Neural translation is entering product territory
What changed in machine translation in 2014 and why it matters for companies dealing with large volumes of text.
Machine translation has existed for a long time, and most companies have dealt with it - usually with disappointing results. Statistical systems produced acceptable output for simple texts and fell apart on anything more complex than an appliance manual. That experience formed a settled opinion: machine translation is a supporting tool, not something you show to customers.
That opinion is now worth revisiting. Not because the technology suddenly became perfect, but because its capabilities have shifted meaningfully over the last year or two. The same wave that is reshaping translation quality is visible in how language models began handling meaning differently in search.
What changed technically
Machine translation systems based on neural networks produce qualitatively different results for certain language pairs and text types. The difference is most noticeable where coherence matters: a neural model "sees" context over a wider window and loses the thread between sentences less often.
This does not mean neural translation is better everywhere. On technical texts with specialised terminology, older approaches still compete. But on coherent narrative text - marketing copy, support content, general documentation - the gap in favour of neural systems has become perceptible.
Where this is practically applicable
Companies that need to work with text in multiple languages fall into a few categories.
The first is those who translate for understanding, not for publication. Internal documents, incoming correspondence from overseas partners, monitoring of foreign news. The quality bar here is lower, and machine translation already does this well enough.
The second is those who want to accelerate professional translators. Machine translation as a draft that a human edits is a different economic model from translating from a blank page. When the initial quality is high enough, both speed and cost change materially.
The third is those thinking about translation for publication without post-editing. This is the most demanding case, and quality for a specific language pair and text type needs to be verified separately.
What not to expect
Machine translation does not replace a human translator where brand voice matters, where legal precision of phrasing or cultural adaptation is needed. The technology has improved, but "sounds right in another language" is still human work.
It is also important to understand: quality varies considerably by language pair. Translating from English to German and from English to Russian are different tasks with different levels of tool maturity.
Questions for evaluating fit
Before investing in machine translation as a tool, a few questions are worth answering:
- What volume of text do we translate each month, and what types are they?
- Which language pairs do we need?
- What is the current process - professional translators, in-house staff, or no translation at all?
- What is the quality measure for our case - internal comprehension or publication?
- Do we have a glossary of key terms that must be translated consistently?
The technology moves fast. An opinion formed three years ago is probably already out of date.