuanced interpretation, and critical thinking.


It might have modest beginnings, but neural machine translation is the most advanced translation technology in the modern era. Companies can leverage NMT’s high accuracy, flexibility, scalability, adaptability, cost efficiency, and customizability. 

It allows them to meet strategic objectives. Realizing globalization, internationalization, and localization goals is a cinch with an NMT providing the backbone for high-precision content.

We can help you get the best transcription solution if you want to deliver highly accurate translated content to your customers. We can be partners in your ongoing pursuit of excellence in the global market.