Sci-fi enthusiasts are so familiar with the idea of the universal translator that it is ubiquitous in the world of fiction. And now that it’s becoming a reality, we often think more about whether it works rather than how it works.
As I’m creating a sci-fi world based on our world today, the details are important.
So, let’s dig into the particulars and talk about the processes behind machine translation.
Rather watch a video?
Check out the YouTube version of this blog post.
Both manual and automatic, machine translation models have been made possible thanks to a 9th century Arabic cryptographer called Al-Kindi, who crafted techniques in systemic language translation. But, from there, not much advancement was made. …Not until the 1930s, when a machine translator consisting of a typewriter, a film camera, and a set of language cards was prototyped.
Fast forward another handful of decades to the early 2000s, and you’ll see when machine translation began to form into what we know today.
Three Major Types of Machine Translation Approaches
To be clear, when I’m talking about machine translation, I’m mostly talking about AI translation. But not all machine translation is built on a foundation of artificial intelligence. There are still a few modern machine translation methods that are a bit more manual.
Rule-based machine translation parses the source language. Then, it references dictionaries and style guides to convert the text to be translated into the target language. This type of translation is pretty good when the language gets technical or jargony, but programmers have to manually update the dictionaries regularly to prevent errors from creeping in.
It’s not an AI solution, but still part of the mix of machine translation options.
Statistical machine translation is a model based on math rather than linguistic rules. This model uses algorithms to analyze large amounts of human translations that already exist. It looks for statistical patterns. Then, it predicts a translation based on the statistical likelihood that a specific word or phrase will be with another word or phrase in the target language.
This type of model can be used for features like “auto-suggest” on phones and word processors. And it’s usually pretty accurate as long as it has a huge language repository to analyze.
Neural machine translation uses neural nets, as well as statistical translation methods, to learn and improve language usage. While other machine translation models parse source language passages into sentences and phrases, neural networks consider the whole input.
These networks are made of interconnected nodes that reference giant datasets when processing the source language (or input data). Each node makes a change to transform the source language into the language it’s being translated into. Once all the nodes have had a turn, the output node gives the final result.
Because the neural machine translation approach is holistic, this model is thought to produce better quality translations, and it’s the backbone of AI translation.
How AI Translation Neural Networks Learn New Information
Neural machine translation is considered to be the most accurate translation model because neural networks have been designed to work like human brains.
Essentially, neural networks learn from their mistakes and do a multitude of things more accurately than other machine learning designs.
But just like human brains, neural networks need to understand their world. To facilitate neural network learning, data scientists typically start by feeding labeled datasets to them. These datasets show the machine what to do by providing both an input and a reasonable output. Basically: Question and answer.
The neural network slowly builds knowledge from these datasets. Once trained, the network is fed unfamiliar input and given freedom to process the information based on its training.
And, of course, there comes a day when the neural net has been deemed smart enough to interact with people in the real world.
Though that day is upon us, we’re not yet at the point where universal translators are as common as cell phones. Maybe soon, though.
Until next time, stay curious.
Comments