I recently moderated a session on incorporating the European Commission’s eTranslation MT solution into Trados workflows at ETUG 2022. ETUG 2022 is annual meeting of the European Trados Users Group. It brings together translators, language technologists from corporate language services and representatives from RWS to talk about all aspects Trados. RWS unveil roadmaps for their products and there are use case presentations, like in my session.
Catherine Lane and Daniel García-Magariños from the Language Technology and Innovation Unit within DG Translation at the European Central Bank demonstrated how they have approached incorporating eTranslation into Trados workflows.
For me, as a banking supervision translator, it made moderating the session simpler, but interventions from the floor from the automotive industry also provided valuable inputs on some of the considerations for use of MT in workflows.
The ECB Experience
The presentation was split into two parts. Catherine addressed setting up the Finance engine, including the QA of imported data, and language combinations and available engines. Daniel demonstrated the tool developed by the ECB for importing machine translations TMX files.
Catherine dealt with how the ECB has applies rules about the level of confidentiality of documents that can be sent to eTranslation. Mitigations are in place (e.g. files downloaded from eTranslation, not by e-mail, and deleted immediately from the system after download). These measures are necessary for ensuring that the files do not remain in the system for any longer than is necessary.
Catherine also addressed issues about onboarding of translators – they had used an eLearning module to handle some of the training. Currently it is still an additional aid to complement existing server-based human translation TMs, and not a direct replacement, and serving more as a starting point where existing TMs did not include good fuzzy matches for sentences.
Currently translations are only delivered for one engine at a time. However, it is possible to have translations into multiple languages. I meant to ask about pivot languages for exotic combinations – e.g. for Finnish-Maltese does MT output involve an intermediate step through English?
MT’s Top Model
Another consideration is about which engines to use. For my area of work, I would probably use 2-3 engines (e.g. Bundesbank Neural, Finance, Formal). This would require running the process three times at the moment. Depending on the text type, however, the Formal engine (e.g. for legal texts) might prove the most useful. The Finance engine would prove more useful for financial texts.
As Daniel explained, processing power also means that there is currently not direct way to access eTranslation from inside Trados. Instead, eTranslation translates the document and the output made available to download. Downloaded files are imported into a separate Translation Memory for MT results. In the translation project a 20% penalty applies. The TM settings are “lookup” and “concordance” enabled, but “update” disabled. This essentially means it is a read-only translation memory.
The ECB’s “eTranslator importer” helps ensure that the files land in the right place and domain-specific fields appended to each TU. This includes extra field content about the engine used. The Translation Memory is cleared regularly.
Averse – Ambivalent – Evangelist
Three attitudes towards MT emerged in the discussion about the uptake among translators. I called them “averse”, i.e. those who opposed the use of MT, “ambivalent” i.e. nice to have but not a deal-breaker, and “evangelist”. There has been some move away from “averse” towards “ambivalent”. Possibly this is due to the emergence of NMT, thereby overcoming the aversity displayed towards statistical Machine Translation.
A similar project from the automotive industry mentioned that their own project had only given access to more experienced translators. Less-experienced translators might lack the depth of knowledge to identify that a fluent sounding TU was in fact incorrect.
I am in the “ambivalent” camp. The potential uses for eTranslation in my setting in banking supervision are evident. I am aware of the fact that there is still a considerable post-editing of the MT required. My direct concern is needing to pseudonymise all mentions of the entities in question. Similarly for any placeable values (e.g. about total assets etc.), but doing so negates the productivity gain.
I find MT output is very rigid in its word order, whereas I like to invert sentences to in turn negate the use for a passive in English
However, I can understand and appreciate that texts carefully prepared for translation (check out search results for “writing for translation” to get an idea), mean a greater productivity gain. This might in turn improve unnecessary verbosity and lead to clearer writing.
Takeaways from the session
A few take-aways from the break-out session on integrating eTranslation into the translation workflow of the European Central Bank
- Any institution, agency and authority with access to eTranslation can use this approach.
- There are a number of domain-specific engines. Currently eTranslation only uses a single engine per request. Different engines seem better suited to different text types.
- eTranslation works for all EU official languages, and there are also some other non-EU languages (e.g. Ukrainian, Chinese, Arabic)
- ECB used language data from central banks and supervisory authorities to build the Finance engine.
- A 20% penalty to MT output means that eTranslation output only comes into play where there are no human-translated and verified TUs.
- From an assessment of translation quality for pure MT out, language combinations with the largest number of TUs achieve the best results.
- Translators fall into three camps “averse”, “ambivalent”, “evangelist”. Some sceptics (averse) are becoming more enthusiastic, partially due to the advent of NMT.
- Future developments include tools for anonymization or pseudonymisation – essential when using names of entities etc.
- Translator experience level may contribute to gains from these workflows.