Machine Translation: Where It Works, Where It Doesn’t

It is the opinion of many translation industry experts that machine translation (MT) will never be able to fully replace the skills of a qualified translator. Why is that? Well, it comes down to the fact that MT software has yet to advance to the point (and may never advance to the point), where it is able to translate with certain cultural implications, context, and market perception in “mind.”


Culture is a living, breathing, and ever-evolving factor that translation technology will always have a very difficult time keeping up with—and when it comes to globalizing a company, being mindful of culture is not only a pivotal advantage, but also a necessary asset. That being said, there do exist certain content types that—whether it be due to orientation, reader expectation, or quantity—make sense to translate using MT. Lionbridge has developed three levels of MT services to accommodate these circumstances.

Raw MT

Lionbridge’s raw MT is 100% computer generated translation—without the involvement of any human edition. This means that thousands of words are able to be translated in minutes. However, the implementation of raw MT translation allows for the existence of mistakes, such as spelling errors, mistranslations, inconsistencies, and poor overall style.

Light Post-Editing (LPE) MT

Lionbridge’s LPE MT involves text being firstly machine translated, and secondly read and edited by a linguist. The quality level offered in LPE is “good enough/fit for purpose”, and the average productivity is 4,000 words per day.

While the addition of a human editor ensures that the text will at least be accurate and comprehensible, it will not be stylistically compelling, it may sound like it was generated by a computer, and it still may contain errors in punctuation, grammar, syntax, and among other possible inconsistencies.

Heavy Post-Editing (HPE) MT

Like the LPE MT process, Lionbridge’s HPE MT service means that text is machine translated and then read and edited by a human. The implementation of HPE offers quality that is “close to that of human translation/publishable quality”, and the average productivity is 3,000 words per day. In contrast to the LPE, HPE’s final products are stylistically fine, though the style may not be as good as that achieved by a native-speaking human translator. Also, the syntax will be normal, and both grammar and punctuation are corrected during post-editing.

So, what texts are suitable for MT?

To start, HPE MT is the only MT service which could be used externally, in client communication, etc. Other than that, MT services are mainly used for repetitive, technical texts, where style is unimportant—or texts with little or no limited number of ambiguous words wherein a word can have more than one meaning. Until machines can independently think and learn, all other content-types should be left in the hands of good old-fashioned, expert human translators.