Neural Machine Translation Makes Multilingual Content Affordable
Does your company have a lot of content and is looking to expand to multiple countries?
You can use neural machine translation to translate content like user guides, brochures, press releases, and reference manuals into multiple languages.
In this article, we’ll explore what neural machine translation is and the advantages of incorporating it into your business.
Watch the Inside the Bot Episode: Secure Neural Machine Translation with Philip Staiger
Translation Accuracy Evolution
Computer-based translation has gone through several iterations over the years. It started with rule-based, moved to statistical, and now AI-based neural networks allow engines to generate accurate and quick translations.
A major advantage of using neural networks is customization. According to Philip Staiger, Technical Account Manager at Systran, “Specializing the engine can make a huge impact in how cost-effective the machine translation can actually be.”
It can run in the cloud or run locally with no internet connection. This ability allows for entirely secure environments where security is of the utmost importance.
Customizing translation with glossaries and dictionaries with a narrow domain focus allows AI engines to use the new terminologies with immediate results. For companies, this saves them thousands in annual translation budgets.
Part of what makes content challenging is that it needs to be handled by a subject matter expert. It’s not only translating it from one language into another but also understanding the content.
You can translate the correct words, but without the proper context, the meaning changes or is lost completely. From this, we can see there are two parts to the translation puzzle: language ability and domain knowledge.
Luckily, specializing the engine trains it to become a subject matter expert in a domain and improves its language accuracy within a domain.
The most significant improvement you can make to the quality of translation comes with using translation memories. Translation memories are full sentences used to train an AI engine that works to refocus or specialize the engine.
Many companies have existing data—years of collected material from previous translations. These “translation memories” are used to teach the system what it should or should not translate, recognize elements that vary, and allow the engine to become domain specific to a particular topic.
Training and customizing the AI engine in a narrow domain reduced the time for post-editing and human verification, resulting in time and savings for a business.
You need a minimum of around two to three thousand sentences to start specializing the engine. This data is specific to your business expertise, not general, like translating common words, but narrow in focus.
Translating a user guide for video editing software is an example. It’s not the material in its entirety that needs to be translated, only the updated sections. There might only be 30% changes to the material between each updated version you release each year, v12, v13, v14, etc.
Since the engine is specialized, it knows how to translate those updated sections from the language standpoint and the subject matter expertise level. It understands the context, which allows it to interpret the new material accurately.
Quality and Neural Fuzzy Adaption
The metric to evaluate the quality of translations is typically based on a validated human reference—thousands of previously translated sentences. Comparing new material to old is based on a percentage match, with 100% being an identical match.
Using neural fuzzy adaptation is another tool that helps engines learn on the fly and reduce the time needed for human verification. “Fuzzy matches” contain similar phrasing and terminology, allowing the engine to infer meaning from close matches.
Sometimes, a business doesn’t need 100% accuracy—95% or 90% is just fine. This is based on the type of content and where it will be published. A user guide, for instance, posted in a knowledge base posted on a forum with 90% accuracy may serve the job its intended purpose.
Machine Translation Opportunities
E-discovery, Biotech, and banking have a lot of opportunities. Some other examples are manufacturing in non-English speaking countries, public announcements at airports, internal communications, and even video games that translate into many different languages.
Subject matter experts will still need to review documents with legal or medical consequences, but machine translation can be used to take over much of the translation effort.
“Every new generation of technology brings changes. It’s fascinating when you can find other ways for humans to contribute and make it even richer,” -Philip Staiger
In a well-focused domain, it may only take a few hours to train and have a high accuracy in the 90 percental. For a business, that can save thousands of dollars in translation budgets.
About Jason Dzamba
Director of Media Relations, Productivity Strategist, and Host of Inside the Bot Podcast, Jason uses a process-driven to help leaders optimize their actions and achieve their most important goals. His creative outlet is painting abstract art and producing music. He lives in Miami, Florida, with his three kids.
Related Blog Posts
Does your firm have a bot workforce? Mike Reynolds, Leader of Service Digitization at KeyBank shares his recommendations on scaling banking processes using digital workers…
Savvy firms are using intelligent automation to handle the rising interest rates, inflation, and uncertainty of world events that contribute to market slowdowns…
Leaders who want to innovate should ask themselves how they can embed innovation in everything they do. Learn Cristina Duta’s perspective on adopting intelligent automation…