In the news
Phrase, a popular TMS provider that recently merged two offerings (Phrase and Memsource) under one umbrella brand, announced its very own machine translation engine, Phrase NextMT. Phrase claims this is the first machine translation made specifically with the TMS use case in mind.
It can improve fuzzy matches with MT, use terminology including morphology support, and provides advanced tag handling.
MT vendors, who are they?
With this announcement, the number of commercial MT providers on the market reaches 55 (check our recent report for the full list).
We often see MT systems built by cloud computing providers (Alibaba, Amazon, Baidu, Google, IBM, Microsoft, Tencent). Also, a few MT systems are built by so-called “platform LSPs,” which provide full-service localization services, often based on their own TMS (RWS, Lengoo, Lilt, Translated, Transperfect, Unbabel). There are also some MT systems spun off from LSPs or other businesses doing a lot of translations on their own and looking at how to better monetize their data assets.
This is not an often case where a TMS vendor presents their home-brewed machine translation solutions and becomes an MT vendor, having a wide range of other commercial and open source MT options on board.
TMS + AI = ?
We already have seen that a close relationship between TMS and AI yields many customer benefits. Memsource (before it became Phrase) had the first seamless integration between TMS and MTQE. MateCat + ModernMT unlocked dynamic training. Systran worked with XTM to launch their Neuro-Fuzzy Augmentation feature. Intento also works closely with many TMS vendors to ensure our connectors can properly process TMS markup to make the content more digestible by MT.
Phrase announced three key features of their new solution:
- Translation memory adaptation. Per Phrase description, “only non matching parts of fuzzy matches are machine-translated for increased translation quality.” This looks like a different approach to the same problem from what’s implemented by Systran in their Neuro-Fuzzy Adaptation and described in their 2022 paper as “similar sentence translations are provided dynamically to guide translation of a given sentence.”
- Advanced glossary support that can “handle morphological inflection and goes beyond search and replace substitution.” This is a great addition to the list of MT systems providing morphology support for glossaries (so far, we know about DeepL, Microsoft, Systran, and Yandex).
- Automated tag placement. There are two different approaches to handling inline tags for MT: either turn them into a special type of token and translate the tokenized sentence with tags, or remove them completely, translate the plain text, and insert tags after the translation is done via word alignment. Phrase takes the latter approach — it’s the same as what we do at Intento when unhappy with tag handling of a specific MT provider for a specific purpose (often it’s SRT translation).
We are very curious to see how it will go. We often see new MT emerging for the unhandled language combinations or domains, while here it’s for better handling the use-case (TMS integration).
- How will it affect the availability of other MT providers in Phrase (most of them explicitly prohibit customers from building competitive technology)
- Will the data for the model come mostly from Phrase customers, or will it be acquired from the data vendors, such as TAUS?
- How PhraseMT compares to other MT systems in the language pair support, translation quality, and other performance features?
- Will it be available outside of Phrase? The press-release states it will be available to Phrase TMS customers only, but for sure many of those would like to connect it to other systems directly.
- Would end-users consider bundling TMS and MT as a convenient solution or too much of a vendor lock?