The European Association for Machine Translation (EAMT) held its 21st annual conference in Alacant, Spain, on May 28–30, 2018.1 The organizers made a great effort to invite professional translators to the conference with the goal of bringing machine translation (MT) researchers, developers, and vendors closer to the actual individuals using MT systems. I had the honor of being invited as a keynote speaker and embraced the opportunity to present ideas on “human-centered translation technology.”2
Is There Space for the Human in the Modern Translation Pipeline?
I’ve attended many MT conferences over the past decades and witnessed the considerable developments in MT. I’ve also seen the considerable and enduring divide between the MT research and development community and the translator training, research, and professional translation community.
During a webinar presented in April 2018 by the Translation Automation User Society (TAUS),3 the modern translation pipeline was characterized as data-driven, self-learning, invisible, and autonomous. This leads to the question: Is there space for the human in this pipeline? I believe so, and my mission at EAMT’s conference was to convince translation technology developers, and MT researchers and developers in particular, to pay more attention to the humans who use the output from their systems. Which humans did I have in mind? Of course, professional translators are the first cohort to consider, but I also made a plea for thinking about end users.
Let’s start with professional translators. At the TAUS webinar mentioned above, some large global companies confirmed that they now publish “fully automatic useful translation” online (i.e., raw MT output that hasn’t been edited by a translator). However, they also confirmed that they are still benefiting from traditional translation networks and that translation memory (TM) was their first line of support for translation, followed in second place only by MT. With much recent hype about MT, it’s easy to lose sight of the fact that TM technology is still very much relied upon.
Many translators have been using TM tools for many years, and there is no doubt that such tools are highly beneficial in certain contexts. However, in the research domain, several studies of professional translators at work have reported that TM editing environments are still not without their faults. Not all translators will agree with this, but the findings suggest that some translators are irritated by, for example, being forced by TM user interfaces (UI) to focus on “segments” rather than entire texts. Instability of tools, as well as bugs, continue to annoy translators, as does the perceived “complexity” of the interfaces. Although some TM environments have undergone sleek redesigns recently, we still find that the TM UI can be very “busy.”
A workplace study I helped conduct, using eye tracking technology to observe how professional translators worked with their normal tools, found that translators looked at the “target text window” 61% of the time.4 However, this window occupied only 19% of the screen space available in the UI when it was in the default configuration mode of the tool. Of course, the translator could customize the UI to make this window bigger, but the information presented in other parts of the UI are also important for the translator. However, my research colleagues and I believe that interfaces in general could be better designed, possibly even made adaptive, so that the part that requires the most attention at any particular point is given the most dominance.
If research on the translation process and translators has taught us one thing, it’s that we don’t behave like automatons. Though there are agreed-upon general approaches to translation, each translator has his or her own nuances in terms of process, product, and preferences. However, translation technology development hasn’t really embraced this fact to date, especially not MT development where the MT output is often sent to the translator regardless of its suitability or quality. To place the human in the center of the picture, I believe we need to look at the potential of “personalization.” What follows is somewhat overly optimistic thinking, so bear with me.
Personalization
Personalization means tailoring a product or service to better fit the user. In areas such as education, personalization has been shown to have positive effects on learners and learning. Personalization is based on learning about user needs, interests, preferences, expertise, workload, and tasks. Context is highly relevant and user modeling is key.
If we think about translation for a moment, it’s a task that can be modeled. Context is also highly relevant. So, I believe that we could, theoretically, produce translation tools that are personalized not just to individual translators, but to individual translation tasks. For example, if TM (or even MT) data were tagged for register (e.g., formal, informal), and we knew that a specific job required a formal register, we could personalize the translation engine so it prioritized the suggestions tagged as “formal” only. A translation engine could also “learn” about the interests of a translator by logging the types of informational searches the translator carries out. The translation engine could also learn about the online resources a translator uses most frequently by logging the time spent on a resource or whether the translator cuts and pastes information from that resource.
As another example related to the use of MT, we know that translators have varying levels of tolerance for MT. We also know that this is context-dependent. For example, it may depend on the language pairs you work with, on the text type you’re translating, or on the time you’ve been given to produce the translation. Using this kind of information, a personalized MT engine could be established for each translation task. MT could be useful for one context (so switch it on) and totally useless for another (so switch it off). One translator might find MT suggestions useful at a certain quality level while another might find them irritating. So, by using quality estimation scores, the engine might “learn” the tolerance threshold for each individual translator for each specific translation context.
It must be acknowledged here that this kind of “machine learning” is contentious and raises ethical issues, so I’m not suggesting that this be done without knowledge or approval from individual translators. It’s also the case that this kind of learning isn’t easy and would take time before the “engine” was sophisticated enough to be useful. As I said, this is overly optimistic thinking, but the aim is to at least move us away from the scenario where any kind of MT output is produced and sent to translators for all contexts.
What about End Users?
The final part of my talk at EAMT2018 focused attention on a group who are largely forgotten when we discuss translation technology: end users (i.e., the people who are the recipients or readers of the translation produced via tools). This isn’t to say that translators forget about their readers, because they don’t! However, we know very little about the impact on end users of different translation modalities, especially raw (unedited) MT or post-edited MT and how it compares with translations produced by a professional translator without the aid of MT. At Dublin City University, we’ve started researching various types of users to see how comprehension, task completion, and attitude is affected if the user is provided with raw or post-edited MT or with “human” translation. For example, some work has been carried out at Dublin City University to see how well users can follow an instructional text that explains how to carry out tasks in MS Excel. Another researcher is investigating comprehension levels when learners are exposed to MT, post-edited, or “human” translations of subtitles in online courses.
Putting Humans at the Center
Investigating the impact of different translation modalities is really putting the human in the center of translation technology. The final end user group in the study I mentioned earlier included those who might need translation as part of humanitarian response efforts. We’re researching the role and need for translation (not just interpreting) in “crisis” and disaster contexts through a project funded by the European Union called the International Network in Crisis Translation, or INTERACT.5 Humanitarian response is often in need of translation, and it’s a sector that is becoming ever more technologized. MT has already been used to assist with communication on the ground during the Haiti earthquake of 2010, and there is every likelihood that it will be used again in the future. If we get it wrong in this context, we get it really wrong. So, for this kind of end user, we need to make sure we’re not just repackaging MT engines as they are now and handing them over for use without any discussion or consideration of the impact on end users in these kinds of contexts.
My mission at EAMT2018, as mentioned previously, was to try to get translation technology researchers and developers to think more about how the output from their systems affects end users—whether professional translators or consumers. At least we’ve begun to open up the conversation.
If you are interested in finding out more about some of the research mentioned above, please visit: http://bit.ly/OBrien-research.
Remember, if you have any ideas and/or suggestions regarding helpful resources or tools you would like to see featured, please e-mail Jost Zetzsche at jzetzsche@internationalwriters.com.
Notes
- For more information on the European Association for Machine Translation 2018 conference, check out http://bit.ly/EAMT2018.
- This article is based on a chapter by Sharon O’Brien and Owen Conlan, “Moving Towards Personalising Translation Technology,” which will soon to be published in Moving Boundaries in Translation Studies (John Benjamins Publishing Company).
- Translation Automation User Society, www.taus.net.
- Teixeira, Carlos, and Sharon O’Brien. “Investigating the Cognitive Ergonomic Aspects of Translation Tools in a Workplace Setting,” Translation Spaces, Volume 6 (John Benjamins Publishing Company, 2017), 79–103, http://bit.ly/Translation-Spaces.
- For more information on the International Network in Crisis Translation, see http://bit.ly/INTERACT-project and @CrisisTrans on Twitter. This project receives funding from the European Union’s Horizon 2020 Research and Innovation Program under the Marie Skłodowska-Curie Grant Agreement No. 734211.
Sharon O’Brien is an associate professor in the School of Applied Language and Intercultural Studies at Dublin City University. She teaches translation, translation technology, localization, and research methods. Her research focuses on human factors in translation technology, especially machine translation and, more recently, the role of translation in crises and disasters. Contact: sharon.obrien@dcu.ie.