Reader differences in navigating English–Chinese sight interpreting/translation

2.1 Reader differences and “uniform” representation of reading in SiT studies

Jakobsen and Jensen (2008) examine six professional translators (including one interpreter) and six translation students’ eye behaviour in four tasks: (1) normal reading, (2) reading for translation, (3) reading during L2-L1 SiT, and (4) reading while translating (p. 106). Task-wise, the difference was non-significant for gaze time (i.e., duration of all fixations), and partly significant for mean fixation duration (MFD). Nonetheless, fixation frequency significantly differentiates between the tasks and groups. On average, practitioners complete each task faster than the students. However, the difference comes from fewer instead of shorter fixations, and the two groups are parallel regarding gaze time. Alves et al. (2011) also look at both fixation count and duration of six translators and six translation students, who conducted L2-L1 reading for (1) comprehension, (2) summarising, and (3) SiT. While MFD turns out to be statistically significantly longest for SiT, the trend is the opposite for fixation count. Interestingly, both indicators show that the students resemble the professionals in their reading behaviour.

Based on the above two studies, Wang and Yan (2018) examine six professional interpreters’ and six interpreting students’ behaviour during L2-L1 reading for (1) comprehension, (2) summarising, (3) translation, and (4) SiT. As task time and fixation count increase in tandem with task complexity (except between Task 2&3), MFD presents a slightly different picture: SiT is significantly longer, while the others are statistically identical. Descriptive statistics show that professionals tend to use less time and have fewer fixations and shorter MFD (for which SiT is an exception in that the professionals’ is actually longer). Later, He and Wang (2021) again examine L2-L1 SiT, using eight professional interpreters and eight interpreting students. Task-wise, similar to Jakobsen and Jensen’s (2008) results, there is no between-group difference in total fixation duration (or gaze time), but professionals do use significantly fewer fixations than their counterparts. This study further delves into local areas of interest (AOIs), i.e., low-frequency words/metaphors, which really showcase the between-group differences: the trainees’ numbers for total fixation duration, fixation count, regression time (i.e., how long the participant looks back at the AOIs), and the number of regressions all significantly exceed the other group.

Findings from the above studies converge on several fronts and show that professional experience is reflected in the task time and fixation count, but not (mean) fixation duration nor gaze time. Furthermore, details are where the expertise shines: in local problem triggers such as low-frequency words, professionals outpace trainees and re-read less. Complementing the above observations, Chmiel and Lijewska (2019) compare the performance of professional interpreters against interpreting trainees in L2-L1 SiT (but note that individual sentences were used instead of full texts). On average, professionals gaze away less than trainees, especially in objective-relative sentences, indicating stronger tolerance of visual interference (see also Jakobsen and Jensen 2008). Gaze time and task time again demonstrate the expertise advantage. That said, quite some similarity is still noted, including working memory capacity and the strategies used to chunk object-relative sentences, for which “professionals were not more autonomous in their reformulation choices” (Chmiel and Lijewska 2019: 392).

Until now the research design sheds light on the effects of experience (although earlier studies tend to use translators and later ones interpreters). Chmiel and Mazur (2013) slightly deviates from this paradigm in comparing the SiT performance of ten interpreting students with one year of training against eight with two years of training in L1-L2 SiT. The design thus indirectly looks at exposure to the task in training (as there were no dedicated SiT training but rather scattered as an activity in class), not professional contexts. It turns out an additional year does not yield significant impact on task time, fixation count, and the processing of local AOIs including sentences with specific structures and low-frequency terms. The similarity even extends to the fixation count in the warm-up session, in which participants perused the material for 10 s, although less experienced students embraced a different reading approach and scanned to cover larger ground.

Some other studies, using the same comparative paradigm, look beyond the whole task of SiT and drill down into different reading stages. McDonald and Carpenter (1981) asked two professional translators and two untrained bilinguals to perform L2-L1 SiT using garden-path sentences embedded in context and report that the first reading pass is generally normal reading, followed by reformulation (second pass), and error rectification (following passes) if the initial understanding of the ambiguous elements seems erroneous. The reading (speed) of professionals and amateurs were largely (statistically) similar, with one exception: one amateur read unusually fast in the first pass (seemingly adopting a different approach); interpreting speed was also comparable across groups. The between-group difference therefore seems to be quality (accuracy), not speed.

The above findings are mostly supported by later studies that find statistical similarities between normal and SiT reading in early stages, which include decoding words and deriving meaning from them, regardless of SiT training or professional interpreting experience. Ho et al. (2020) tease out the effects of training using 18 interpreting students and 18 untrained bilinguals in three L2-L1 tasks – (1) SR, (2) RA, and (3) SiT. Results for task time, fixation count, and mean fixation duration attest to the comparable language proficiency of the two groups in SR, while in RA the trainees use significantly fewer fixations, although results for the other two indicators remain comparable. Contrarily, the trainees exhibit statistically significantly different behaviour, completing SiT in less time and fewer fixations, rendering MFD the only exception, all the while achieving significantly better accuracy and delivery scores. Interestingly, local word-based analysis presents a counterintuitive picture: the two groups fail to differ in both early and later-stage reading processes in all three tasks.^[1] Using the same experimental setup, Ho (2021) compares the performance of 17 experienced interpreters and 18 interpreting students in SiT. The professionals turn out to use statistically similar task time, fixation counts, and MFD as the trainees, scoring significantly higher on accuracy but comparable on delivery. Word-based reading analysis also confirms the between-group similarities in the first and non-first pass reading.

From a quick combing of the SiT literature, three observations emerge: (1) the goal (and reading instructions) differentiates SiT reading (partly) from other reading-related tasks; (2) professionals differ from trainees in aggregate indicators, e.g., total fixation duration/frequency, but rarely in mean fixation duration or even word-based indices such as first fixation duration and gaze duration – in other words, the de facto reading behaviour bears much resemblance; (3) there is only one form of reading, represented (almost) solely by duration.

2.2 Insights from reading research

Numerous studies have actually found that reading is multi-faceted, consisting of several reading processes (Carver 1990; Olivier et al. 2022; Simola et al. 2008). While the conclusion remains that purpose affects reading behaviour, acknowledging the multiplicity of reading is important. An obvious benefit is that we can better understand why fixation duration varies more and sometimes increases dramatically in such more demanding reading-related tasks as SiT, leading to a higher mean duration. This understanding also offers a complementary perspective to explain why the overall fixation counts and regressions are higher in SiT than in normal reading for comprehension.

Carver (1990) systematically reviews relevant studies and concludes that reading consists of five processes, including scanning, skimming, rauding (i.e., normal reading), learning and memorising (pp. 12–22), which are utilised by readers to fulfil the goal of reading. As his focus is on reading speed, each process is associated with a standard (model) speed, represented in the format of Word per minute (Wpm). Here the Word denotes a standard-length word stretching six-character spaces (Carver 1977) – Wpm accounts for the word length variation, which intricately affects reading (Clifton et al. 2016). The representative Wpm and corresponding fixation duration per word for each process are reported by Carver (1990) as follows: scanning (600 Wpm & 100 ms), skimming (450 Wpm & 133 ms), rauding (300 Wpm & 200 ms), learning (200 Wpm & 300 ms), and memorising (138 Wpm & 433 ms). Worth mentioning is that (1) saccades generally shorten as the duration lengthens, and (2) the rauding process varies little across participants and tasks, while the other four processes can “vary more when executed by different individuals and…within individuals on different occasions or conditions” (Carver 1990: 22). The second point raised here is especially worth noting when comparing SiT reading in our study with Carver’s work. We anticipate seeing the same rauding process in our data as well, as the results deriving from different tasks and participants in previous studies have converged on this front. On the other hand, we might see more variations in other processes or even processes that have not been found in previous research as a result of the tasks we use and such background variables as professional experience and training.

Other studies have similarly identified plural reading processes. Simola et al. (2008) report three processes in information search tasks, including (1) scanning (135 ms), manifested by saccades longer than rauding and fewer regressions, (2) normal reading (200 ms), with a saccade that extends around one word, and (3) decision making (175 ms), which has slightly more regressions than rauding but shares similar saccade length. Using a different task that requires participants to judge topic-text relevance, Olivier et al. (2022) have identified four processes: normal reading (304 wpm; 197 ms), fast reading (509 wpm; 118 ms), slow confirmation (263 wpm; 228 ms), and information search (183 wpm; 328 ms) (we added duration to make comparison easier). The two studies here have corroborated Carver’s (1992) claim: Rauding remains similar across participants and tasks, while the other processes can take various forms according to the task in question and therefore fluctuate more.

Discerning what reading processes are involved in SiT complements the current SiT scholarship, as it offers a unique perspective to account for the fluctuations in fixation durations observed in different studies. Namely, fixation duration may reflect cognitive load, which seems to be the focus of the majority of the SiT studies, but it might not be the only reason. When examined together with saccade length and direction, shorter fixation duration with a longer outward regression could indicate a different reading approach, such as preparing to revert back to an earlier region of the text to locate the syntactic asymmetry between the source and the target languages to facilitate reorganising the sentence structure and delivering one’s rendition following normal reading for comprehension (Lijewska et al. 2022). Moreover, as current research heavily centres on fixation duration and frequency, the inclusion of saccade can further help describe the features of SiT reading. Therefore, our study aims to understand SiT reading, drawing on the methodology and findings of the studies introduced in this subsection. Two indicators are considered to identify reading processes, including fixation duration and (outgoing) saccade, which also carries information about directionality. We calculate the number of crossed words (NCW) (Olivier et al. 2022) in place of absolute saccade length, which frequents in previous research. NCW is easier to interpret, as what absolute saccade length means varies based on the length of the fixated word(s). Cluster analysis is used to classify reading processes, inspired by Hyönä et al. (2002).

3.1 Participants

The data of three groups of participants were analysed, including 17 experienced interpreters, 18 interpreting students, and 18 bilinguals without interpreter training. All participants considered themselves to be native Chinese speakers and deemed English as their first foreign language. The interpreters all had at least 150 days of professional experience, while the trainees all received equivalent postgraduate interpreter and SiT training. All participants met the language proficiency requirement (IELTS 6.5), had normal or corrected-to-normal vision, and signed the informed consent. The interpreters turned out to have significantly higher English proficiency (M = 8.08) than the trainees (M = 7.58) and the untrained bilinguals (M = 7.27), but the latter two were statistically equal; working memory size was comparable across groups (details see Ho 2017).

3.2 Materials and research design

Three 175-word English speech scripts were used. All were adapted from different speeches by the same speaker and offered an overview of the World Trade Organization (WTO). As Table 1 shows, the three scripts are considered similar, regarding not just the percentages of passive sentences, Flesch reading ease score, and Flesch-Kincaid grade level, but also the difficulty level rated by both the untrained bilinguals in the original project and the interpreting students recruited in a follow-up project.

Each participant had to conduct SR, RA, and L2-L1 SiT once, each using a different script. The order of the tasks and scripts were counter-balanced within and across groups. Being too long to fit into one computer screen (1,024 × 768 pixels), each script was shown on four consecutive screens (named trials; up to six lines of text per trial). Sentences in every trial were complete, meaning the participants did not have to move to the next one to finish reading a sentence. Going backward to previous screens was not possible. The scripts were projected on a grey background in 22-point Courier. Eye movement data were sampled at 1,000 Hz using Eyelink 1000. Each participant was seated at around 70 cm from the monitor.

3.3 Procedure

The experiment began with an introduction to the research project and the self-paced tasks involved. The participants were advised to conduct the SiT task with an audience in mind, as if they were hired to interpret at a real conference (Setton and Dawrant 2016). They were also informed that they could not scroll back to previous trials once they moved on. Nine-point calibration then followed, and, if successful, the participants would engage with all three tasks in turn. Each task was preceded by a warm-up task of the same nature and followed by two comprehension questions to ensure genuine engagement. Eye-tracking accuracy was recalibrated in between tasks and trials when necessary.

3.4 Data analysis

The data from a total of 53 participants mentioned previously in Section 3.1 were included. For the purpose of the current study, fixations shorter than 80 ms and longer than 1200 ms were removed (Drieghe et al. 2008; Ma et al. 2021; White 2008) and 6.8 % of the data were hence deleted. Two features of each fixation were then chosen for cluster analysis: fixation duration and NCW, which represents the length of outward saccade plus its direction. Around 3 % of the data without NCW were further deleted. Finally, a total of 57,805 fixations underwent cluster analysis using the Partitioning Around Medoids (PAM) algorithm (Kaufman and Rousseeuw 1990). We analysed the full dataset together instead of examining each task separately, although we acknowledge each task may require distinct processes. The rationale is that the fundamental processes, such as decoding words and integrating meaning, are still shared across tasks (see Section 2.2 on overlapping processes between studies on various reading-related tasks). SiT studies investigating reading passes have also shown that reading is similar between normal reading for comprehension and SiT in the first pass, substantiating the claim of shared processes to some extent (Ho et al. 2020; Ho 2021; Lijewska et al. 2022), hence our decision of one single cluster analysis. Another major reason relates to how we used cluster analysis, which would single out unique processes if they were indeed peculiar to a certain task, as we did not pre-determine how many clusters our data should have but were rather informed by PAM (more details see below).

Clustering means partitioning a dataset into “clusters”, with data points within the same cluster sharing more similarity than those in different ones. The idea is to identify clusters that are as distinctive as possible from each other (Schubert and Rousseeuw 2019). Prior to any analysis, the dataset was first examined using the Hopkins statistic (H), which “is used to assess the clustering tendency of a dataset by measuring the probability that a given data set is generated by a uniform data distribution. In other words, it tests the spatial randomness of the data” (Kassambara 2017: 123). An H-value close to 0.5 means the data set is uniformly distributed, hence no possibility of finding genuine clusters. The smaller the H-value, the more possible the dataset can be meaningfully clustered. Our data turned out to be highly clusterable (H = 0.018).

PAM was used in our study as a widely used algorithm and a robust alternative to k-means clustering. For k-means, a cluster centre is the mean value of all data points in the said cluster, whereas for PAM the centre (called medoid) of a cluster is the one presenting minimal dissimilarity to all the rest data points, making it “the most centrally located point in the cluster” (Kassambara 2017: 48), hence less sensitive to outliers. To know how many clusters will best partition the dataset and achieve best compactness for each cluster (i.e., minimal within-cluster difference) while maximising the average distance between clusters (Brock et al. 2008), silhouette width index (Rousseeuw 1987) was utilised. The silhouette width for each fixation can range from −1 to +1. The higher the value, the better, whereas anything below 0 indicates misclassification (ibid.).

Based on previous studies on identifying reading processes using fixation duration and saccade length (e.g., Carver 1990; Simola et al. 2008), we anticipated the possibility to find three to seven different reading processes and therefore examined the average silhouette width of each of these possibilities. We then picked the one with the highest average silhouette width and used the REMOS algorithm (version 2) (Lengyel et al. 2021) to optimise cluster allocation and ensured all data points ended with positive silhouette widths.

Each fixation in our data was assigned to only one cluster. The mean frequency of each cluster by task and by group was then calculated and then log-transformed before conducting between-subject and within-subject analysis using repeated measures ANOVA.

4.1 Cluster analysis

The analysis resulted in five distinct clusters (i.e., reading processes). Table 2 presents the mean, median, and standard deviation (SD) of fixation durations and NCWs for each cluster. The number of clusters found in this study intuitively corroborates Carver’s (1990) findings, but the features of some clusters are worlds apart.

Fixation count shows that Cluster 1 & 2 are the majority, accounting for around 36 % and 45 % of all fixations respectively, while the frequencies of the other three clusters range from infrequent to rare. Cluster 1 has an MFD of 150.4 ms (SD = 33.21 ms; Mdn = 156 ms), akin to what Carver (1990) defines as skimming. On the other hand, Cluster 2 has an MFD of 260 ms (SD = 39.31 ms; Mdn = 255 ms). This seems to sit between the process of rauding (200 ms) and learning (300 ms) stated by Carver (1990), whose results about rauding have further been corroborated by Simola et al. (2008) and Olivier et al. (2022) even in tasks of different nature.

Starting from Cluster 3 onward, appearance is much more infrequent, ranging from around 15.3 % for Cluster 3 (M = 423.97 ms; SD = 62.44 ms; Mdn = 409 ms), 2.8 % for Cluster 4 (M = 735.85 ms; SD = 142.18 ms; Mdn = 689 ms), to 0.8 % for Cluster 5 (M = 264.24 ms; SD = 158.58 ms; Mdn = 222.5 ms). These three types of fixations are what really differentiate our findings from previous studies. While Cluster 3 can still claim to be similar, in terms of how demanding it is, to the memorising process proposed by Carver (1990), Cluster 4 is nowhere close to any of the categories presented in the literature about reading processes. Cluster 5 seems to be reflecting a categorically different process when compared with Cluster 2, with the SDs of the two clusters being wide apart.

The uniqueness of Cluster 5 is further substantiated by Figure 1 below. It is expected to see the SD widening from Cluster 1 to Cluster 4, with the increasing variability reflecting the more demanding nature of the process in question. However, Cluster 5 shows the utmost level of fluctuation, suggesting that this type of fixation indicates a different cognitive process that leads to wide variability across tasks and participants.

Figure 1:

Boxplots showing the mean (square), median (horizontal line), and the distribution of each cluster.

A look at NCW can shed more light on what Cluster 5 represents. For NCWs, all except Cluster 5 have a median of 1 NCW, which means the fixation jumps onto the immediately following word for half of the time. Cluster 5 stands out on this indicator with a mean NCW of −31.83 (Mdn = −30), and a much wider SD of 10.99. We further checked the range and realised it is exclusively a “regressive” cluster, with the shortest regression shooting across 17 words to earlier parts of the text.

After examining the MFD and NCW of each cluster, we argue that Cluster 1 reflects the skimming process, while Cluster 2, the most common type of fixation, is probably rauding (in Carver’s terms) – or normal reading for comprehension – albeit the MFD is slightly higher than what Carver (1990) reports (but in Rayner 2009 the repoted range of SR overlaps to a large extent with our findings; also see Shreve et al. 2010, which uses bilingual reading and shows our findings are comparable). Cluster 3 & 4 could represent similar processes that are equally or more demanding than the memorising process mentioned by Carver (1990). Here, we would hypothesise that, due to the formality of the source scripts and the context in which these speeches take place in real life, the two clusters could be lumped together to indicate problems experienced by the participants, with Cluster 4 reflecting problem-triggers of a greater magnitude – semantically or syntactically. We therefore label these two clusters as “problem-solving” processes. Lastly, Cluster 5 could be “anchoring” fixations. That is, this type of fixation signals long regressions for the reprocessing or integration of syntactically dense sentences, which could stretch across four lines of text on the screen in this study.

4.2 Differences and similarities in the reading processes across groups and tasks

Table 3 presents the mean frequency distributions of the five clusters for each group across the tasks. This information, together with Figure 2 on the distribution of clusters within each group, mainly helps us tackle the first research question – what reading processes are involved in SiT and how they differ in SR and RA respectively – but also partly addresses the second question on how different groups vary in the combination of processes when tackling different tasks. Table 3 shows that the total frequency of fixations in SR and RA sits in the range around 200–300, but SiT is on a different level: around 400 to 800. Using SR as the baseline, the between-group gaps in frequency seem to be the smallest in RA and dramatically widen in SiT, though mainly between untrained bilinguals and the rest two groups. Another observation is that untrained bilinguals almost consistently had the highest mean fixation count for all clusters in all tasks, followed by interpreting trainees, and then experienced interpreters. There were only few exceptions, including Cluster 4 in SR and RA – where trainees had the highest number – and Cluster 5 in SiT, which experienced interpreters used almost equally often as trainees. As all the interpreters except one received the same training as the trainees,^[2] a clear message is that training here does exert an obvious influence on the overall reading behaviour, leading to much fewer fixations in SiT.

Figure 2 visualises the use of various reading processes across groups in the three tasks. The percentage is presented to focus on the relative proportion of fixation counts for the five clusters within each group and each task.

Figure 2:

Fixation distributions (in percentage) of the five clusters for each group across tasks.

As Figure 2 illustrates, the pattern of distribution for all tasks is largely similar. Cluster 2 (rauding) is the most common type, followed by Cluster 1 (skimming), Cluster 3 (problem-solving), Cluster 4 (effortful problem-solving), and then Cluster 5 (anchoring). The last two clusters show minimal presence.

Using SR as the baseline, we can see how percentages of processes vary when there is an additional subtask of articulation in RA and when additional subtasks of interlingual reformulation and articulation are required in SiT. Across all tasks, the gap between Cluster 1 & 2 in RA seems to be the largest of all tasks; on the contrary, the same gap in SiT appears to be the smallest. In addition, Cluster 3 in RA accounts for the largest proportion, followed by SiT and then SR. In terms of Cluster 4, there is an obvious uptick in RA and SiT when compared with SR.

We used repeated measures ANOVA to examine the effects of group and task on the frequency of each cluster. Omega squared was chosen to report effect size and was manually calculated following the instructions of Mellinger and Hanson (2017). Table 4 demonstrates three major phenomena: (1) Cluster 1 & 5 report effects of group and task, without interaction; (2) Cluster 2 & 3 show effects of group and task with interaction; (3) Cluster 4 only has a significant effect of task with interaction between group and task.

The variable of group has an significant impact on the frequency of almost all clusters: Cluster 1, F(2, 50) = 8.16, p = <0.001, ω² = 0.213; Cluster 2, F(2, 50) = 12.5, p = <0.001, ω² = 0.303; Cluster 3, F(2, 50) = 8.65, p = <0.001, ω² = 0.224; Cluster 5, F(2, 50) = 3.53, p = 0.037, ω² = 0.087. Cluster 4 is the only non-significant cluster, F(2, 50) = 2.09, p = 0.134, ω² = 0.04.

On the other hand, all clusters show statistically significant effects of tasks. That is, the task conducted does affect the use of reading processes differently, including Cluster 1, F(2, 100) = 113.05, p = <0.001, ω² = 0.809, Cluster 2, F(2, 100) = 95.98, p = <0.001, ω² = 0.782, Cluster 3, F(2, 100) = 172.21, p = <0.001, ω² = 0.866, Cluster 4, F(2, 100) = 95.63, p = <0.001, ω² = 0.781, and Cluster 5, F(2, 100) = 8.64, p = <0.001, ω² = 0.224. The effect size is surprisingly large throughout, except for Cluster 5. On top of this, how reading processes are used differently in each task is significantly related to some extent with which group the participant belongs to for Cluster 2, F(4, 100) = 6.54, p < 0.001, ω² = 0.295, Cluster 3, F(4, 100) = 7.73, p < 0.001, ω² = 0.337, and Cluster 4, F(4, 100) = 2.87, p = 0.027, ω² = 0.124.

4.3 Reader differences come from training, not professional experience

For this subsection, we delve into the post-hoc analysis (with Bonferroni correction) of the between-subject effects. There is no interaction effect for Cluster 1, which means the difference between groups observed does not vary across tasks. The untrained bilinguals had significantly more Cluster 1 fixations (skimming) than the trainees, t(50) = 3.019, p = 0.012, and the experienced interpreters, t(50) = 3.823, p = 0.001, respectively. The latter two groups were statistically identical, p = 1.

Cluster 2 presents a more complicated relationship between task and group, as shown in Table 5. It turns out that there was no between-group difference in SR or RA. By contrast, in SiT the untrained bilinguals used significantly more Cluster 2 fixations (rauding) than trainees, t(50) = 4.996, p = <0.001, and experienced interpreters, t(50) = 5.169, p = <0.001. Similar to Cluster 1, there was no difference between the latter two groups in all tasks, p = 1 throughout.

For Cluster 3, the detailed comparisons are included in Table 6. The main group effect has three contributors. The untrained bilinguals resorted to problem-solving processes significantly more than experienced interpreters in SR, t(50) = 3.3899, p = 0.049. More importantly, in SiT, the cluster frequency of the bilinguals was much higher than that of the trainees, t(50) = 4.057, p = 0.006, and the interpreters, t(50) = 5.8035, p = <0.001.

The disparities between groups are most noticeable from Cluster 1 to Cluster 3, and there is no group effect for Cluster 4. As for Cluster 5, although the main effect was significant, post-hoc analysis shows no difference between any two groups (p = 0.067 for BIL vs. INT, p = 0.091 for BIL vs. PRO, and p = 1 for INT vs. PRO).