Continuous improvement in machine translation (MT) technology means that MT engines are expected to get ever more effective. One of the areas where this is already happening is fuzzy matches for MT.
MT as a technology has reached a certain level of maturity in terms of baseline quality and customization, which implies some serious setbacks for traditional translation memory (TM) technology.
Jourik Ciesielski, MT Specialist, Nimdzi Insights
As mentioned in the latest edition of the Nimdzi Language Technology Atlas, there is ongoing research into combining MT and TM in a single segment. That is to say that part of a given segment may be found in a TM, and therefore presented to the linguist, but another part (missing from the TM) comes from MT.
One of the latest examples of how it can work is the recently announced Phrase NextMT, a new MT engine by Phrase that considers TM fuzzy matches from Phrase TMS. In cases where a partial match is found in a TM, the NextMT engine will translate only the non-matching parts of the segment. NextMT also includes advanced glossary functionality that can handle morphological inflection.
As a result of this development, Phrase now offers both a TMS and an engine that can handle the resultant fuzzy matches all in one application. As noted by Jourik Ciesielski in “The Present and Future of Machine Translation: Your Up-to-the-Minute Guide to This Key Piece of Language Technology” report, the localization industry will benefit from a closer connection between TM and MT technology:
Fuzzy TM matches are a well established concept in the industry, but in the end they are nothing more than translations with a certain error rate. MT is an excellent resource to review and potentially correct those error rates. With the addition of NextMT to Phrase Translate, Phrase has a complete MT feature set that responds to every market requirement, from AI-driven engine selection and instant quality estimation to easy customization and glossaries.
Jourik Ciesielski, MT Specialist, Nimdzi Insights
Another notable example is a new “AI-enhanced TM” feature demoed by XTM and Systran in October 2022. The feature makes “use of mid-level fuzzy matches as an input to an MT engine.” For each segment that contains both TM matches and new words in XTM, Systran MT engine helps translate the latter for linguists: words and phrases that are not present in the TM. The resulting translation has both input coming from XTM’s TM and new translation coming from Systran MT. Moreover, by leveraging 75-84% fuzzy matches from the TM, MT is able to increase the quality of the given segment automatically without additional post-editing needed.
In the future, this functionality might become another global MT innovation: part of a given segment may be found in a TM, but another part comes from MT.
