How (un)safe is machine translation?

Note: This is a revised version of a text previously published at the eMpTy Pages blog under the heading “The Data Security Issues Around Public MT – A Translator Perspective”, with an extensive introduction by blog editor Kirti Vashee and some reader comments. This version is slightly updated.

Some time ago there were a couple of posts on this site discussing data security risks with machine translation (MT), notably by Kirti Vashee and by Christine Bruckner. Since they covered a lot of ground and might have created some confusion as to what security options are offered, I believe it may be useful to take a closer look with a more narrow perspective, mainly from the professional translator’s point of view. And although the starting point is the plugin applications for SDL Trados Studio, I know that most of these plugins are available also for other CAT tools.

About half a year ago, there was an uproar about Statoil’s discovery that some confidential material had become publicly available due to the fact that it had been translated with the help of a site called translate.com (not to be confused with translated.net, the site of the popular MT provider MyMemory). The story was reported in several places; this report gives good coverage.

Does this mean that all, or at least some, machine translation runs the risk of compromising the material being translated? Not necessarily – what happened to Statoil was the result of trying to get something for nothing; i.e. a free translation. The same thing happens when you use the free services of Google Translate and Microsoft’s Bing. Frequently quoted terms of use for those services state, for instance, that “you give Google a worldwide license to use, host, store, reproduce – – – such content”, and (for Bing): “When you share Your Content with other people, you understand that they may be able to, on a worldwide basis, use, save, record, reproduce – – – Your Content without compensating you”. This  should indeed be offputting to professional translators but should not be cited to scare them from using services for which those terms are not applicable.

The principle is this: If you use a free service, you can be almost certain that your text will be used to “improve the translation services provided”; i.e. parts of it may be shown to other users of the same service if they happen to feed the service with similar source segments. However, the terms of use of Google’s and Microsoft’s paid services – Google Cloud Translate API and Microsoft Text Translator API – are totally different from the free services. Not only can you select not to send back your finalized translations (i.e. update the provider’s data with your own translations); it is in fact not possible – at least not if you use Trados Studio – to do so.

Google and Microsoft are the big providers of MT services, but there are a number of others as well (MyMemory, DeepL, Lilt, Kantan, Systran, SDL Language Cloud…). In essence, the same principle applies to most of them. So let us have a closer look at how the paid services differ from the free.

Google’s and Microsoft’s paid services

Google states, as a reply to the question Will Google share the text I translate with others: “We will not make the content of the text that you translate available to the public, or share it with anyone else, except as necessary to provide the Translation API service. For example, sometimes we may need to use a third-party vendor to help us provide some aspect of our services, such as storage or transmission of data. We won’t share the text that you translate with any other parties, or make it public, for any other purpose.”

And here is the reply to the question after that, Will the text I send for translation, the translation itself, or other information about translation requests be stored on Google servers? If so, how long and where is the information kept?: “When you send Google text for translation, we must store that text for a short period of time in order to perform the translation and return the results to you. The stored text is typically deleted in a few hours, although occasionally we will retain it for longer while we perform debugging and other testing. Google also temporarily logs some metadata about translation requests (such as the time the request was received and the size of the request) to improve our service and combat abuse. For security and reliability, we distribute data storage across many machines in different locations.”

For Microsoft Text Translator API the information is more straightforward, on their “API and Hub: Confidentiality” page: “Microsoft does not share the data you submit for translation with anybody.” And on the “No-Trace” page: “Customer data submitted for translation through the Microsoft Translator Text API and the text translation features in Microsoft Office products are not written to persistent storage. There will be no record of the submitted text, or portion thereof, in any Microsoft data center. The text will not be used for training purposes either. – Note: Known previously as the “no trace option”, all traffic using the Microsoft Translator Text API (free or paid tiers) through any Azure subscription is now “no trace” by design. The previous requirement to have a minimum of 250 million characters per month to enable No-Trace is no longer applicable. In addition, the ability for Microsoft technical support to investigate any Translator Text API issues under your subscription is eliminated.

Other major players

As for DeepL, there is the same difference between free and paid services. For the former, it is stated – on their “Privacy Policy DeepL” page, under Texts and translations – DeepL Translator (free) – that “If you use our translation service, you transfer all texts you would like to transfer to our servers. This is required for us to perform the translation and to provide you with our service. We store your texts and the translation for a limited period of time in order to train and improve our translation algorithm. If you make corrections to our suggested translations, these corrections will also be transferred to our server in order to check the correction for accuracy and, if necessary, to update the translated text in accordance with your changes. We also store your corrections for a limited period of time in order to train and improve our translation algorithm.”

To the paid service, the following applies (stated on the same page but under Texts and translations – DeepL Pro): “When using DeepL Pro, the texts you submit and their translations are never stored, and are used only insofar as it is necessary to create the translation. When using DeepL Pro, we don’t use your texts to improve the quality of our services.” And interestingly enough, DeepL seems to consider their services to fulfil the requirements stipulated – currently as well as in the coming legislation – by the EU Commission (see below).

Lilt is a bit different in that it is free of charge, yet applies strict Data Security principles: “Your work is under your control. Translation suggestions are generated by Lilt using a combination of our parallel text and your personal translation resources. When you upload a translation memory or translate a document, those translations are only associated with your account. Translation memories can be shared across your projects, but they are not shared with other users or third parties.”

MyMemory – a very popular service which in fact is also free of charge, even though they use the paid services of Google, Microsoft and DeepL (but you cannot select the order in which those are used, nor can you opt out from using them at all) – uses also its own translation archives as well as offering the use of the translator’s private TMs. Your own TM material cannot be accessed by any other user, and as for MyMemory’s own archive, this is what they say, under Service Terms and Conditions of Use:

“We will not share, sell or transfer ’Personal Data’ to third parties without users’ express consent. We will not use ’Private Contributions’ to provide translation memory matches to other MyMemory’s users and we will not publish these contributions on MyMemory’s public archives. The contributions to the archive, whether they are ’Public Data’ or ’Private Data’, are collected, processed and used by Translated to create statistics, set up new services and improve existing ones.” One question here is of course what is implied by “improve” existing services. But MyMemory tells me that it means training their machine translation models, and that source segments are never used for this.

And this is what the SDL Language Cloud privacy policy says: “SDL will take reasonable efforts to safeguard your information from unauthorized access. – Source material will not be disclosed to third parties. Your term dictionaries are for your personal use only and are not shared with other users using SDL Language Cloud. – SDL may provide access to your information if SDL plc believes in good faith that disclosure is reasonably necessary to (1) comply with any applicable law, regulation or legal process, (2) detect or prevent fraud, and (3) address security or technical issues.”

Is this the whole truth?

Most of these terms of services are unambiguous, even Microsoft’s. But Google’s leaves room for interpretation – sometimes they “may need to use a third-party vendor to help us provide some aspect of [their] services”, and occasionally they “will retain [the text] for longer while [they] perform debugging and other testing”. The statement from MyMemory about improving existing services also raises questions, but I am told that this means training their machine translation models, and that source segments are never used for this. However, since MyMemory also utilizes Google Cloud Translate API (and you don’t know when), you need to take the same care with both MyMemory and Google.

There is also the problem with companies such as Google and Microsoft that you cannot get them to reply to questions if you want clarifications. And it is very difficult to verify the security provided, so that the “trust but verify” principle is all but impossible to implement (and not only with Google and Microsoft).

Note, however, that there are plugins for at least the major CAT tools that offer possibilities to anonymize (mask) data in the source text that you send to the Google and Microsoft paid services, which provides further security. This is also to some extent built into the MyMemory service.

But even if you never send back your translated target segments, what about the source data that you feed into the paid services? Are they deleted, or are they stored so that another user might hit upon them even if they are not connected to translated (target) text?

Yes and no. They are generally stored, but – also generally – in server logs, inaccessible to users and only kept for analysis purposes, mainly statistical. Cf. the statement from MyMemory.

My conclusion, therefore, is that as long as you do not return your own translations to the MT provider, and you use a paid service (or Lilt), and you anonymize any sensitive data, you should be safe. Of course, your client may forbid you to use such services anyway. If so, you can still use MT but offline; see below.

What about the European Union?

Then there is the particular case of translating for the European Union, and furthermore the provisions in the General Data Protection Regulation (GDPR), to enter into force on 25 May 2018. As for EU translations, the European Commission uses the following clause in their Tender specifications:

”Contractors intending to use web-based tools or any other web-based service (e.g. cloud computing) to execute the /framework contract/ must ensure full compliance with the terms of this call for tenders when using such services. In particular, the provisions on confidentiality must be respected throughout any web-based process and the Union’s intellectual and industrial property rights must be safeguarded at all times.” The commission considers the scope of this clause to be very broad, covering also the use of web-based translation tools.

A consequence of this is that translators are instructed not to use “open translation services” (beggars definition, does it not?) because of the risk of losing control over the contents. Instead, the Commission has its own MT-system, e-Translation. On the other hand, it seems possible that the DG Translation is not be quite up-to-date as concerns the current terms of service – quoted above – of Google Cloud Translate API and Microsoft Text Translation API, and if so, there may be a slight possibility that they might change their policy with regard to those services. But for now, the rule is that before a contractor uses web-based tools for a EU translation assignment, an authorisation to do so must be obtained (and so far, no such requests have been made).

As for the GDPR, it concerns mainly the protection of personal data, which may be a lesser problem generally for translators (at least if you don’t handle texts such as medical records, legal cases, etc.). In the words of Kamocki & Stauch on p. 72 of Machine Translation, “The user should generally avoid online MT services where he wishes to have information translated that concerns a third party (or is not sure whether it does or not)”. If you do handle personal data, you should forget about MT since the new regulation requires you to have a contract with the data processor (i.e. the MT service provider), and I doubt that for instance Google or Microsoft will be bothered.

Offline services and beyond

There are a number of MT programs intended for use offline (as plugins in CAT tools), which of course provides the best possible security (apart from the fact that transfer back and forth via email always constitutes a theoretical risk, which some clients try to eliminate by using specialized transfer sites). The drawback – apart from the fact that being limited to your own TMs – is that they tend to be pretty expensive to purchase.

The ones that I have found (based on investigations of plugins for SDL Trados Studio) are, primarily, Slate Desktop translation provider, Transistent API Connector, and Tayou Machine Translation Plugin. I should add that so far in this article I have only looked at MT providers which are based on providers of statistical machine translation or its further development, neural machine translation. But it seems that one offline contender which for some language combinations (involving English) also offers pretty good “services” is the rule-based PROMT Master 18.

However, in conclusion I would say that if we take the privacy statements from the MT providers at face value – and I do believe we can, even when we cannot verify them – then for most purposes the paid translation services mentioned above should be safe to use, particularly if you take care not to pass back your own translations. But still I think both translators and their clients would do well to study the risks described and advice given by Don DePalma in this article. Its topic is free MT, but any translation service provider who wants to be honest in the relationship with the clients, while taking advantage of even paid MT, would do well to study it.

Share
You can leave a response, or trackback from your own site.

2 Responses to “How (un)safe is machine translation?”

  1. Tradutor says:

    Good post Mats, I think there is one big different point, for Human translation you have to pay, Machine translation is free. Maybe neuron nets make free translation more better in the future, we will see.

    • Mats Dannewitz Linder says:

      Not sure what you mean here. Yes, machine translation can be free, but then the confidentiality is compromised (and usually the quality is debatable). If you want confidentiality and better quality, you have to pay. Usually, this solution is applied by professional translators but can of course also be used by the clients.

Leave a comment

Powered by WordPress | Designed by: backlink indexing | Thanks to Mens Wallets, warcraft gold and buy backlinks