Why Eye-Tracking Matters in Machine Translation Post-Editing

Eye Tracking in Machine Translation

Eye-tracking is a type of sensor technology that can detect our presence, follow in real-time where we are looking, what we are looking at, and for how long. The technology has already advanced to a point where eye trackers in screens, webcams, and eyeglasses have been deployed across a range of use cases to capture and measure eye positions and movements.

As a research tool, eye-tracking — either alone or with other biometric sensors — is also being used in market research, website testing, gaming and UX, learning and education, neuroscience and psychology, medical research, and elsewhere.

The foundation of eye-tracking technology is the eye-mind hypothesis: the close relationship between a person’s gaze and what that person is thinking about. Thus, eye-tracking can give researchers valuable information about the cognitive processes involved in a specific task.

An increase in the number of fixations (i.e., the periods of time where the eyes are relatively still), the higher durations of such fixations, increased pupil dilation, and so on, can all be used as indicators of increased cognitive effort.

Applications in Translation Studies

In recent years, eye-tracking has attracted considerable interest in Linguistic Studies mainly because researchers have been able to focus on the entire process rather than on the final product thanks to eye-tracking technology.

Thus, researchers have used the technology to answer a number of questions related to translation, interpreting, audiovisual translation, ergonomics, machine translation (MT), and post-editing.

SlatorCon Remote December 2022 | Super Early Bird Now $98

SlatorCon Remote December 2022 | Super Early Bird Now $98

A rich online conference which brings together our research and network of industry leaders.

Register Now

For instance, it has been used to study cognitive effort when interacting with MT output, monolingual post-editing, as well as comparing processes with each other: post-editing with translation-from-scratch, standard post-editing with interactive post-editing, and post-editing to a native language (L1) with post-editing to a non-native language (L2).

Marking Errors and More

Eye-tracking has also been found to be a promising automatic or semi-automatic MT evaluation technique, mainly because it correlates very well with human evaluation. Furthermore, unlike other techniques, it measures processes that are largely unconscious and can thus be fairly objective.

It can also be used to mark errors in MT output — as MT errors trigger more fixations and longer gaze-times than correct passages — and find which error types are more difficult to identify and require more cognitive effort to be corrected.

According to a study by Stephen Doherty, Sharon O’Brien, and Michael Carl, eye-tracking can be used to extend the basis of MT evaluation by involving end-users; therefore expanding MT evaluation activity into the fields of user reception of MT output.

This means that MT developers will be able to automatically collect data on what the actual end-user has difficulty with by monitoring their reading behavior and how they interact with MT output.

Additionally, eye-tracking is a potential research tool for developing fair post-editing pricing models. Given that there is no industry standard, many different ways are being used to calculate prices in post-editing, including per source or target text character / word / line, per hour, according to edit distance (HTER), calculating MT segments to be equivalent to fuzzy match segments, and so on. But none of these take into account the actual effort involved in a post-editing task.

“Post-editing effort can be used to evaluate the quality of machine translation (e.g., Aziz et al. 2013) and develop a suitable pricing model for post-editing” — Sanjun Sun, Associate Professor, Beijing Foreign Studies University

According to a study by Hans Krings, Professor of Applied Linguistics at the University of Bremen, post-editing effort is multidimensional. Therefore, other aspects should be taken into account, such as the time needed to post-edit a text to the expected level of quality (temporal effort), the number of edits performed (technical effort), and the cognitive processes required to identify and correct any issues in MT output (cognitive effort).

Another study pointed out that this is mainly because some errors in MT output can easily be identified but require many edits, while other errors may require a few keystrokes to be corrected but involve considerable cognitive effort.

Eye tracking has now become a powerful tool in scientific research and its visibility in the market is steadily increasing. Given that technological advancements in the area are expected to continue, the global eye-tracking market is projected to grow 26% over the next four years, from USD 368m in 2020 to USD 1.75bn by 2025.