Psychon Bull Rev DOI 10.3758/s13423-014-0682-6
BRIEF REPORT
Skipped words and fixated words are processed differently during reading Michael A. Eskenazi & Jocelyn R. Folk
# Psychonomic Society, Inc. 2014
Abstract The purpose of this study was to investigate whether words are processed differently when they are fixated during silent reading than when they are skipped. According to a serial processing model of eye movement control (e.g., EZ Reader) skipped words are fully processed (Reichle, Rayner, Pollatsek, Behavioral and Brain Sciences, 26(04):445–476, 2003), whereas in a parallel processing model (e.g., SWIFT) skipped words do not need to be fully processed (Engbert, Nuthmann, Richter, Kliegl, Psychological Review, 112(4):777–813, 2005). Participants read 34 sentences with target words embedded in them while their eye movements were recorded. All target words were three-letter, lowfrequency, and unpredictable nouns. After the reading session, participants completed a repetition priming lexical decision task with the target words from the reading session included as the repetition prime targets, with presentation of those same words during the reading task acting as the prime. When participants skipped a word during the reading session, their reaction times on the lexical decision task were significantly longer (M = 656.42 ms) than when they fixated the word (M = 614.43 ms). This result provides evidence that skipped words are sometimes not processed to the same degree as fixated words during reading. Keywords Reading . Eye-movements . Word recognition During reading, approximately 30 % of words are skipped or never directly fixated (Rayner, 1998). Given this prevalence, word skipping has become an important phenomenon for models of eye-movement control to explain. While skipped
M. A. Eskenazi : J. R. Folk (*) Department of Psychology, Kent State University, Kent, OH 44224, USA e-mail: [email protected]
words reach some level of lexical access, there is considerable debate as to how deeply skipped words are processed. On one extreme, skipped words are processed by a “best guess,” and on the other extreme, skipped words are fully identified in the parafovea (Drieghe, Rayner, & Pollatsek 2005). The purpose of the current study was to determine if there are measurable differences in the way that skipped words are processed compared with fixated words. The two most prominent models of eye-movement control (EZ Reader and SWIFT) have different explanations of how words are skipped, and thus the degree of lexical processing of skipped words has implications for each of these models. In the EZ Reader Model, each word must be fully processed before any lexical processing of the next word can begin (Reichle, Rayner, and Pollatsek 2003; Pollatsek, Reichle, & Rayner 2006). Lexical access is split into two stages: a familiarity check (L1), and the completion of lexical access (L2). The completion of L1 indicates that lexical access is imminent and that it is safe to program a saccade to wordn+1. The completion of L2 initiates a covert shift of attention to wordn+1 to begin lexical processing. Word skipping occurs when lexical access is completed or imminent on wordn+1 during the covert shift of attention while the eyes are still fixated on wordn. Therefore, all words, whether they are fixated or skipped, go through the same stages of lexical access according to the EZ Reader Model. This is an important assumption for the model, because word identification explains how the eyes move forward through the text. Unlike the EZ Reader Model, in the SWIFT model, the default target of the saccade is not always wordn+1 (Engbert, Nuthmann, Richter, & Kliegl 2005). The SWIFT Model is a parallel-processing model in which up to four words can be processed at the same time. Each of these words competes to be the target of the saccade from wordn in a dynamic field of activation where lexical access for each word increases to a certain threshold of lexical access, based on each word’s
Psychon Bull Rev
difficulty. After reaching threshold, processing decreases until lexical activity is zero, similar to completion of L2. The next saccade will be directed towards the word in the dynamic field of activation that is the farthest from reaching zero lexical activity. Therefore, if wordn+2 is farther from reaching lexical completion than wordn+1, wordn+2 will be the target of the saccade, and wordn+1 will be skipped. Although, wordn+1 is usually identified, there is no requirement that its lexical activity reaches zero (Engbert et al., 2005). The assertion in the SWIFT model that skipped words do not need to be fully processed is in direct contrast with the assumption of the EZ Reader model that all words are fully processed whether they are skipped or fixated. There is support for both arguments. Engbert et al. (2005) argue that a higher probability of regressions to skipped words is evidence for incomplete lexical access (Vitu & McConkie, 2000). However, there is also evidence that at least some level of lexical semantic access has been achieved on skipped words, because predictable words are skipped more often than unpredictable words (Rayner, Binder, Ashby, & Pollatsek 2001). The question remains as to whether skipped words are processed to the same degree as fixated words. To answer this question, the current study used a repetitionpriming paradigm with the assumption that a word that had not achieved full lexical access would be a weaker prime than a word that had achieved full lexical access. A target word embedded in a sentence during a silent reading task served as the prime, and a second presentation of that word in a lexical decision task was the lexical decision target. The target word/prime and the lexical decision target were the same word, consistent with a repetitionpriming paradigm. In the repetition-priming effect, a single presentation of a stimulus increases the processing speed of a future presentation of the same stimulus. This is a robust effect that can persist at delays up to 24 hours (Scarborough, Cortese, & Scarborough 1977). We have two different hypotheses according to the two different models. According to the EZ Reader Model, both fixated and skipped words go through the same two-stage (L1 and L2) process to achieve lexical access. Therefore, fixated words and skipped words should act as equally effective primes, such that reaction times will not be different on the lexical decision task when the target word/prime was either fixated or skipped during the reading task. However, according to the SWIFT Model, skipped words do not have to reach full lexical access. Therefore, fixated words should be better primes than skipped words, because fixated words will achieve lexical access, whereas skipped words on average will have a lower level of lexical access. In the RPLDT, the average response time when the target word/prime was skipped should be longer than when the target word/prime was fixated.
Method Participants Twenty-six students from Kent State University participated in this study. All had normal or corrected vision and were native English speakers. None reported any reading disability. An additional 26 participants completed only the lexical decision task, serving as a control group for repetition-priming effects. All participants received course credit for participation in the study. Materials and design Eye-tracking Target words were embedded in sentences that contained a neutral context, and the target word was never in the first three or last three words of the sentence. For example, the target word lab was used in the sentence “Her friend had a fancy lab inside of the new building.” To increase the likelihood that a sufficient proportion of target words would be skipped during the eye-tracking part of the study, all target words were three-letter nouns and low frequency (M = 10.54, SD = 6.97 from CELEX) (Baayen, Piepenbrock, & van Run 1995). To ensure that target words were unpredictable and thus would not be guessed through top-down processing without being fully processed, a cloze task was administered to 100 participants who did not participate in other parts of the study. Each sentence context that preceded the target word was presented on a computer screen, one at a time, with the target word and any subsequent context missing. Participants were instructed to type the words that best completed the sentence. No participant chose any of the target words, confirming that the target words were unpredictable from the sentence context The word prior to the target word (wordn-1) was always a five-letter adjective; this was so that wordn-1 would likely be fixated to ensure that the target word would be close enough to the fixation that it could be skipped through parafoveal processing. Each of the 34 target words were embedded into its own sentence structure, which resulted in a total of 34 experimental sentences. There also were 10 filler sentences. Repetition priming lexical decision task The repetition priming lexical decision task (RPLDT) included 100 words. The lexical decision targets were the same as the 34 target words/ primes from the eye-tracking portion of the study. The presentation of the target word in the eye-tracking portion served as the prime condition. Sixteen filler words and 50 orthographically legal and pronounceable nonwords also were included in the RPLDT. A no-prime control condition was created by including participants who did not read the sentences in the eye-tracking portion. Thus, none of the lexical decision targets were primed for this group.
Psychon Bull Rev
Procedure Eye-tracking A bite plate was made for each participant as they read and signed the consent form. Participants rested their mouths on the bite plate during the eye-tracking session to minimize head movements. The eye-tracker was calibrated and aligned for each participant, which took approximately 5 minutes. Before reading each sentence, participants viewed two rows of five boxes on the computer screen. The upper leftmost box was a different color and indicated that the first letter of the sentence would appear there. When participants looked at this box, the experimenter would press a button to display the sentence to control the onset of reading. The study began with a practice session that included five sentences presented one at a time on a computer screen. Participants were instructed to read each sentence silently for comprehension and to press a button when they were done reading. For sentences that were followed by a comprehension question, participants were instructed to press a “true” button if the answer was true or a “false” button if the answer was false. After the practice session, participants continued on to the experimental session that included 34 experimental sentences and 10 filler sentences. Presentation was randomized. Each filler sentence was followed by a true or false comprehension question to ensure that participants read for comprehension. No participant scored lower than an 80 % on the true or false questions (M = 98 % correct). Repetition priming lexical decision task Immediately after completing the eye-tracking portion of the study, participants began the RPLDT. Participants were not told ahead of time that this task would follow the eye-tracking portion of the study. Participants began with a practice session consisting of ten letter strings. A fixation point of a plus sign was presented in the center of the screen for 750 ms, and then the fixation point was replaced by a word or nonword. Participants were instructed to press quickly and accurately a “yes” button if the letter string was a real word or a “no” button if the letter string was not a real word. After the practice session, the experimental session began consisting of the 34 lexical decision targets, 16 filler words, and 50 legal nonwords presented in random order. The control participants only completed the lexical decision task. Apparatus Data from the eye-tracking session were recorded using a Fourward Technologies (2001) Dual-Purkinje Image eye monitoring system Generation 6.1. This system recorded data using binocular vision and measured eye position from the right eye every millisecond. A computer was connected to this system to present the stimuli and store the data. The stimuli were presented one at a time on a VGA color monitor, and each sentences had a maximum of 72 characters. Participants
were seated approximately 32 inches (81 cm) away from the computer screen. One degree of visual angle consisted of four characters. Participants rested a hand on a button box as they read and pushed a button to end the presentation of a sentence and to answer the comprehension questions. The RPLDT was presented using E-Prime.
Results Each analysis was run using a Linear Mixed Effects (LME) Model using the statistics program R (R Project, 2011). Both the participants and items were included in the maximally random structure with random intercepts and random slopes. All p values were calculated using a Markov chain Monte Carlo sampling. Fixed effects were Skipping (skip or fixate), Regression (regression or no regression), or Group (control or experimental) depending on the analysis. Dependent variables included lexical decision reaction times and probability of making a regression into the target region. Variables that were calculated as percentages (such as regressions) were specified as having a binomial distribution and run as a mixed-effects logistic regression. All fixation time measures that were below 120 ms and above 1,000 ms were excluded from analyses. Fixation durations of these lengths are generally caused by eyeblink or track loss and thus not relevant to analyses. Fixations in this range accounted for 3.39 % of the data. The region for the target wordn was defined as the space before the target word until the last letter of the target word. Since the role of word skipping was central to the majority of analyses, it was essential only to include instances where the skipped word was processed on the previous fixation through parafoveal processing. To ensure this, we only included cases in which the target word was skipped when the previous fixation was on wordn-1. Participants launched from wordn-1 before skipping wordn on 75.88 % of all trials. Therefore, 24.22 % of trials were not included in analyses, because we could not ensure that a skipped word achieved any level of lexical access when it was not parafoveally processed. In the RPLDT, all incorrect responses were removed from analyses, accounting for 2.45 % of the data. Means and standard deviations were calculated for each participant after removing any responses below 200 ms and above 3000 ms. Any scores 2.5 standard deviations above or below a participant’s mean were removed, accounting for 10.4 % of the data (Yap, Balota, & Tan 2013). Eye-tracking data The only independent variable was whether the target was skipped or fixated. The target word was skipped 47.14 % of the time, and when the word was fixated
Psychon Bull Rev
the average first fixation duration was 249.91 ms and the average gaze duration was 265.6 ms. A regression into the target word region was defined as the probability of fixating the target word after the word had already been passed either by a skip or fixation. Participants were significantly more likely to make a regression into the target word after it had been skipped (M = 14.43%, SD = 35.19) than after it had been fixated (M = 4.68%, SD = 21.15), β = −0.98, SE = 0.39, z = −2.45, p < 0.05. Repetition priming We found a significant repetition priming effect regardless of whether the prime word was skipped or fixated in the eye-tracking portion of the study.1 Reaction times were significantly faster when the word was skipped during reading (M = 656.42, SD = 173.91 ms) compared with the control group (M = 726.73 ms, SD = 262.21 ms), β = 72.24, SE = 40.17, t = 2.03, p < 0.05. Reaction times were also significantly faster when the word was fixated during reading (M = 614.43 ms, SD = 138.15 ms) compared with the control group (M = 726.73 ms, SD = 262.21 ms), β = 110.87, SE = 37.40, t = 2.96, p < 0.05 (Fig. 1). Importantly, participants responded significantly faster in the lexical decision task when they fixated the target word (prime) (M = 614.43 ms, SD = 138.15 ms) than when they skipped the target word (prime) (M = 656.42, SD = 173.91 ms), β = 37.08, SE = 14.85, t = 2.50, p < 0.05. Because a regression was made to the target word after it was skipped on 14.43 % of all trials in the eye-tracking part of the study, some of the prime words that were in the skipped condition were actually fixated. However, the same effect remained even when we only included trials in which a skipped prime word was never fixated (i.e., no regression to the skipped word) in the eye-tracking part of the study in that participants had significantly faster lexical decision times to words that were fixated in the eye-tracking (prime) portion of the study (M = 614.43 ms, SD = 138.15 ms) than when words were never fixated (M = 661.72, SD = 162.79 ms), β = 40.49, SE = 15.05, t = 2.69, p < 0.05.
Discussion Our findings indicate that skipped words are processed differently than fixated words during silent reading. Evidence for this comes from two main sources: 1) participants took longer
1 The control group and the experimental group did not have different baseline reaction times as revealed by an analysis comparing the reaction times on the filler words for each group, β = 10.23, SE = 52.21, t = 0.20, p = 0.88. The fact that there were no baseline differences allows us to directly attribute any differences in reaction times to exposure to the word (skip, fixate, or no exposure control).
to respond to a word in the lexical decision task when its prime was skipped than when its prime was fixated while reading, and 2) skipped words were more likely to be the target of a regression than fixated words. The first piece of evidence suggests that skipped words do not reach the same level of lexical access as fixated words. The second suggests that skipped words are sometimes misidentified or not fully processed and require a direct fixation via a regression. It appears as though these results are most consistent with the SWIFT Model. SWIFT explains the difference in processing of fixated and skipped words through the assertion that “words need not to be fully identified in order to be skipped” (Engbert et al., 2005, p. 22). This occurs because word skipping is the result of competition between all words in the processing gradient. If wordn+2 is further from reaching lexical access than wordn+1, wordn+2 will be the target of the saccade. Therefore wordn+1 will be skipped without reaching full lexical access. To explain our data, when a word is skipped and is not fully processed it will not prime a future presentation of that word as well as a fully processed fixated word. The second main finding, that skipped words were more likely to be the target of a regression than fixated words, also is consistent with SWIFT. In SWIFT, regressions are caused by incomplete lexical access. If the hypothesis that skipped words do not always reach full lexical access is true, then this pattern of results is to be expected. Computational models as well as experimental data have supported this claim (Engbert, et al., 2005; Vitu & McConkie, 2000). An alternative explanation is that other visual or linguistic aspects of the word can explain the difference in regression probabilities (Vitu, McConkie, & Zola 1998). We find this to be unlikely for our study, because all of the target words in this experiment were controlled for visual and linguistic characteristics (length, frequency, predictability, class). Therefore, incomplete lexical access is a more fitting explanation. Interestingly, a regression was made only 14.43 % of the time when the target word was skipped. This means that either incomplete lexical access occurred rarely or that incomplete lexical access does not always require a regression to be made. Another alternative explanation is that mislocated fixations could account for the higher regression probability to skipped words. A mislocated fixation occurs when the eyes overshoot their intended target due to oculomotor error and is very likely for short words. EZ Reader could easily explain this finding through mislocated fixations. However, much more interesting and novel are the results from the repetition priming lexical decision task. The conclusion that skipped and fixated words are processed differently is inconsistent with the skipping mechanisms in EZ Reader, which is described as requiring “complete identification of the [skipped] word in the parafovea” (Drieghe et al., 2005, p. 964). Occasionally words are skipped, because they are guessed based on their high predictability (Pollatsek et al., 2006). However, the target words in this
Psychon Bull Rev Fig. 1 Mean reaction times in milliseconds (ms) on the target word during the repetition priming lexical decision task when the word was fixated, skipped, or never read (control)
study were unpredictable from the previous context, so it is unlikely that predictability played a role in skipping.2 We propose a slight alteration to account for the current data. It is possible that completion of L1 is sufficient to identify a skipped word. Although L2 is “imminent” when a saccade is programmed to skip a word, L2 might not always be achieved. For the most part, this is not problematic as reading continues uninterrupted. However, increased regression probabilities to skipped words indicate a problem with word identification, and slower reaction times in the RPLDT indicate less processing of skipped words. It is likely that L2 often is completed on skipped words but that the effect is especially pronounced in the current data. The time to reach L2 from L1 is a linear function of the frequency and predictability of the currently attended word (Reichle et al., 2003). The current study used low-frequency and unpredictable words, which maximized the time to reach L2 from L1. As the time it takes to reach L2 increases, the probability of reaching L2 decreases. This interpretation can explain how low-frequency unpredictable skipped words did not prime as well as low-frequency unpredictable fixated words in the repetition priming lexical decision task. This account is consistent with recent work by Choi and Gordon (2013) who attribute word skipping to an implicit lexical decision. This means that words are skipped based on being processed enough that they can be recognized as a real word that fits the overall representation of the sentence. This is
2 Our cloze task norms revealed that no words (not just our target words) were predictable in the sentence contexts, with the highest cloze task probability at 15% for any one response.
sufficient to activate L1 and initiate a saccade program to skip, but it still leaves open the question whether L2 is ever achieved. Our results are also consistent with an account of shallow processing during reading such that although eye movement behavior appears to be normal, not all words are being processed to the same degree (Daneman, Lennertz, & Hannon 2007). A counter argument to the notion that skipped words sometimes do not reach L2 is that attention cannot move forward to wordn+2 until L2 has been achieved on wordn+1. While this is certainly true for fixated words (Henderson & Ferreira, 1990), we argue that this is not necessary for skipped words. If attention were to move forward to wordn+2 when wordn+1 was skipped, then we should expect to see preview benefit on wordn+2. However, there is no evidence that preview benefit is obtained on wordn+2 (Rayner, Juhasz, & Brown 2007). Therefore, it is not problematic for the EZ Reader Model to assume that L2 might not always be achieved on skipped words. While this study has provided evidence that skipped words and fixated word are processed differently, it leaves an open question as to what exactly it means to have identified a word without achieving full lexical access. Interestingly, SWIFT does not provide details about what it means to have incomplete lexical access, and Pollatsek et al. (2006) acknowledge that the conception of lexical access in EZ Reader is an “oversimplification.” These important models have been successful at accurately accounting for much eye movement behavior during reading, but we agree with Pollatsek et al. (2006) that future research should investigate the nature of lexical access in reading, especially as it relates to fixated and skipped words.
Psychon Bull Rev
References Baayen, R. H., Piepenbrock, R., & van Run, H. (1995). The CELEX lexical data base, Release 2 on [CD-ROM]. University of Pennsylvania, Philadelphia: Linguistic Data Consortium. Choi, W., & Gordon, P. C. (2013). Coordination of word recognition and oculomotor control during reading: The role of implicit lexical decisions. Journal of Experimental Psychology: Human Perception and Performance, 39(4), 1032–1046. Daneman, M., Lennertz, T., & Hannon, B. (2007). Shallow semantic processing of text: Evidence from eye movements. Language and Cognitive Processes, 22(1), 83–105. Drieghe, D., Rayner, K., & Pollatsek, A. (2005). Eye movements and word skipping during reading revisited. Journal of Experimental Psychology: Human Perception and Performance, 31(5), 954–969. Engbert, R., Nuthmann, A., Richter, E. M., & Kliegl, R. (2005). SWIFT: a dynamical model of saccade generation during reading. Psychological Review, 112(4), 777–813. Henderson, J. M., & Ferreira, F. (1990). Effects of foveal processing difficulty on the perceptual span in reading: implications for attention and eye movement control. Journal of Experimental Psychology: Learning, Memory, and Cognition, 16(3), 417–429. Pollatsek, A., Reichle, E. D., & Rayner, K. (2006). Tests of the EZ Reader model: Exploring the interface between cognition and eyemovement control. Cognitive Psychology, 52(1), 1–56.
Rayner, K. (1998). Eye movements in reading and information processing: 20 years of research. Psychological Bulletin, 124(3), 372–422. Rayner, K., Binder, K. S., Ashby, J., & Pollatsek, A. (2001). Eye movement control in reading: Word predictability has little influence on initial landing positions in words. Vision Research, 41(7), 943–954. Rayner, K., Juhasz, B. J., & Brown, S. J. (2007). Do readers obtain preview benefit from word n + 2? A test of serial attention shift versus distributed lexical processing models of eye movement control in reading. Journal of Experimental Psychology: Human Perception and Performance, 33(1), 230–245. Reichle, E. D., Rayner, K., & Pollatsek, A. (2003). The EZ Reader model of eye-movement control in reading: Comparisons to other models. Behavioral and Brain Sciences, 26(04), 445–476. Scarborough, D. L., Cortese, C., & Scarborough, H. S. (1977). Frequency and repetition effects in lexical memory. Journal of Experimental Psychology: Human Perception and Performance, 3(1), 1–17. Vitu, F., & McConkie, G. W. (2000). Regressive saccades and word perception in adult reading. Reading as a Perceptual Process, 301–326. Vitu, F., McConkie, G. W., & Zola, D. (1998). About regressive saccades in reading and their relation to word identification. Eye Guidance in Reading and Scene Perception, 101–124. Yap, M. J., Balota, D. A., & Tan, S. E. (2013). Additive and interactive effects in semantic priming: Isolating lexical and decision processes in the lexical decision task. Journal of Experimental Psychology: Learning Memory and Cognition, 39(1), 140–158.
BRIEF REPORT
Skipped words and fixated words are processed differently during reading Michael A. Eskenazi & Jocelyn R. Folk
# Psychonomic Society, Inc. 2014
Abstract The purpose of this study was to investigate whether words are processed differently when they are fixated during silent reading than when they are skipped. According to a serial processing model of eye movement control (e.g., EZ Reader) skipped words are fully processed (Reichle, Rayner, Pollatsek, Behavioral and Brain Sciences, 26(04):445–476, 2003), whereas in a parallel processing model (e.g., SWIFT) skipped words do not need to be fully processed (Engbert, Nuthmann, Richter, Kliegl, Psychological Review, 112(4):777–813, 2005). Participants read 34 sentences with target words embedded in them while their eye movements were recorded. All target words were three-letter, lowfrequency, and unpredictable nouns. After the reading session, participants completed a repetition priming lexical decision task with the target words from the reading session included as the repetition prime targets, with presentation of those same words during the reading task acting as the prime. When participants skipped a word during the reading session, their reaction times on the lexical decision task were significantly longer (M = 656.42 ms) than when they fixated the word (M = 614.43 ms). This result provides evidence that skipped words are sometimes not processed to the same degree as fixated words during reading. Keywords Reading . Eye-movements . Word recognition During reading, approximately 30 % of words are skipped or never directly fixated (Rayner, 1998). Given this prevalence, word skipping has become an important phenomenon for models of eye-movement control to explain. While skipped
M. A. Eskenazi : J. R. Folk (*) Department of Psychology, Kent State University, Kent, OH 44224, USA e-mail: [email protected]
words reach some level of lexical access, there is considerable debate as to how deeply skipped words are processed. On one extreme, skipped words are processed by a “best guess,” and on the other extreme, skipped words are fully identified in the parafovea (Drieghe, Rayner, & Pollatsek 2005). The purpose of the current study was to determine if there are measurable differences in the way that skipped words are processed compared with fixated words. The two most prominent models of eye-movement control (EZ Reader and SWIFT) have different explanations of how words are skipped, and thus the degree of lexical processing of skipped words has implications for each of these models. In the EZ Reader Model, each word must be fully processed before any lexical processing of the next word can begin (Reichle, Rayner, and Pollatsek 2003; Pollatsek, Reichle, & Rayner 2006). Lexical access is split into two stages: a familiarity check (L1), and the completion of lexical access (L2). The completion of L1 indicates that lexical access is imminent and that it is safe to program a saccade to wordn+1. The completion of L2 initiates a covert shift of attention to wordn+1 to begin lexical processing. Word skipping occurs when lexical access is completed or imminent on wordn+1 during the covert shift of attention while the eyes are still fixated on wordn. Therefore, all words, whether they are fixated or skipped, go through the same stages of lexical access according to the EZ Reader Model. This is an important assumption for the model, because word identification explains how the eyes move forward through the text. Unlike the EZ Reader Model, in the SWIFT model, the default target of the saccade is not always wordn+1 (Engbert, Nuthmann, Richter, & Kliegl 2005). The SWIFT Model is a parallel-processing model in which up to four words can be processed at the same time. Each of these words competes to be the target of the saccade from wordn in a dynamic field of activation where lexical access for each word increases to a certain threshold of lexical access, based on each word’s
Psychon Bull Rev
difficulty. After reaching threshold, processing decreases until lexical activity is zero, similar to completion of L2. The next saccade will be directed towards the word in the dynamic field of activation that is the farthest from reaching zero lexical activity. Therefore, if wordn+2 is farther from reaching lexical completion than wordn+1, wordn+2 will be the target of the saccade, and wordn+1 will be skipped. Although, wordn+1 is usually identified, there is no requirement that its lexical activity reaches zero (Engbert et al., 2005). The assertion in the SWIFT model that skipped words do not need to be fully processed is in direct contrast with the assumption of the EZ Reader model that all words are fully processed whether they are skipped or fixated. There is support for both arguments. Engbert et al. (2005) argue that a higher probability of regressions to skipped words is evidence for incomplete lexical access (Vitu & McConkie, 2000). However, there is also evidence that at least some level of lexical semantic access has been achieved on skipped words, because predictable words are skipped more often than unpredictable words (Rayner, Binder, Ashby, & Pollatsek 2001). The question remains as to whether skipped words are processed to the same degree as fixated words. To answer this question, the current study used a repetitionpriming paradigm with the assumption that a word that had not achieved full lexical access would be a weaker prime than a word that had achieved full lexical access. A target word embedded in a sentence during a silent reading task served as the prime, and a second presentation of that word in a lexical decision task was the lexical decision target. The target word/prime and the lexical decision target were the same word, consistent with a repetitionpriming paradigm. In the repetition-priming effect, a single presentation of a stimulus increases the processing speed of a future presentation of the same stimulus. This is a robust effect that can persist at delays up to 24 hours (Scarborough, Cortese, & Scarborough 1977). We have two different hypotheses according to the two different models. According to the EZ Reader Model, both fixated and skipped words go through the same two-stage (L1 and L2) process to achieve lexical access. Therefore, fixated words and skipped words should act as equally effective primes, such that reaction times will not be different on the lexical decision task when the target word/prime was either fixated or skipped during the reading task. However, according to the SWIFT Model, skipped words do not have to reach full lexical access. Therefore, fixated words should be better primes than skipped words, because fixated words will achieve lexical access, whereas skipped words on average will have a lower level of lexical access. In the RPLDT, the average response time when the target word/prime was skipped should be longer than when the target word/prime was fixated.
Method Participants Twenty-six students from Kent State University participated in this study. All had normal or corrected vision and were native English speakers. None reported any reading disability. An additional 26 participants completed only the lexical decision task, serving as a control group for repetition-priming effects. All participants received course credit for participation in the study. Materials and design Eye-tracking Target words were embedded in sentences that contained a neutral context, and the target word was never in the first three or last three words of the sentence. For example, the target word lab was used in the sentence “Her friend had a fancy lab inside of the new building.” To increase the likelihood that a sufficient proportion of target words would be skipped during the eye-tracking part of the study, all target words were three-letter nouns and low frequency (M = 10.54, SD = 6.97 from CELEX) (Baayen, Piepenbrock, & van Run 1995). To ensure that target words were unpredictable and thus would not be guessed through top-down processing without being fully processed, a cloze task was administered to 100 participants who did not participate in other parts of the study. Each sentence context that preceded the target word was presented on a computer screen, one at a time, with the target word and any subsequent context missing. Participants were instructed to type the words that best completed the sentence. No participant chose any of the target words, confirming that the target words were unpredictable from the sentence context The word prior to the target word (wordn-1) was always a five-letter adjective; this was so that wordn-1 would likely be fixated to ensure that the target word would be close enough to the fixation that it could be skipped through parafoveal processing. Each of the 34 target words were embedded into its own sentence structure, which resulted in a total of 34 experimental sentences. There also were 10 filler sentences. Repetition priming lexical decision task The repetition priming lexical decision task (RPLDT) included 100 words. The lexical decision targets were the same as the 34 target words/ primes from the eye-tracking portion of the study. The presentation of the target word in the eye-tracking portion served as the prime condition. Sixteen filler words and 50 orthographically legal and pronounceable nonwords also were included in the RPLDT. A no-prime control condition was created by including participants who did not read the sentences in the eye-tracking portion. Thus, none of the lexical decision targets were primed for this group.
Psychon Bull Rev
Procedure Eye-tracking A bite plate was made for each participant as they read and signed the consent form. Participants rested their mouths on the bite plate during the eye-tracking session to minimize head movements. The eye-tracker was calibrated and aligned for each participant, which took approximately 5 minutes. Before reading each sentence, participants viewed two rows of five boxes on the computer screen. The upper leftmost box was a different color and indicated that the first letter of the sentence would appear there. When participants looked at this box, the experimenter would press a button to display the sentence to control the onset of reading. The study began with a practice session that included five sentences presented one at a time on a computer screen. Participants were instructed to read each sentence silently for comprehension and to press a button when they were done reading. For sentences that were followed by a comprehension question, participants were instructed to press a “true” button if the answer was true or a “false” button if the answer was false. After the practice session, participants continued on to the experimental session that included 34 experimental sentences and 10 filler sentences. Presentation was randomized. Each filler sentence was followed by a true or false comprehension question to ensure that participants read for comprehension. No participant scored lower than an 80 % on the true or false questions (M = 98 % correct). Repetition priming lexical decision task Immediately after completing the eye-tracking portion of the study, participants began the RPLDT. Participants were not told ahead of time that this task would follow the eye-tracking portion of the study. Participants began with a practice session consisting of ten letter strings. A fixation point of a plus sign was presented in the center of the screen for 750 ms, and then the fixation point was replaced by a word or nonword. Participants were instructed to press quickly and accurately a “yes” button if the letter string was a real word or a “no” button if the letter string was not a real word. After the practice session, the experimental session began consisting of the 34 lexical decision targets, 16 filler words, and 50 legal nonwords presented in random order. The control participants only completed the lexical decision task. Apparatus Data from the eye-tracking session were recorded using a Fourward Technologies (2001) Dual-Purkinje Image eye monitoring system Generation 6.1. This system recorded data using binocular vision and measured eye position from the right eye every millisecond. A computer was connected to this system to present the stimuli and store the data. The stimuli were presented one at a time on a VGA color monitor, and each sentences had a maximum of 72 characters. Participants
were seated approximately 32 inches (81 cm) away from the computer screen. One degree of visual angle consisted of four characters. Participants rested a hand on a button box as they read and pushed a button to end the presentation of a sentence and to answer the comprehension questions. The RPLDT was presented using E-Prime.
Results Each analysis was run using a Linear Mixed Effects (LME) Model using the statistics program R (R Project, 2011). Both the participants and items were included in the maximally random structure with random intercepts and random slopes. All p values were calculated using a Markov chain Monte Carlo sampling. Fixed effects were Skipping (skip or fixate), Regression (regression or no regression), or Group (control or experimental) depending on the analysis. Dependent variables included lexical decision reaction times and probability of making a regression into the target region. Variables that were calculated as percentages (such as regressions) were specified as having a binomial distribution and run as a mixed-effects logistic regression. All fixation time measures that were below 120 ms and above 1,000 ms were excluded from analyses. Fixation durations of these lengths are generally caused by eyeblink or track loss and thus not relevant to analyses. Fixations in this range accounted for 3.39 % of the data. The region for the target wordn was defined as the space before the target word until the last letter of the target word. Since the role of word skipping was central to the majority of analyses, it was essential only to include instances where the skipped word was processed on the previous fixation through parafoveal processing. To ensure this, we only included cases in which the target word was skipped when the previous fixation was on wordn-1. Participants launched from wordn-1 before skipping wordn on 75.88 % of all trials. Therefore, 24.22 % of trials were not included in analyses, because we could not ensure that a skipped word achieved any level of lexical access when it was not parafoveally processed. In the RPLDT, all incorrect responses were removed from analyses, accounting for 2.45 % of the data. Means and standard deviations were calculated for each participant after removing any responses below 200 ms and above 3000 ms. Any scores 2.5 standard deviations above or below a participant’s mean were removed, accounting for 10.4 % of the data (Yap, Balota, & Tan 2013). Eye-tracking data The only independent variable was whether the target was skipped or fixated. The target word was skipped 47.14 % of the time, and when the word was fixated
Psychon Bull Rev
the average first fixation duration was 249.91 ms and the average gaze duration was 265.6 ms. A regression into the target word region was defined as the probability of fixating the target word after the word had already been passed either by a skip or fixation. Participants were significantly more likely to make a regression into the target word after it had been skipped (M = 14.43%, SD = 35.19) than after it had been fixated (M = 4.68%, SD = 21.15), β = −0.98, SE = 0.39, z = −2.45, p < 0.05. Repetition priming We found a significant repetition priming effect regardless of whether the prime word was skipped or fixated in the eye-tracking portion of the study.1 Reaction times were significantly faster when the word was skipped during reading (M = 656.42, SD = 173.91 ms) compared with the control group (M = 726.73 ms, SD = 262.21 ms), β = 72.24, SE = 40.17, t = 2.03, p < 0.05. Reaction times were also significantly faster when the word was fixated during reading (M = 614.43 ms, SD = 138.15 ms) compared with the control group (M = 726.73 ms, SD = 262.21 ms), β = 110.87, SE = 37.40, t = 2.96, p < 0.05 (Fig. 1). Importantly, participants responded significantly faster in the lexical decision task when they fixated the target word (prime) (M = 614.43 ms, SD = 138.15 ms) than when they skipped the target word (prime) (M = 656.42, SD = 173.91 ms), β = 37.08, SE = 14.85, t = 2.50, p < 0.05. Because a regression was made to the target word after it was skipped on 14.43 % of all trials in the eye-tracking part of the study, some of the prime words that were in the skipped condition were actually fixated. However, the same effect remained even when we only included trials in which a skipped prime word was never fixated (i.e., no regression to the skipped word) in the eye-tracking part of the study in that participants had significantly faster lexical decision times to words that were fixated in the eye-tracking (prime) portion of the study (M = 614.43 ms, SD = 138.15 ms) than when words were never fixated (M = 661.72, SD = 162.79 ms), β = 40.49, SE = 15.05, t = 2.69, p < 0.05.
Discussion Our findings indicate that skipped words are processed differently than fixated words during silent reading. Evidence for this comes from two main sources: 1) participants took longer
1 The control group and the experimental group did not have different baseline reaction times as revealed by an analysis comparing the reaction times on the filler words for each group, β = 10.23, SE = 52.21, t = 0.20, p = 0.88. The fact that there were no baseline differences allows us to directly attribute any differences in reaction times to exposure to the word (skip, fixate, or no exposure control).
to respond to a word in the lexical decision task when its prime was skipped than when its prime was fixated while reading, and 2) skipped words were more likely to be the target of a regression than fixated words. The first piece of evidence suggests that skipped words do not reach the same level of lexical access as fixated words. The second suggests that skipped words are sometimes misidentified or not fully processed and require a direct fixation via a regression. It appears as though these results are most consistent with the SWIFT Model. SWIFT explains the difference in processing of fixated and skipped words through the assertion that “words need not to be fully identified in order to be skipped” (Engbert et al., 2005, p. 22). This occurs because word skipping is the result of competition between all words in the processing gradient. If wordn+2 is further from reaching lexical access than wordn+1, wordn+2 will be the target of the saccade. Therefore wordn+1 will be skipped without reaching full lexical access. To explain our data, when a word is skipped and is not fully processed it will not prime a future presentation of that word as well as a fully processed fixated word. The second main finding, that skipped words were more likely to be the target of a regression than fixated words, also is consistent with SWIFT. In SWIFT, regressions are caused by incomplete lexical access. If the hypothesis that skipped words do not always reach full lexical access is true, then this pattern of results is to be expected. Computational models as well as experimental data have supported this claim (Engbert, et al., 2005; Vitu & McConkie, 2000). An alternative explanation is that other visual or linguistic aspects of the word can explain the difference in regression probabilities (Vitu, McConkie, & Zola 1998). We find this to be unlikely for our study, because all of the target words in this experiment were controlled for visual and linguistic characteristics (length, frequency, predictability, class). Therefore, incomplete lexical access is a more fitting explanation. Interestingly, a regression was made only 14.43 % of the time when the target word was skipped. This means that either incomplete lexical access occurred rarely or that incomplete lexical access does not always require a regression to be made. Another alternative explanation is that mislocated fixations could account for the higher regression probability to skipped words. A mislocated fixation occurs when the eyes overshoot their intended target due to oculomotor error and is very likely for short words. EZ Reader could easily explain this finding through mislocated fixations. However, much more interesting and novel are the results from the repetition priming lexical decision task. The conclusion that skipped and fixated words are processed differently is inconsistent with the skipping mechanisms in EZ Reader, which is described as requiring “complete identification of the [skipped] word in the parafovea” (Drieghe et al., 2005, p. 964). Occasionally words are skipped, because they are guessed based on their high predictability (Pollatsek et al., 2006). However, the target words in this
Psychon Bull Rev Fig. 1 Mean reaction times in milliseconds (ms) on the target word during the repetition priming lexical decision task when the word was fixated, skipped, or never read (control)
study were unpredictable from the previous context, so it is unlikely that predictability played a role in skipping.2 We propose a slight alteration to account for the current data. It is possible that completion of L1 is sufficient to identify a skipped word. Although L2 is “imminent” when a saccade is programmed to skip a word, L2 might not always be achieved. For the most part, this is not problematic as reading continues uninterrupted. However, increased regression probabilities to skipped words indicate a problem with word identification, and slower reaction times in the RPLDT indicate less processing of skipped words. It is likely that L2 often is completed on skipped words but that the effect is especially pronounced in the current data. The time to reach L2 from L1 is a linear function of the frequency and predictability of the currently attended word (Reichle et al., 2003). The current study used low-frequency and unpredictable words, which maximized the time to reach L2 from L1. As the time it takes to reach L2 increases, the probability of reaching L2 decreases. This interpretation can explain how low-frequency unpredictable skipped words did not prime as well as low-frequency unpredictable fixated words in the repetition priming lexical decision task. This account is consistent with recent work by Choi and Gordon (2013) who attribute word skipping to an implicit lexical decision. This means that words are skipped based on being processed enough that they can be recognized as a real word that fits the overall representation of the sentence. This is
2 Our cloze task norms revealed that no words (not just our target words) were predictable in the sentence contexts, with the highest cloze task probability at 15% for any one response.
sufficient to activate L1 and initiate a saccade program to skip, but it still leaves open the question whether L2 is ever achieved. Our results are also consistent with an account of shallow processing during reading such that although eye movement behavior appears to be normal, not all words are being processed to the same degree (Daneman, Lennertz, & Hannon 2007). A counter argument to the notion that skipped words sometimes do not reach L2 is that attention cannot move forward to wordn+2 until L2 has been achieved on wordn+1. While this is certainly true for fixated words (Henderson & Ferreira, 1990), we argue that this is not necessary for skipped words. If attention were to move forward to wordn+2 when wordn+1 was skipped, then we should expect to see preview benefit on wordn+2. However, there is no evidence that preview benefit is obtained on wordn+2 (Rayner, Juhasz, & Brown 2007). Therefore, it is not problematic for the EZ Reader Model to assume that L2 might not always be achieved on skipped words. While this study has provided evidence that skipped words and fixated word are processed differently, it leaves an open question as to what exactly it means to have identified a word without achieving full lexical access. Interestingly, SWIFT does not provide details about what it means to have incomplete lexical access, and Pollatsek et al. (2006) acknowledge that the conception of lexical access in EZ Reader is an “oversimplification.” These important models have been successful at accurately accounting for much eye movement behavior during reading, but we agree with Pollatsek et al. (2006) that future research should investigate the nature of lexical access in reading, especially as it relates to fixated and skipped words.
Psychon Bull Rev
References Baayen, R. H., Piepenbrock, R., & van Run, H. (1995). The CELEX lexical data base, Release 2 on [CD-ROM]. University of Pennsylvania, Philadelphia: Linguistic Data Consortium. Choi, W., & Gordon, P. C. (2013). Coordination of word recognition and oculomotor control during reading: The role of implicit lexical decisions. Journal of Experimental Psychology: Human Perception and Performance, 39(4), 1032–1046. Daneman, M., Lennertz, T., & Hannon, B. (2007). Shallow semantic processing of text: Evidence from eye movements. Language and Cognitive Processes, 22(1), 83–105. Drieghe, D., Rayner, K., & Pollatsek, A. (2005). Eye movements and word skipping during reading revisited. Journal of Experimental Psychology: Human Perception and Performance, 31(5), 954–969. Engbert, R., Nuthmann, A., Richter, E. M., & Kliegl, R. (2005). SWIFT: a dynamical model of saccade generation during reading. Psychological Review, 112(4), 777–813. Henderson, J. M., & Ferreira, F. (1990). Effects of foveal processing difficulty on the perceptual span in reading: implications for attention and eye movement control. Journal of Experimental Psychology: Learning, Memory, and Cognition, 16(3), 417–429. Pollatsek, A., Reichle, E. D., & Rayner, K. (2006). Tests of the EZ Reader model: Exploring the interface between cognition and eyemovement control. Cognitive Psychology, 52(1), 1–56.
Rayner, K. (1998). Eye movements in reading and information processing: 20 years of research. Psychological Bulletin, 124(3), 372–422. Rayner, K., Binder, K. S., Ashby, J., & Pollatsek, A. (2001). Eye movement control in reading: Word predictability has little influence on initial landing positions in words. Vision Research, 41(7), 943–954. Rayner, K., Juhasz, B. J., & Brown, S. J. (2007). Do readers obtain preview benefit from word n + 2? A test of serial attention shift versus distributed lexical processing models of eye movement control in reading. Journal of Experimental Psychology: Human Perception and Performance, 33(1), 230–245. Reichle, E. D., Rayner, K., & Pollatsek, A. (2003). The EZ Reader model of eye-movement control in reading: Comparisons to other models. Behavioral and Brain Sciences, 26(04), 445–476. Scarborough, D. L., Cortese, C., & Scarborough, H. S. (1977). Frequency and repetition effects in lexical memory. Journal of Experimental Psychology: Human Perception and Performance, 3(1), 1–17. Vitu, F., & McConkie, G. W. (2000). Regressive saccades and word perception in adult reading. Reading as a Perceptual Process, 301–326. Vitu, F., McConkie, G. W., & Zola, D. (1998). About regressive saccades in reading and their relation to word identification. Eye Guidance in Reading and Scene Perception, 101–124. Yap, M. J., Balota, D. A., & Tan, S. E. (2013). Additive and interactive effects in semantic priming: Isolating lexical and decision processes in the lexical decision task. Journal of Experimental Psychology: Learning Memory and Cognition, 39(1), 140–158.
