When I was a lot younger, I was fascinated by the idea of a Universal Translator. I first saw it in sci-fi movies. Characters could communicate with alien species from all across the galaxy. And I wondered how long it’d take for machines to be able to translate all of Earth’s languages instantly.
As it turns out, I didn’t have to wait long. Today, tools like Amazon Translate, Google Translate, and Microsoft Translator can deliver that functionality by translating text on the fly.
This article will describe what machine translation is, uncover its history, and explain how it works. We’ll also explore different types of machine translation and how they can help your business reach international audiences.
What is machine translation?
As the name implies, machine translation (MT) is a computerized process that translates text from one language to another. It does this by looking for grammatical patterns, using learning algorithms, and referring to previous translations.
There was a time when automated translations weren’t very accurate or reliable. The main problems were context and homonyms. If you didn’t know, the latter are words spelled the same but can mean different things based on the words around them. Bark is a homonym. It means one thing when used with “dog” and something completely different when used with “tree.”
The example shows why word-by-word translations don’t always work. The machine needs to identify homonyms and factor context into its calculations. It’s tricky, but technology is improving and adapting very fast. And you’ll be surprised at how far we’ve come.
How does machine translation work?
Unsurprisingly, there’s no one-size-fits-all answer to this question. While there are several ways to use MT today, most still mimic the same strategies human translators have used for centuries.
When humans translate a text, they use their knowledge of the source language to understand the intended meaning behind the words. Then, they use their knowledge of the target language to convey the meaning of the source text as closely as possible.
A computer is much more methodical by nature. Machine translation follows a predefined set of rules or steps. The first is to filter and organize the source text. Next, it references a database of similar passages translated from the source to the target language. Its objective is to spot patterns that it can use to understand which words to select and how phrases are structured.
Translating large amounts of text by hand takes much time and effort. By automating the process, machine translation aims to produce more accurate results in less time than a human can. So far, the first part has largely been achieved, but the second part is still a work in progress for many tools.
A brief history of machine translation
Believe it or not, the idea of machine translation has a history long before computers. Like many other fields, what we see today came from humble beginnings long ago.
The original idea dates back to Al-Kindi, a 9th-century polymath born in present-day Iraq. He oversaw the translation of ancient Greek scientific and philosophical texts into Arabic. The experience had a profound effect on him. Among many other achievements, he also found ways to translate languages using cryptanalysis, frequency analysis, probabilities, and statistics.
In 1629, René Descartes suggested a universal language of symbols to express similar ideas in different languages. The idea of a translating machine was conceived and patented in the 1930s. Interestingly, it was done independently by two other people. Georges Artsrouni filed his patents in France, while Petr Petrovich Troyanskii did the same in Moscow.
Warren Weaver presented the first proposals for computer-based machine translation in 1949. Since then, many others have added to this growing body of knowledge.
Different approaches to machine translation
The approaches to MT correlate roughly to different stages of the technology’s development, beginning with Weaver almost 75 years ago. The tradition of building on the body of knowledge continued. As a result, some elements of previous models survived and were adapted to enrich later approaches.
Rule-based machine translation
Weaver’s proposals in the 1950s were for rule-based machine translation (RBMT), which follows a predefined set of rules. For about 30 years, that was the only way computers could translate text.
It’s often called the classic method because it relies on linguistic rules and dictionaries that human language experts must manually create, maintain, and update constantly.
Due to its manual nature, RBMT is time-consuming to implement and maintain. It also struggles with ambiguous terms, which can result in lower-quality translations with stilted or “machine-like” output.
Statistical machine translation
Computing power increased dramatically in the 1980s, and prices started dropping. This led to a new approach called statistical machine translation (SMT). It uses statistical models and large volumes of bilingual data to translate text from one language to another.
An SMT system learns to translate by analyzing the statistical relationships between original texts and existing translations of those texts. Its output for a new translation is then put through a language model to give the translated language a more natural flow.
A big drawback of this approach is the cost. Creating the parallel data and training the system is laborious and time-consuming. It’s also less accurate when working with language pairs with differences in word order, such as English to Japanese.
Neural machine translation
The next technological advance led to the most sophisticated of the three, where artificial intelligence (AI) appears. Where previous models were phrase-based systems, the neural machine translation (NMT) model considers entire sentences at each step. This allows for discernment of context and results in far more accurate translations.
The most remarkable feature of NMT is its ability to learn and adapt to new contexts with each piece of text it translates. This makes NMT an ideal solution for quickly, accurately, and flexibly translating lots of text.
Computer-assisted translation tools
But things still needed improvement. So, an earlier development involving professionals using technology led to the development of computer-assisted translation (CAT) tools. The idea was to let technology help translators with some parts of the process.
CAT tools break up source text into smaller, more manageable segments (often sentences). Next, human translators work on individual segments, referring to the original text when necessary. Then, the tool stitches the segments back together and spits out the translated text for review.
Most CAT tools include glossaries, translation memories, and other advanced features to further streamline the process.
How generative AI changed the game
Using AI for translation is nothing new. But, older AI-powered tools needed larger volumes of data to improve their contextual understanding of language. Generative AI doesn’t need all that data. It enables zero-shot translation and powers the ChatGPT evolution by solving complex problems with limited data.
Most MT systems let you feed them glossaries of commonly translated terms, but they still require occasional maintenance and other manual updates. But that’s not needed with AI. Generative AI systems automatically add new terms and save them in a glossary.
A big problem for most translation systems is low-resource languages with less available data to learn from. But generative AI models can learn and recognize content they haven’t seen before, meaning they can translate these languages with greater accuracy.
Finding the right machine translation system
Machine translation has made great strides, especially with recent advancements in AI. But when choosing a solution for your business, it’s important to consider the balance between cost, speed, and accuracy. If you want the fastest, most accurate solution, it won’t come cheap.
MT is often good enough for dealing with individual words or something relatively straightforward, like UI content. For technical or complex types of translation and those full of industry-specific lingo, a more advanced method would be best.
The key is to find the solution with the speed, accuracy, and price point that makes sense for your needs.
Many businesses might opt for an advanced MT solution in critical or customer-facing areas and use a free tool where speed and accuracy aren’t as consequential.
The bottom line is that standard MT is a fast and cost-effective option for translating large quantities of content. However, AI-powered tools can give you more accurate, contextual, and on-brand translations.
This may help you decide which type of solutions are best for different applications:
- Complex/high-risk translations. A combination of AI and human translations, with post-editing by humans and AI-powered linguistic quality assurance (LQA)
- Medium-risk translations. AI translations complemented with post-editing by humans and AI-powered LQA
- Non-critical/minimum viable translations. MT with AI-powered LQA.
Machine translation and humans in tandem
Even before MT became mainstream, lost-in-translation mistakes were an issue for international businesses. For example, KFC’s global tagline didn’t quite work in China, where “Finger-lickin’ good” was mistakenly translated to “Eat your fingers off”. It’s not clear if that was due to human or machine error.
Either way, there’s no need to rely on one or the other. They work best in tandem. Indeed, many language professionals already offer services that combine MT, generative AI, human translators, and human post-editors. Depending on what they’re translating, some may even add special features and tools.
With tools like Lokalise AI, teams can quickly shorten, rephrase, optimize, and translate content in bulk. Leveraging AI-powered linguistic quality assurance (LQA) means human professionals can focus their time and energy on other parts of the process.
What can machine translation do for your business?
One thing that has become clear in recent years is that we live in a global economy. The pandemic showed us that we’re all interconnected and language barriers can be overcome.
Gone are the days of setting up individual websites for every language you support. The same goes for presentations, training materials, marketing campaigns, and other company literature in the languages of all the countries where you operate.
Managing all these different things was a major chore and a big reason for all the now humorous lost-in-translation problems. Today’s tech makes MT much easier to manage and get right. The best way to show that is with a couple of use cases.
Internal communication
For global teams, writing all company emails and interoffice communications in English can lead to misunderstandings—even where it’s a native language. It’s the same with other languages. The French spoken in West Africa is different from the French spoken in France, for example. And the complexities ramp up very quickly the further afield you go.
That’s where machine translation can help. It can translate your messages (or voice) on the fly to avoid disrupting the conversation. It can translate replies back into your language almost instantly. Speed is essential when communicating urgent messages globally, and you need rapid responses.
Other use cases
There are many other places where machine translation can help you, each with its own needs in terms of the balance between cost, speed, and accuracy:
- Conference calls. Speed is critical because slow translation disrupts conversations. It’s up to you whether cost or accuracy is the second most important.
- Technical and legal documentation. Accuracy is paramount, but you must have flexibility with cost and speed.
- Product instructions. Speed shouldn’t be a problem here, allowing you to find the right balance between cost and speed.
- Marketing material. This is another use case in which time is usually a minor issue. You can focus on either cost or accuracy.
- Websites. Speed is the most critical factor here, followed by accuracy. Because the pages are live, you can quickly fix typos and other issues. The speed does come at a higher cost.
Let Lokalise help you grow your business.
Machine translation is unlikely to completely replace humans for some time, if ever. But in the meantime, it can help speed up and improve the translation process.
In this interconnected world, you don’t have to learn a new language to reach new audiences. Today’s AI-powered tools and tech are breaking down borders. With the proper localization and translation tools to streamline and simplify your workflow, you can reach new audiences without breaking a sweat.
Don’t take our word for it, though. Start your free trial today with Lokalise and see for yourself.
FAQ
Machine translation usually produces acceptable translations for uncomplicated content, but it may fail to translate specialized, technical, or complex material. Human translation is strongest in these cases and can offer higher accuracy, cultural sensitivity, and more natural-sounding results.
t’s generally accepted that neural machine translation is currently the most accurate, flexible, and fluent form of machine translation technology.
Machine translation has been improving since its inception. Regular advancements in AI technology and the use of style guides and glossaries have enabled it to continue improving.
In some applications, yes. Today, highly accurate translations requiring few post-edits are possible, but they’re still imperfect. This is especially true for complex texts that don’t use formulaic language.
How do you eat an elephant? One bite at a time. It’s the same with translation. Human and machine translators work on text in bite-sized chunks of 5-10 words. They’ll constantly refer to dictionaries, glossaries, and other sources to convert the source text into the target language.
The short answer is yes. But it’s not something you should trust to translate important text because the results are not as accurate or precise as a specialized translation tool or a professional translator would be.
