Cloud Hosting

Artificial Intelligence

Artificial intelligence (AI) for translation is something Google and other companies have provided for individuals. It can be accessed on your phone. However, translation is still a much larger and complex issue than many people realize. The business community has many complex and unique needs that add to the challenge of accurate and reliable translation, and AI is showing increasing capability.

One of the keys to business translation is the simple reality that each business sector has its own terms, phrases, and even idioms. A generic translation system in the cloud, trained widely by crowd sourcing or other public methods, wont have the accuracy required for business translation. In addition, the cloud itself is still a problem. Much of a businesses goals involve protecting intellectual property (IP). To do that, they want their information to stay on-premises, behind their firewalls.

Now throw in the complexity of privacy requirements such as the European Unions GDPR and Californias CCPA. Increasingly, governments are setting rules for where citizens data must be kept and what may be shared. The location and anonymization of information also adds to the challenge of a company understanding multilingual business.

Then theres collaboration. Just about everyone in business uses some electronic communications, whether it be emails and text messaging or more formal chat systems. Enhancing those applications with accurately, instant, translation can improve a global companys internal communications and drive success.

While ediscovery was an obvious entry point into business translation with AI, began JP Barazza, CIO, SYSTRAN. Imagine global development groups in such industries as high tech and biotech. They can become more efficient with the assistance of strong translation. Customer support is another horizontal where translation can be of great use.

As with cloud based models, the SYSTRAN system is using unsupervised learning. However, it uses a far more curated data set in order to train systems for each industry. The neural network is only a component of the logic of the system. Because of the specific terminology in many languages, procedural logic is used in pre- and post-processing around the network to help with the clear rules and terminology of business sectors. After all, theyre easier to manage for clearly defined linguistic convention while the neural network can handle the fluidity of the overall language.

One example of the need for rules is how names are used in American English and French French (yes, I had to do that). In the US, we regularly use a leaders name, such as President Biden. In France, news reports usually dont use names, but refer to titles, such as The President of the United States. Think about two-way translation. While it is simple to translate from English by dropping the name and expanding the title, said Jean Senellart, CEO. If we add a name when going from French into English, what happens when the president changes? The system would continue to add the previous presidents name until there was enough data to retrain the system. We made the decision to keep the French reference style when translating to English in order to remain accurate. The use of explicit rules is a clean way of addressing that issue.

That combination of neural network and procedural rules also provides flexibility to the company. A core system can be trained, with different plug-ins around it for different companies. That allows both a simpler development cycle and a cleaner way to provide updates. Specific corporate and industry rules can be added without having to retrain the deep learning system.

Increased accuracy is necessary for business. Consumers are willing to accept errors, as long as the general meaning is conveyed through translation, said Mr. Barazza. Business needs accuracy. Its not just for regulatory and contract compliance, a lack of accuracy can slow product development, lower safety, and create dissatisfied customers.

Because of that need for accuracy, and due to the state of the industry, theres another component to the solution. We are not yet at a point where the automated systems can be completely trusted. Humans must review the translations.

Within the system, at this point translation is complex and is focused on a small enough group of languages, so they are using pairwise engines. For example, one engine translates from English to French and the other translates from French to English. Training the systems uses a weird form of back propagation. In a single engine, back propagation means correcting results and feeding them back in as input. In translation, that means translating results back through the second engine, then correcting. Its more complex that that (at least for me), but I understand the basics of a very interesting loop where both engines help train each other.

That is the way translations are now done, but there is a change that will happen. That style means a lot of individual engines and the larger the number of languages, the increased permutations mean vast increase in the number of engines. One solution have been to use English as an intermediate language, translating everything through it to limit the different engines. That adds inefficiencies and inaccuracies. Facebook has recently announced a single model that can translate in all direction for multiple languages. While individuals are more comfortable with errors so thats a great place to test out such a model, eventually the technology will strengthen and corporate translation will benefit.

Business also drives non-AI design issue. SYSTRAN is not a pure cloud play. They must by hybrid, as on-premises computing is often required to meet privacy and other regulations.

Because of the state of the evolution of systems, including a lack of transparency in deep learning, no company is going to exclusively use AI driven translation for business and government. It will be used in respect to the 80/20 rule, where basic translation will save significant time and effort while humans will still be required to review and edit final versions of business and governmental translations.

Translation tools have made great advances in the last decade. Given the less rigid requirements in translation between individuals, it is no surprise that the initial focus has been on personal use. Technology has now advanced so that addressing the more formal requirements of business and governmental translations is now being addressed. Its early, but its looking good.

Read more here:
Translation Is Trickier For Business, And Artificial Intelligence Can Help - Forbes