Machine Translation vs Human Translation
Where is the translation industry going? Will translation professionals still have a job 30 years from now or are they a dying breed destined to be replaced by Machine Translation (MT)? Can we measure translation quality, processes, and resources? Is data the key to success?
These are just a few of the far-reaching questions touched on during the TAUS (Translation Automation User Society) Roundtable in Barcelona that I attended on 12th May together with Andrea Tabacchi our Technology Director.
The underlying question that ran through all the separate presentations was the following: does the future of translation need human contribution and to what extent? The roundtable focused on how our profession will be affected by future technology developments and on how we must respond to change.
Starting with an analysis by Celia Rico, Universidad Europea, of the use of Machine Translation among LSPs, we have seen how the focus today is not on obtaining ‘perfect’ MT output quality, but rather ‘useful’ MT output and the role of the post-editor has become crucial. However post-editors will not have to do all the work alone, as they will be supported by technology, for example integrating MT systems in the translation environment, managing terminology more effectively to improve MT output quality, controlling source text formats and authoring and dealing with a huge variety of domains.
MTradumàtica, created by the Tradumàtica Research Group, Dept. of Translation of the Universitat Autònoma de Barcelona, is proof of this new approach. MTradumàtica is a web-based Stastical Machine Translation (SMT) engine that responds to some of the main concerns regarding Machine Translation: privacy, adaptability, customization. All these issues are addressed by empowering the translator through technology: the final result is high quality translation, which can be obtained by integrating the human contribution with the latest LT developments.
How to measure Machine Translation quality?
But how can we establish machine translation quality and, more in general, how do we know if we are applying the correct processes? ‘Measuring translation’ is the key, as Luigi Muzii, sQuid, declared, where “measuring translation” in our profession is intended as measuring every piece of translation data that can influence our decision-making processes, from translation quality levels to productivity per vendor, from most profitable domains to high-volume customers. All this information allows us to make correct decisions based on previous experience, in the form of data.
Data-driven decision-making is the future. Technology has already made some important steps in this area and we saw some interesting examples at the roundtable, ranging from data-sharing platforms to measuring tools for different purposes.
Apertium is definitely a valid tool. Developed within the project “Open-Source Machine Translation for the Languages of Spain” funded by the central Spanish Government, Apertium is a free/open-source Machine Translation platform, where language resources can be shared mainly for MT purposes. What makes Apertium very interesting is that, being open-source, it can be integrated into any multilingual content management system, and because it is open it is constantly growing thanks to shared contributions from the community.
The TAUS DQF (Dynamic Quality Framework), presented by Paola Valli, TAUS Product Manager for the Quality Dashboard, is another example. Integrated in most CAT tools through an API this tool allows users to measure key factors in addition to Machine Translation quality and translation quality, such as the post-editor performance, productivity and can also provide a categorization of error types. All these measuring tools facilitate the lives of both project managers and post-editors as they can be used to evaluate each step of the translation workflow and also compare the performance of projects and translators over time. In addition, translators and post-editors have full access to the metrics connected to the projects, so this is again another way of involving them directly in the measuring process.
As with the DQF, APIs can be created for any software tool and integrated into customer systems. Diego Bartolomé, tauyou, made this clear: customers do not want to adapt to unknown environments, they want us to adapt to their environments, this is why we must be ready to “API-fy” our industry in order to be able to respond to customer process automation requirements as well as to broaden our customer base. In this API-economy the human contribution to the process seems to be reduced, because processes will become more and more streamlined and managed by machines.
Automation can also be applied to a different step of the translation process: vendor selection. For example within Arancho Doc, our Dynamic Vendor Rating helps us assign translation projects through an automated process based on automatic resource ranking. The most suitable resource is selected for a given customer by this tool based on a number of data collected for that specific resource. Data is collected for each crucial factor: delivery rating, quality score, availability and so on. The result is a data-centric approach which proves to be successful.
However quality is not purely data/technology-driven but also powered by crowd motivation. Xavier Maza, IDISC, and Olga Blasco, Rosetta Foundation, explained why. From both commercial and pro bono perspectives it appears that motivation can have a postivie effect on quality levels. Words like trust, empowerment, and fair pay are still factors that can determine the quality of the final product. This is even more true if these values are combined with data-driven technology, and particularly in the case of pro bono approaches, such as Rosetta Foundation’s Trommons system, where the word ‘fair pay’ can be replaced by ‘motivation’.
For more information, please visit: http://www.aranchodoc.com/machine-translation-science-fiction-or-tomorrows-reality-taus-roundtable-2016-barcelona/