Could machines be closer to taking over the world (of patent translations) than we thought they would ever be?

04/08/2022

In view of the UPC system, two Reddie & Grose attorneys whose mother tongue is not English, have attempted to gauge how good machine translations are compared with their own human translations.

Languages at the EPO

Under EPO practice, an application is most often filed in one of its official languages – English, French and German. If filed in another language of an EPC-contracting state, a translation of the specification into one of the three EPO official languages has to be provided within two months of the date of filing. 

The thus selected official language becomes ‘the language of the proceedings’, and is used during prosecution of the application for all communications between the applicant and the EPO. When the EPO informs the applicant that it intends to grant them a patent, the applicant must file translations of the claims in the other two official languages. For post-grant validation at national level, the EPC-contracting states have the right to require for the specification to be translated into one of their national languages.

Translations in the UPC

Ever since the UP system negotiations began, issues relating to languages have been the subject of some intense debate. Some countries worried that extending the tri-language regime in force at the EPO beyond the limits of its jurisdiction could have been prejudicial to individuals whose mother tongue was not one of the EPO’s official languages. For this reason, the EU reverted to the so-called ‘enhanced cooperation’ process, by which Spain and Italy were excluded from the UP system. Italy subsequently reviewed its position and eventually joined the UP system. 

The rules relating to the language requirements for Ups were published in the EU Official Journal as EU Council Regulation No. 1260/2012 on 17^th December 2012 and have not been reviewed or updated since. 

These rules provide for a transitional period, during which a request for unitary effect must be accompanied by a full translation of the specification of the patent into English where the language of the proceedings before the EPO is French or German, or into any official language of the Member States that is an official language of the Union where the language of the proceedings before the EPO is English. 

This provision was introduced at the time as a way to ensure that during the transitional period all European patents with unitary effect would be made available in English, which is the language most commonly used in the field of international technological research and for scientific publications. Additionally, this guaranteed that during the transitional period a fraction of all European patents with unitary effect would effectively be made available in other official languages of the participating Member States. 

Interestingly enough, the rules published in 2012 also indicated that the transitional period should terminate as soon as high quality machine translations into all official languages of the Union become available, subject to a regular and objective evaluation of the quality by an independent expert committee established by the participating Member States. Given the state of technological development at the time, the rules also indicated that the duration of the transitional period was not expected to exceed 12 years after entry into force of the UPC Agreement. 

At the same time, the rules very clearly established that machine translations should serve for information purposes only, and should not have any legal effect, and further advised that, during the transitional period, the translations to be filed with the request for unitary effect should not be carried out by automated means. This was so the high quality of human-made translations would contribute to the training of translation engines at the EPO.  

It is fair to imagine that the legislators who came up with the 12-year transitional period in 2012 assumed that, by now, the UP system would have been up and running. And yet, 10 years later, as we prepare to potentially see the first European patents with unitary effect be granted towards the end of the year or early in 2023, we wonder if it may not be time to reassess those translation requirements, and reconsider the state of technological development as far as machine translations are concerned. 

Machine translations

The research activity in the field of machine translations began at the MIT, the University of Washington, and at the University of California in the early 1950s. Impressed as they were with the potential of machine translation, the US government supported this technology with a view to expanding the capacities for translation of scientific and intelligence data. However, no later than in 1964, a special committee – the Automatic Language Processing Advisory Committee (ALPAC) – formed by the US government to investigate and evaluate the situation with the machine translation researches recommended that further investments in machine translation research be suspended, as there was “no immediate or predictable prospect of useful machine translation”. Machine translation was deemed to be slower, less accurate, and twice as expensive as human translation, and so the ALPAC report recommended that focus be placed instead on automatic dictionaries and researches. 

As is often the case, technology has progressed much faster than the seven members of ALPAC could have imagined, and in the 1980s the research activity took off again with renewed intensity. New methods and strategies of machine translation emerged in the 1990s. Work on statistical methods inspired researchers because of the increasing power of computers, memory capacity, and lower costs. Translation began to be done by estimating the likelihood that a word in a source sentence language corresponds to a word in the target language sentence, in view of aligned phrases and words in the parallel texts. More recently, the emerging of deep learning models has further boosted this area of research. 

But how far have machine translations actually got, as of today? 

Exemplary translations of claim 1 of G1/21

Alex Frank and Andrea Civera, two Reddie & Grose attorneys whose mother tongue is not English, have attempted to gauge that by personally translating claim 1 of EP 1609239 B1 (which was at the centre of the recent G1/21 decision) into their respective native languages (German and Italian) and then comparing their own human translations with machine translations obtained from Patent Translate (powered by EPO and Google) and from Google Translate. The results are shown in the Tables below.   

Concluding remarks

Differences between each machine translation and the corresponding human translation appear in bold in the tables above. It is immediately apparent that both Italian and German machine translations match almost verbatim the human translation, the only minor discrepancies coming down to essentially formal aspects.

Of course we don’t mean to argue that a single example, i.e. claim 1 of G1/21, can be taken as an indication of a general trend, but we do feel that almost ten years after their publication the rules of EU Council Regulation No. 1260/2012 should be reviewed; perhaps the duration of that provisional period or even the absolute veto on machine translations for information purposes could be reconsidered.   

This article is for general information only. Its content is not a statement of the law on any subject and does not constitute advice. Please contact Reddie & Grose LLP for advice before taking any action in reliance on it.