Interpreting is a form of translation where an immediate and singular production in a target language is produced based on the one-time presentation of an utterance in a source language (Kade, 1968; Pöchhacker, 2016). Since Interpreting Studies (IS) take interpreting as its object of investigation, it is seen as a subdiscipline of the larger domain of Translation Studies (Pöchhacker, 2010). Salevsky (1993) classifies IS into the theoretical, descriptive, and applied domains, which covers topics including process, product and performance, practice and profession, and pedagogy (Pöchhacker, 2016). Among them, process research, that is, studies on the identification, decoding, transferring, and producing of linguistic units from one language to another (Nida, 1964), is a dominant theme in empirical IS (Pöchhacker, 2016). According to Gile (2009a) effort model, changes in the processing load of interpreters have an impact on the product; therefore, it is of high relevance to investigate and understand the process.

As interpreting is a complex linguistic activity (Hale, 2010; Stachowiak-Szymczak, 2020) and researchers do not have immediate access to interpreters’ brain, they resort to various methods to look into the “black box” of interpreting processing. These methods can be classified as offline or online ones. Offline methods refer to methods that are temporally separate from the main task, usually after it (Godfroid, 2019). Such an offline method is retrospective protocol, popular in interpreting studies (Dimitrova and Tiselius, 2014), boasts operational convenience, and yields qualitative data on the cognitive mechanisms behind certain phenomena, thereby providing answers to “why” they take place (Dimitrova and Tiselius, 2009). However, retrospective protocols are prone to subjectivity and lower accuracy, as there may be a gap between what is perceived or reported and what really happened (Ericsson and Simon, 1984; Ivanova, 2000). Moreover, offline methods fail to capture the moment-to-moment interpreting process and may miss crucial details of it.

Online methods, in contrast, enable the concurrent, real-time observation and collection of data during the task under study (Godfroid, 2019). They primarily include two subcategories: performance methods, such as analysis of ear-voice span (EVS), disfluencies, and omissions, and psychophysiological methods, such as eye-tracking, event-relatesd potential (ERP), and positron emission tomography (PET) (Tirkkonen-Condit and Jääskeläinen, 2000; Timarová, 2010; Seeber, 2013; Godfroid, 2019). Between these two subcategories, performance data are more readily available, as collection and analysis of them usually do not require additional and complicated apparatuses. For instance, researchers have easy access to performance data with recordings of the interpreting tasks. But compared with psychophysiological methods, such as functional magnetic resonance imaging (fMRI) that provides information on the process- brain activities, performance methods are more subjective, as researchers need to infer the interpreting process from the product. According to Seeber (2013), psychophysiological methods are more objective, since they record ocular, cerebral, and other physiological responses, which cannot be consciously controlled. Furthermore, they generate concurrent data that allow for moment-to-moment investigation of interpreting processes. With some psychophysiological methods, however, the trade-off for these merits is low ecological validity, that is, the representativeness or naturalness of experimental situations, materials or task, or how much they resemble the actual, real-life conditions (Kvavilashvili and Ellis, 2004; Spinner et al., 2013). Since the closer the resemblance, the more generalizable findings from the interpreting study can be, high ecological validity is usually a desirable feature (Kvavilashvili and Ellis, 2004). But such a virtue is hard to come by for some psychophysiological methods. For example, in studies with ERP, participants need to stay stationary, but real-life interpreting often involves head movements, hand gestures, or even body movements of interpreters. Therefore, the constraints posed on interpreters by ERP apparatus would influence the interpreting process and make it unnatural, or less authentic.

Among popular psychophysiological methods in IS, such as eye-tracking, PET (Rinne et al., 2000; Tommola et al., 2000), ERP (Proverbio et al., 2004; Koshkin et al., 2018), and fMRI (Hervais-Adelman et al., 2011; Becker et al., 2016; Van De Putte et al., 2018), eye-tracking has the advantage of higher ecological validity, high temporal and spatial accuracy and easier operation (Godfroid, 2019; Tiselius, 2020). To start with, the emergence of more flexible and less intrusive eye trackers, such as eye-tracking glasses, allows for movements of the interpreters (Mele and Federici, 2012). Therefore, it is possible to document eye movements in a more normal or actual state of interpreting. Secondly, eye trackers offer high temporal resolution of milliseconds and high spatial resolution at the sub-lexical level (Godfroid, 2019). Another advantage of eye-tracking over other psychophysiological methods is its operational convenience, such as instant configuration, execution, and calibration (Conklin and Pellicer-Sánchez, 2016). Taken together, it is a non-intrusive, relatively easy-to-operate method that allows for real-time, fine-grained investigation into the dynamics of interpreting processing (Hvelplund, 2014; König et al., 2016; Tiselius, 2020).

By definition, eye tracking refers to “an experimental method of recording eye motion and gaze location across time and task” (Carter and Luke, 2020, p. 50) that provides significant information about the operation of the human brain (Rayner, 1978; Leigh and Zee, 2015). Eye movements generally include fixations, saccades, and other types of ocular motion such as vergence and pupil dilation or constriction (Carter and Luke, 2020). The hypotheses that link eye movements to cognitive activities are Just and Carpenter’s (1980) eye-mind and immediacy assumptions, which postulate that what is being fixated is what is being processed, and there is no appreciable lag between the two. Despite limitations of these assumptions, for example, they cannot explain the phenomenon of mind wondering independent of eye movements (Shepherd et al., 1986; Smallwood and Schooler, 2006), they are regarded as the theoretical underpinnings of eye-tracking (Seeber, 2015). The history of eye-tracking studies dates back to the 1870s, when Javal first identified two major types of eye movements during reading: fixations and saccades (Yan et al., 2013; Płużyczka, 2018). The method first became popular in psychology and psycholinguistics, data from which were seen as “the gold standard for experiments” (Rayner, 2009, p. 1474). It was then introduced in IS by McDonald and Carpenter (1981), who investigated the interpretation and parsing of ambiguous idiomatic phrases, as well as error detection and correction in sight translation, a variant of interpreting (Herbert, 1952; Agrifoglio, 2004; Chen, 2015; Pöchhacker, 2016). Later, it was increasingly adopted in linguistics, as evidenced by a growing number of studies in this field using eye-tracking in the 21st century (Conklin and Pellicer-Sánchez, 2016), and is now gaining ground in IS (Korpal, 2015; Ma, 2017; Tiselius, 2020).

However, despite the increasing popularity of this method in IS, research that delineates a comprehensive and systematic picture of eye-tracking in IS seems to be lacking. We argue that it is necessary to have a clear and comprehensive overview of what can be done with this method, what has been done, and what still needs to be done. With this landscape of eye-tracking interpreting research, researchers can better understand how their research fits into it. This is especially valuable for researchers unfamiliar with this method, and aiming to explore new territories with eye-tracking in IS. Moreover, it is of theoretical relevance to conduct a synthesis on current interpreting studies with eye-tracking and on findings from them. For instance, these findings can offer valuable information on the underlying interpreting mechanisms, and can be applied in building new interpreting models or validating existing ones, such as Gile’s (2009a) Effort Model. In addition, these findings, for example, on eye movement patterns in case of difficulties, in the employment of specific strategies, and their differences between trainees and professionals also contribute to interpreting training. Despite the insights afforded by previous reviews of eye-tracking IS, they are somewhat limited in breadth and scope, as the focus is either several selected studies (Moratto, 2020), or a specific research dimension—methodological issues (Korpal, 2015), or a specific theme—cognitive load (Su et al., 2021), or a specific interpreting mode—sight interpreting (Ma, 2017). Thus, this is the first attempt at a comprehensive review of eye-tracking in IS in its entirety. With this review, we aim to identify common practices and problems, as well as explore the potential of eye-tracking in IS. Specifically, we (1) review publication trend of previous studies; (2) analyze the design features of these empirical papers, including participants, interpreting modes, and eye trackers used; (3) summarize their contributions—key themes discussed and eye-tracking measures used; (4) discuss the limitations and their implications, and (5) propose potential avenues for future research. This is of academic, theoretical, and practical relevance, as it can help researchers get an overview of extant literature, design future studies, and gain insight on what eye-tracking stands to offer, deepen our understanding of interpreting mechanisms, and inform interpreting training, for example, through identification of interpreting difficulties and differences of eye movement patterns between professionals and trainees.

This study follows Cooper’s (2016) comprehensive and widely adopted framework of research synthesis with the following stages: formulating the problem, searching the literature, gathering information form studies, evaluating the quality of studies, analyzing and integrating the outcomes of studies, interpreting the evidence, and presenting the results. Since the aim of this study is to identify how eye-tracking is employed in IS, empirical articles are selected, and to ensure the representativeness of data, we choose peer-reviewed journals as the data source.

Despite their contribution to IS studies, two limitations identified in some studies in this review deserve our attention.

The first one concerns the incomplete reporting of key information about data collection and processing, which undermines data quality and research replicability. For instance, some studies (Dragsted and Hansen, 2009) neglected essential information about eye trackers’ specifications, including sampling rate, precision, accuracy, and latency (Conklin et al., 2018). These four fundamental attributes determine eye-tracking quality, and the right choices concerning them need to be made for the right research. In reading research, for example, eye-tracking systems with a sampling rate of 500 Hz or 1,000 Hz are preferred, and if a low sampling rate (the criteria is under 250 Hz) is adopted for studies with small AOIs, a multitude of data are required to compensate for inaccuracies arising from it (Holmqvist et al., 2011; Conklin et al., 2018). Hence, a detailed account of information about data collection constitutes a basis on which to judge data quality and robustness of the research. Another rationale for increasing transparency in reporting eye trackers’ specifications is that differences in them are found to yield different results. For instance, Leube et al. (2017) found that sampling rate affected saccade detection. More specifically, in a reading task, the 120 Hz mobile eye tracker boasted a significantly higher saccades detection rate and a more accurate estimation of the mean saccade duration than the 60 Hz one. Therefore, reporting of information concerning data collection, including but not limited to eye trackers’ specifications can contribute to research comparability and replicability. In addition, important information on eye-tracking data processing, such as data filter settings, is found missing in some studies (Bosch-Baliarda et al., 2020), and these settings have a direct impact on results. For example, it has been found that effects were significant when inclusion criterion for fixations is set at longer than 140 ms, but insignificant at 50 ms (Yan et al., 2013). Most eye-tracking software allows researchers to decide filter settings, primarily, fixation duration and maximum distance between 2 gaze samples. In order to improve research replicability, information on them needs to be reported, as the ability to replicate findings is crucial to scientific advancement (Open Science Collaboration, 2015).

For these reasons, it is recommended by Godfroid and Hui (2020) that eye-tracking studies should include information on items such as eye trackers (e.g., sampling rate, precision), experiment set-up (e.g., screen size), design (e.g., font size and spacing of the stimuli), data analysis protocol (e.g., data preprocessing), and data quality.

The second limitation is insufficient triangulation, which impairs research validity. Every measure or method has inherent biases, and the adoption of a single approach or measure may lead to premature and erroneous conclusions. Triangulation, a process where researchers seek convergence with multiple data sources and endeavors to offset biases, enhances research validity (Greene et al., 1989; Han, 2018). In the case of eye tracking interpreting research reviewed here, some studies were found to draw a conclusion with a single method of eye-tracking, or a single measure from it. For instance, Korpal and Stachowiak-Szymczak (2018) used mean fixation duration as the sole indicator of cognitive effort. Fixation duration is indeed seen as indicative of cognitive effort (Rayner, 1998; Holmqvist et al., 2011), the use of a single measure, however, is likely to raise doubt on the research results, especially when eye-tracking can yield rich data from a wide array of measures, some of which are also indices of cognitive load, such as fixation count and pupil size. In addition, Korpal and Stachowiak-Szymczak (2018) used the single method of eye-tracking to extrapolate cognitive load on participants. However, this method has its share of weaknesses despite various merits. Eye movements are not the impeccable representation of cognitive mechanisms, because oftentimes cognitive shifts can happen independent of eye movements, as is illustrated by the phenomenon of wandering or drifting (Shepherd et al., 1986; Smallwood and Schooler, 2006). Therefore, eye-tracking is believed to reflect a “proximation of the relationship between visual focus and cognitive focus” (Hvelplund, 2014, p. 209). In light of this situation, it will be most potent when data obtained is cross-checked and supplemented by those yielded from other methods. Therefore, Korpal and Stachowiak-Szymczak’s (2018) study would have been more convincing if results from eye-tracking on cognitive load were confirmed by other methods, such as stimulated recall (Seeber, 2013). In summary, in order to strengthen validity, multiple eye-tracking measures need to be triangulated with other online measures such as ERP, or off-line measures, such as questionnaires, interpreting product analysis, and stimulated recall.

Based on our analysis, in future research, eye-tracking should generate increased scholarly attention on the investigation of multimodality in interpreting, especially for interpreting modes whose multimodal feature has traditionally been neglected. The necessity to process multimodal information is seen as one of the defining characteristics of interpreting (Seeber, 2017), but eye-tracking has primarily been applied to examine text-reading in sight translation, and future research will see more study into multimodal processing in DI, SI, SLI, and CI for the following reasons. Firstly, with technological development, more affordable, flexible and mobile eye trackers with higher precision would come along (Cognolato et al., 2018). This makes it easier to probe, for example, how interpreters deal with both nonverbal messages and speech of the speakers in DI, and how interpreters handle both text and speech in SI with text. Moreover, with the popularity of automatic speech recognition and of distance interpreting in the wake of COVID-19, interpreters are more likely to be faced with multimodal input than before. It is not uncommon that service providers transfer speech to text and project them on screen real time for simultaneous interpreters and audience (Li, 2020), thereby offering them both aural and visual information. In light of these new trends, Pöchhacker (2019) identified intermodality as one of the five novel features of interpreting. International Association of Conference Interpreters (2019) also issued guidelines on distance interpreting in 2019, where they recommended large and clear enough screens for interpreters to see texts, images, speakers, participants, and the conference room clearly. Against this backdrop, growing popularity of applying eye-tracking into the study of multimodal interpreting processing is likely.

In addition, it would be interesting to investigate with eye-tracking how interpreting processes differ between different language pairs and between long and short segments. Studies can also be conducted to examine the processing of interpreting abbreviations, enumerations, idiomatic expression, and accented and technical speech. Furthermore, eye-tracking has great potential for interpreting testing. Recent years has seen an increasing number of studies that used eye-tracking to evaluate L2 reading, listening, and speaking assessment, and they mainly focused on test validity (Conklin and Pellicer-Sánchez, 2016; Godfroid and Hui, 2020). Similarly, eye-tracking can be a powerful tool for interpreting assessment. For example, researchers can compare online eye movements of disparate groups, such as professional interpreters and interpreting trainees, for the same test. The results provide insight into the discriminating power of this test, in other words, test validity. Lastly, in future studies, saliency approach could be adopted in IS with eye-tracking and contribute to the cross-fertilization of both disciplines of IS and computer science. As a data analysis method that can be used to predict eye movement patterns like spatial location of individual fixations and their sequential order (Foulsham and Underwood, 2008), it has a wide variety of models that are extensively applied (Borji et al., 2013), but it is rarely adopted in the studies we review. Saliency models goes beyond simply describing eye movements by predicting where interpreters might look, thus deepening our understanding of interpreting. Additionally, since interpreting is a complex activity (Hale, 2010; Stachowiak-Szymczak, 2020), results from IS with eye tracking constitute valuable information for comparison between and evaluation of the performance of different saliency models (He et al., 2019).

The purpose of the current review was to synthesize papers on IS with eye-tracking so as to depict a comprehensive and systematic picture of their contributions, limitations and implications, as well as future directions. We have found that in general, comparison of different participant groups, in particular, expert-novice comparison, instead of focusing on a single group is the most popular design. We have argued for the recruitment of extra participants, given the recurring data loss in eye-tracking, and cautioned against overgeneralizing findings to interpreters on the whole, when only trainees are enrolled for investigation. In addition, STR is the most popular mode for examination. When it comes to the choice of eye trackers, remote ones are the most preferred type, and it is necessary to adopt the right eye tracker in alignment with research questions, so as to achieve a balance between ecological, external validity, and reliability. This review has also shown that eye-tracking is used for two strands of research: cognitive investigation and observation of oculomotor behaviors. The former theme includes evaluation of cognitive load and exploration of cognitive processing, and the latter covers topics of utilization of visual input, viewing patterns, and the role of gaze. As far as eye-tracking measures are concerned, we have noted that fixations were the most favored metrics for all of the foregoing themes apart from the study into the role of gaze. Researchers are also found to adopt novel measures, such as saccadic amplitude to index cognitive effort, as well as tailored measures for standardization. We have argued that in order for researchers to build on each other’s work and for scientific progress, detailed reporting of key information on data collection and processing, such as sampling rate of eye trackers, set-up information and data filtering criteria are necessary for future studies. Furthermore, the results would be more convincing with more triangulation of multiple measures and methods. In terms of future work, it would be interesting to see how more advanced eye-tracking systems can offer new insight into the multimodal processing of previously underexplored interpreting modes, such as SI, CI, DI, and SLI. Further research could also be conducted to determine the interpreting processing of different language pairs, speech of different lengths, abbreviations, enumerations, idiomatic expression, and accented and technical speech. Another interesting research agenda would be to investigate how eye-tracking can be applied in interpreting testing, especially in the examination of test validity. It is also recommended that saliency approach be applied in future studies. One major limitation of this study needs to be acknowledged. We included only peer-reviewed empirical papers and English publications for the review. Notwithstanding this limitation, the present study, as the first attempt at a comprehensive review of IS with eye-tracking, should prove valuable for practitioners, educators, and researchers alike. Findings reported here will offer insight into the nature of interpreting, have pedagogical implications for interpreting educators, and inform researchers’ efforts in using eye-tracking in conducting interpreting research.

TH is responsible for the conceptualization, design, data collection and analysis, and writing of the manuscript. XW is responsible for the conceptualization, data collection and analysis, visualization, and writing of the manuscript. HX is responsible for the conceptualization, design, and suggestions of revisions. All authors contributed to the article and approved the submitted version.

This study was funded by Zhejiang Conservatory of Music (G001A3032033).

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.