Neural Machine Translation
Neural machine translation (NMT) is a still-novel approach to automatic translation based on artificial neural networks, vast “webs” of elements interconnected not unlike neurons in animal brains.
As of 2020, neural machine translation is almost universally accepted as the most accurate, fluent, and versatile approach to automatic translation.
AISA’s Neural Machine Translation
AISA’s NMT is trained in Southeast Asian languages and contexts, with the aim to provide our clients with the most accurate machine translations in the region.
If you wish to check out the quality of our in-house NMT, simply sign up an account at our Translation Portal. It’s FREE!
Main Benefits of Adopting Neural Machine Translation
With machine translation, you don’t have to worry about your content getting stale while it’s being translated. This is especially crucial for fast-paced sectors such as mobile app development, where updates can happen several times a week.
NMT finally clears machine translation of its notorious clumsiness and non-natural structures. Especially with low- to mid-complexity texts such as online reviews, the translation quality is almost indistinguishable from a human!
Through just a bit of engineering, you can make your app, website, or video game automatically “feed” its texts to an NMT engine hosted in the cloud — and have the translations pushed back right to your repo. That’s a huge work-saver！
Businesses That Benefit Most From Neural Machine Translation
- App & video game developers
- Website & blog owners
- E-commerce providers & platforms
- E-learning providers & platforms
- Media companies
Choose The Right Neural Machine Translation Engine
Here are some guidelines to make sure you make the most from your NMT experience:
- Specialised over generic. If an engine translates “all to all” languages, there’s a high chance that the quality will go down. NMT engines specialising in a few languages, such as AISA MT, will catch the nuances of those languages much better.
- Proprietary over public.Public MT engines such as ones from Google or Microsoft are trained on the very texts you feed them to translate. If these texts are sensitive, such as in a contract, they may get accidentally outed to other users of the same engine.
- Consider post-editing. If your text is critical, it might make sense to order a post-editing service to rule out any mistakes. It may cost you more than just an MT package, but for some types of content the cost of a mistake can be much higher.
A Bit of Tech Talk: Neural vs Statistical Machine Translation
Since its invention in the mid-2010s, NMT has become the most advanced machine translation technology. It surpasses the runner-up, statistical machine translation, in almost all respects, from fluency to generalization.
Both MT technologies share the same basic approach: Engines are created based on large bilingual “corpora” — sets of texts that are already available in both languages. The difference is as follows:
- Statistical MT engines do a “brute-force” matching of words and phrases that it finds in both languages. This is effective for “simple” matchings such as “I love you” versus “我爱你”, but gets much worse in less obvious examples.
- Neural MT engines, on the other hand, use state-of-the-art deep learning algorithms. This allows them to encode and decode meanings that go far beyond superficial word structures and be truly context-dependent.
Fluent and Natural
The output of statistical MT often looks as a “Frankenstein’s monster” made up of disparate patches. Neural machine translation, on the other hand, reads natural and is often indistinguishable from natural language.
The output quality of statistical MT decreases exponentially for niche subject matters. Neural engines, in contrast, generalise very well and are able to provide correct output even for unusual input text structures.
Works for all Languages
Statistical engines work poorly for languages who don’t share the same sentence structure. Neural machine translation, though, doesn’t have such an issue and can handle even dissimilar languages such as English vs Chinese.