Neural Machine Translation (NMT) systems are changing the world right now! Unlike previous MT technologies, NMT provides higher quality translation and is improving fast. Very fast. At its current rate of development, NMT is set to drastically change the traditional, human-based translation industry in as little as one to three years, impacting over 600,000 linguists and 21,000 language service providers (LSPs).
For customers, the age of NMT will lead to massive reductions in pricing and improvements in speed (even where humans are required).
To take full advantage of the NMT wave, however, LSPs and businesses need the capability to automatically handle a complex process involving simultaneous projects, high workloads, quality control, etc.
One Hour Translation (OHT) and companies like Booking.com are already doing this using a hybrid (NMT + human) translation approach.
Why Neural Networks (NN) matter
Deep learning and artificial intelligence are changing the world right now. The applications of these technologies encompass several industries. Neural Networks (NNs), the foundation underlying these technologies, is used for smarter image recognition, autonomous driving, more capable digital personal assistants, x-ray analysis, improved voice recognition, mastery over games like Go and Chess, and much more.
A lot has been written about the field, but in brief layman’s terms, NNs imitate the way the human brain is built and how it learns new skills / information. From a technical perspective, an AI system can be viewed as several matrices, layered one on top of the other. The cells in each matrix represent neurons, which are connected to other cells in the next matrix, with varying degrees of “strength.”
The top matrix is fed some input, e.g. pixels of an image. The input “trickles” down to the bottom matrix, layer by layer, via the connections between the cells, and the output from the bottom matrix can be whatever the system is trained to produce, e.g. it can ascertain whether the image contains a car or not.
Conventional information systems from SAP, Salesforce, Facebook, and even the OHT platform have millions of lines of code each! They take years to develop and are subject to endless debugging and improvements so they can continue to function properly.
One of the amazing things about NNs is that the software required to run them is relatively simple, just a few thousands of lines of code, compared to the very long and complex code behind conventional IT systems. This is a very important feature of this technology, and means than making a NN-based system better or “smarter” is much easier than improving traditional software.
What makes a NN “smart” and useful is mainly the data it is “trained” with and not the complexity of the software itself.
So unlike conventional software where the smarts is mostly in the software, with NNs the value and the desired function of the system are mainly a product of the nature, quality, and quantity of the training data.
NN training is a relatively simple process whereby input data is fed to the system and the result is examined versus the desired result. A simple process adjusts the connections between the cells to get a result closer to the desired outcome. Doing this hundreds of thousands or millions of times produces a NN that can handle its designated task well. Clearly, the accuracy of the feedback provided to the NN is crucial for its training.
The exact same NN can be useless before it is trained and extremely valuable afterwards.
Why is this technical mumbo-jumbo important? Because it has one simple meaning. Making NN systems better is easy. All that is needed is more compute power and more quality input. In this manner, unlike conventional software, NNs do not require architects and product designers to think of new ways to improve the system. Nor do they require brilliant engineers to work for years to make the software increasingly better.
Making NN systems better is easy. All that is needed is more compute power and more quality input.
Technology and Translation
In the past 50 years or so, compute power has been growing exponentially. Gordon Moore (co-founder of Intel) predicted as far back as 1965 that the number of transistors on a given piece of silicon would double every two years (Moore’s law).
To better understand the potential of this sort of exponential growth, consider the following: doubling the length of a 1 meter stick 25 times makes it +33,000Km long, almost 3 times the diameter of the earth!
Continuing to double it 25 times more (i.e. 50 times in total), makes it over 1 trillion and 125 billion kilometers long! For comparison, the radius of the solar system (average distance between the Sun and Pluto) is “just” 5.9 billion km.
Applying this to calculations, a computer that could run just 100 calculations per second 50 years ago is now able to process over 3.3 billion calculations per second today, and will do over 112,500 trillion calculations a second in 50 years (assuming the current rate of improvements continue), i.e. 112,500,000,000,000,000 calculations per second.
This improvement in compute power is happening very fast and, with time, the performance increase is more and more substantial due to its exponential nature. These improvements are happening regardless of what the compute power is used for.
The other important factor contributing to making NNs better is the quantity and the quality of data. Consider the fact that 90% of online material today (images, text in various languages, etc.) is less than two years old. And more and more content is being generated all the time.
More and more quality data is available online for NMT training.
How does this apply to translation? In the past two to three years, several companies started using NNs for translation. The results of these efforts are staggering. Over this short period, the quality of NMT in some areas became human like, swiftly surpassing previous translation technologies.
Moreover, NMT systems continue to improve fast, thanks to:
- Increase in compute power because of: Exponential compute power improvements as described above
- More compute power allocated to these systems as they demonstrate great results
- More training material coming fromCrawling the web – source and translations available online
- Translation memories created by business customers
- Proactive translations done for training purposes
- The ability to provide human feedback on a massive scale, e.g.OHT is already running NMT rating and feedback projects (over a million projects to date) for NMT vendors / users.
- Facebook and Google encourage users to provide translation feedback, etc.
The important thing to understand is the rate of change. Previous technologies improved slowly because conventional software improvements depended on human developers. With NMT, the core technology can improve its processes quickly through additional compute power and input data that are easy to add.
The Revolution is here!
In simple terms, NMT is a tsunami approaching quickly; it is not “just another” wave of technology advancement. I predict that over 30% to 50% of all translations in the world will be done using NMT within one to three years (with potentially some level of human intervention).
NMT is a tsunami approaching quickly; it is not “just another” wave of technology advancement
Now consider this: the world’s biggest LSP, TransPerfect, controls less than 2% of the USD 40+ billion global translation market. This means that NMT (with post-editing, quality control, etc) has the potential to totally disrupt the market.
NMT has a direct impact on over 600,000 linguists and over 21,000 translation agencies. Those who manage to leverage the technology will survive; the rest will have difficult times. A similar pattern is developing in the automotive industry, where autonomous cars / trucks (Tesla, etc) and NN-based automation (UiPath) are replacing drivers and office workers.
To ride the NMT wave (and avoid drowning), LSPs and major businesses that use NMT systems, should be able to handle complex translation process. The NMT engine alone is not sufficient for real business use. Using an automotive analogy, NMT is like the car’s engine, while a business solution, like OHT’s Hybrid translation service, is the entire vehicle.
To work properly, a hybrid (NMT+human) translation service should handle hundreds of thousands of projects simultaneously, dynamically select the right NMT, decide what human intervention is needed and where, and make sure quality control is done properly and smoothly. This is a complex, multi-step process.
NMT engines are just like real car engines, heavy blocks of metal that are not very useful on their own. Business customers need the entire vehicle, i.e. NMT engine + procedures around it, to benefit from NMT.
OHT made a substantial shift of focus to become the first hybrid translation agency. Using our Hybrid translation system (NMT + humans), we are already providing high quality, low cost translations to business customers.
We recently released ONES – the first independent, human based NMT evaluation score. Using ONES, we are able to select the perfect NMT engines for specific projects “on the fly.” We are also working with a few of the largest NMT vendors to train their systems by providing high volumes of human translation, rating and evaluation of NMT results, and human feedback and corrections, among others. More importantly, we do so in 100 languages, since one of the main issues with NMT is having enough material for training in languages other than English.
OHT’s estimate is that by the end of this year, 80% of our general translations will be done using NMT, providing human quality at a much lower price and much higher speed than traditional translation.
Bottom-line: NMT is already here. It is taking over fast—very fast—and more and more business customers are beginning to enjoy its benefits.
Interestingly, Blockchain, another fast-growing technology, can assist NMT in reaching market domination.
Where does Blockchain come in?
There is a huge pool of existing business translations that are stored in Translation Memories (TM). These TMs were built over many years and contain tons of data. The data is saved in a way that is perfect for NMT training. Using these TMs, NMT training can leap forward, producing engines perfect for business customers. Furthermore, these TMs are prior investments, so any future income they generate for their owners is pure profit.
So what is missing? To make these TMs readily available, there is a need for an easily accessible repository or some sort of a marketplace. The minimal requirements are to make it easy to search the source content while ensuring that the translation is delivered only upon payment, and to eliminate the need to trust a central vendor that will manage it all. Most users (businesses or individuals) will not upload their TMs to some central system due to privacy issues.
A blockchain-based system could be the ideal solution! Using blockchain-based architecture, it is possible for TM owners to earn income without risking their data privacy. TM owners like companies, translators, etc., will be able to share their data in a way that will only expose it once payment is made. Customers will be able to select the translation data they want based on metadata (like rating), and translators will be able to upload their translations for common phrases and get paid every time someone uses them (pay-per-use).
The financial incentive for such a system is clear, powerful, and already exists. Many of our customers will gladly recoup some of their past translation cost by selling their TMs.
Once such a blockchain-based system is available, NMT builders will be able to purchase the training data they need, in many languages, and make their NMTs better at performing business translations, instantly!
OHT is in the process of building such a marketplace, and will elaborate more about that in a separate post.