Reading disappearing text: Investigating the removal of parafoveal preview
and supporting the linguistic control of eye movements.
Two experiments were carried out in order to examine the effects of disappearing text conditions on normal reading processes. This study was based on previous investigations (Liversedge, Rayner, White, Vergilino-Perez, Findlay & Kentridge, 2003, under revision; Rayner, Liversedge, White & Vergilino-Perez, 2003, in press) and aimed to further the understanding of the findings. The first experiment was designed to eliminate the possibility of the involvement of iconic memory in the relatively undisturbed reading of text shown in the previous experiments. The second experiment altered the initial conditions so that the word to the right of fixation disappeared as well as the foveal word after 60ms fixation as subjects read normal sentences. Eye movements were monitored and analysed in terms of global reading measures as well as local effects on a frequency and a length manipulation. Results were used to draw inferences related to models of eye movements in reading concerning parafoveal preprocessing.
These current experiments aim to further the understanding of the processes involved in reading disappearing text that was first investigated by Liversedge, Rayner, White, Vergilino-Perez, Findlay & Kentridge (2003, under revision) and Rayner, Liversedge, White & Vergilino-Perez (2003, in press). These previous experiments presented subjects with sentences in which the word that was being fixated disappeared after 60 ms of fixation. Single words disappeared and then reappeared after a saccade was made to a different word. The present investigation used the same sentences and paradigm with slight variations of the conditions. In the experiment by Liversedge et al. the aim was to determine if a gap effect (Saslow, 1967; Reuter-Lorenz, Hughes & Fendrich, 1991) is present in reading. If such an effect were active, it was suggested that reading would be facilitated by the removal of words after 60ms. This time allowance was chosen because previous experiments have shown that readers can access sufficient information in order to normally process a word within 60ms before a mask appears (Slowiaczek & Rayner, 1987; Rayner, Inhoff, Morrison, Slowiaczek & Bertera, 1981; Ishida & Ikeda, 1989). Such an effect was not supported; however the results did indicate that lexical processes are critical for ‘driving the eyes’ through text rather than oculomotor systems. This conclusion was drawn because word frequency effects were still observed, with readers fixating for longer on an infrequent word even after it had been removed from the screen.
The findings of the experiments were used to support cognitive control models of eye movements such as the E-Z reader model (Reichle, Rayner & Pollatsek, 1999; Reichle, Rayner & Pollatsek, 2003, in press) and the SWIFT model (Engbert, Longtin & Kliegl, 2002). This was due to the fact that observing a gap effect would have indicated that oculomotor factors were more important in guiding the eyes through the text. The gap effect is assumed to occur due to saliency of visual items on the screen. The fact that readers did not read faster, jumping to the next word when the foveal word disappeared, shows that low-level control models, which merely accept visual characteristics of the text to impact on fixation locations and durations (Rayner & McConkie, 1976), can not account for eye-movements in reading. However, it was suggested that the results of these disappearing text experiments could alternatively be explained by iconic memory. The first of the two present experiments aims to eliminate this possibility by introducing a mask instead of the text merely disappearing, but otherwise following the same nature as the study by Liversedge et al (2003, under revision).
Iconic memory is usually investigated in terms of flashed objects rather than text. However, it can be assumed that the principle extends to text reading and there is a ‘pre-categorical sensory memory for visual material’ which is termed iconic memory (Coltheart, 1983). A visual stimulus is assumed to exist psychologically for a short time after its offset, as originally demonstrated by Sperling (1960, as cited in Coltheart, 1980). Stimuli can be identified after very brief presentation, which leads to the proposition that there is a period of time after the display in which the stimulus must still be somehow available for processing. There are a variety of different types of this visual persistence, including neural persistence, visible persistence and informational persistence, the last of which is that commonly referred to as iconic memory, although there has been some confusion of the term with visible persistence (Coltheart, 1980; Loftus & Irwin, 1998). The primary difference appears to be that iconic memory is ‘phenomenologically unavailable’ (Krekelberg, 2001). In other words, an iconic memory trace is not as such ‘visible’ as visible persistence is; it is a retention of ‘visual information’ although the stimulus image does not continue to be experienced.
In the case of reading disappearing text it is suggested that this memory holds an unidentified representation involving the characteristics of the word which can be mentally projected onto the blank screen and therefore processed after the physical disappearance of the text. The traditional theory behind iconic memory holds that it is created within the first few milliseconds of stimulus onset and once it is fully established, the presence of the physical display is irrelevant (Coltheart, 1983). Iconic memory decays rapidly and this trace has to be transferred to a categorical memory for the ‘memory’ to be reported. We can assume that in the case of reading this rapid decay would still allow adequate time for processing to be completed on a word whose physical stimulus has disappeared. It is worth noting that there are variations of the model of iconic memory, with one version being that described above. The other posits that iconic memory serves as a post-lexical or post-categorical temporary record of the episodic characteristics related to a word or other image, for example the colour it was displayed in and where in relation to other stimuli it was presented (Coltheart, 1980). This type of iconic memory would not interfere with the disappearing text paradigm because the store occurs after lexical access.
Importantly, it has been shown that presenting a backward mask degrades or eliminates iconic memory. Although this was shown with single letters rather than words it seems sensible to assume the same will apply (Sperling, 1963, as cited in Coltheart, 1983). Therefore, the first present experiment displayed a mask where in the previous experiments there would have been a blank space to eliminate the possibility of iconic memory being used. This type of manipulation would disrupt both visible persistence and iconic memory, as it is assumed that the presence of the new visual stimulus, although only consisting of X’s, will overwrite information for the previously shown word. It is predicted that if it is truly linguistic processing of the text, after all visual information has been accessed, that guides the eyes along text in reading, there will still be no disruption to sentence reading times and frequency effects will still be observed, preserving the findings of Liversedge et al (2003, under revision) and Rayner et al (2003, in press).
The second of the two present experiments was interested in what effects would be observed if the word to the right of the fixated word also disappeared. Extending the investigations of Liversedge et al (2003, under revision) and Rayner et al (2003, in press), which found very little disruption to reading with one word disappearing, this second experiment aimed to see if disruption was observed if more text was removed after 60ms. The same sentences were used and under the same conditions subjects experienced the removal of two words; which then reappeared when subjects moved their eyes from the fixated word. Therefore the word to the right of fixation did reappear when a saccade was made into the area that this second word used to occupy. Findings from this manipulation were expected to confer insight into parafoveal preview or parallel processing and to suggest how much subjects relied on processing parafoveal words to the right when tested on the previous one-word disappearing study. The issues of how much information is extracted from the right of fixation and whether readers can process information from more than one word at a time are often linked questions and have recently been reviewed extensively by Starr and Rayner (2001). Rayner, Well, Pollatsek & Bertera (1982) investigated the availability of information to the right of fixation using window techniques and showed that even just preserving three letters to the right of a fixated word improved reading, obviously indicating the importance of some sort of information from the right of fixation to reading processes.
There is particular division in eye-movement research on the subject of specifically what parafoveal information can be accessed (Rayner, 1998), with ‘conflicting evidence as to whether readers can obtain useful linguistic information from the next word at a time when an interword saccade is not yet committed to action’ (Inhoff, Radach, Starr & Greenberg, 2000). It has been under debate as to whether semantic preprocessing occurs, with meaning of the parafoveal word influencing saccadic decisions such as where fixations are placed within the next word (e.g. Inhoff & Rayner, 1980; Hyönä, Niemi & Underwood, 1989; Rayner & Morris, 1992). However, it has generally been accepted that only sub-lexical properties such as orthographic familiarity are capable of influencing the ‘where decision’, landing positions within words; but there is still argument concerning whether lexical properties can affect the ‘when decision’, fixation durations (Starr & Rayner, 2001). This controversy arises from the assumption of sequential attention shift models such as the E-Z reader model (Reichle, Rayner & Pollatsek, 2003, in press) that linguistic information cannot be accessed from parafoveal words until after programming of the saccade has been completed.
According to the E-Z reader model’s account of attention allocation in reading, attention shifts occur serially, along a line of text. The end of the first stage of word identification (L1, previously called a ‘familiarity check’, Reichle, Rayner & Pollatsek, 1999), which is a collection of orthographic information but without full access to semantics, and therefore not constituting word identification, triggers the programming of the next saccade. Based on word length information, this programming takes time, having a labile stage during which it can be altered to skip the following word if it turns out that this word is short and has been identified parafoveally. The culmination of the second stage of word identification (L2, previously thought of as ‘lexical access’, Reichle et al, 1999) triggers the shift of attention to the next word. Due to the amount of time taken to programme the saccade, the attention shift usually reaches the parafoveal word before the eye movement is completed. This means that processing difficulty effects of the foveal word on parafoveal preview occur because if the foveal word is easy to access, more time before the saccade is executed is allocated to the parafoveal word. An infrequent word, for example, would take longer in L2 processing than a frequent one and so there is less time to preview the word to the right.
This mechanism overall, suggests that attention is directed only at the foveal word until after it has been identified. The difficulty of the parafoveal word should not affect foveal processing times because attention only shifts to it after the saccade has already been programmed and this will take a set amount of time. In opposition to this model are proposers of parallel processing. Inhoff et al (2000), for example, argue that ‘all words within the range of effective vision are attended and subjected to lexical analyses’. They suggest a ‘gradient shift model’ in which resources are spread over this region according to processing difficulty and visual resolution. This still holds that the foveally fixated word receives most attention, but alters the sequential shift models by suggesting that saccades are initiated towards the point of current attention rather than away from. This is because they are only programmed when the allocation of attention in one fixated region has changed to be directed more towards the right (or left). This explains processing difficulty effects by saying that fewer resources are allocated to and therefore less information is obtained from the parafovea when the foveal word is less frequent or otherwise hard to process.
This idea of parallel processing, supported by evidence of parafoveal-on-foveal effects, is maintained largely by the work of Kennedy on ‘parafoveal cross-talk’ (Kennedy, 2000 a; 2000 b; Kennedy et al, 2002). Kennedy (2000 b) argues that the principal reason for assuming sequential shifts of attention between word units is the ‘seductive but fallacious’ belief that linguistic processes on spoken and written language must work similarly. In Kennedy (2000 a, also reviewed in Kennedy, 2000 b), the claim is made that ‘inspection of a foveal stimulus is generally shorter, and refixation rate lower, when the next word is long rather than short’ and that there are ‘effects of parafoveal word frequency on foveal inspection time’ when the eyes are particularly close to the parafoveal word. However, these results were observed in paradigms which were not close to natural reading, presenting words that did not form sentences and using laboratory tasks which did not suggest normal reading processes. Specifically, sentential context has been shown to influence fixations (e.g. Ehrlich & Rayner, 1981; Rayner, 1998), and this cannot be examined in tasks like those used by Kennedy.
An extension to this work in Kennedy et al (2002) acknowledged discrepancies in work supporting parallel processing of foveal and parafoveal words, pointing out that directions of effects often differ, for example between the studies by Inhoff et al (2000) and Kennedy (2000 a). This 2002 repeat study of Kennedy (2000 a) manipulated length and frequency for both foveal and parafoveal words and constraint of initial three letters for foveal words, examining five-word sequences compared to the previous three-word sequences. The results in this case were used to advocate a process monitoring type model. This still involves parallel processing to some extent but Kennedy et al claim that it explains the complicated patterns observed in this experiment better than a full parallel processing model. It still remains that other studies have found reverse effects and this experiment again involved processes not similar to natural reading. In particular the model by Engbert, Longtin & Kliegl (2002) is favoured by the authors.
This model claims to be based loosely on the gradient model of Inhoff et al (2000) and a saccade generation model by Clark (1999) which suggest saccade targets are not exactly programmed but instead are ‘continuously defined by the pattern of activity’ across an attentional window. The SWIFT model (saccade-generation with inhibition by foveal targets, Engbert et al, 2002) proposes three basic principles. The first is that ‘lexical information processing is spatially distributed over an attentional window’, similar to a parallel model; the second is that ‘saccade timing is separated from saccade target selection’, altering sequential shift concepts slightly and the third is that ‘saccade generation is an autonomous random process with inhibition by foveal targets’. However, this model is still a model of cognitive control, with the state of lexical processing determining movement of the eyes. The emphasis is on the dynamic nature of the model whereby the decisions about when and where to move the eyes changes over time, according to previous processing. Parafoveal information can be accessed in a similar way to a parallel processing model; with each word having a lexical activity value which associates how much processing is done on it during a fixation. A two-stage identification process, partly involving pre-processing, similar to the E-Z reader model, is responsible for increasing and decreasing lexical activity, depending on a word’s difficulty, related to length and frequency. This value then determines if a word is refixated or skipped.
In terms of disappearing text, it seems that if parafoveal information were unavailable after 60ms for this model, the system would have severe problems. Lexical activities would be suddenly and vastly changed by text disappearing and in the two-word disappearing text condition there would be no more available information in the attentional window. It is hard to see how this model would then make a new saccade with no access to activation information. As the model assumes random generation of saccades which are only inhibited by high processing loads, the model seems to have difficulty explaining previous disappearing text results where frequency effects were still found when text had disappeared.
With respect to the present experiments, as noted above, the results are anticipated to aid understanding of parafoveal processing by examining how its disruption in the two-word disappearing paradigm affects reading. In terms of the E-Z reader model, the findings of the one-word disappearing experiments could be explained by the fact that processing of the foveal word completes L2 just at the 60ms, roughly, and so when this word disappears reading continues relatively normally, with attention transferring to the parafoveal word until a saccade occurs, which would already have been programmed using low level visual information. Frequency effects would remain because it is stage L1 that is lengthened by frequency and this is affected by how much parafoveal preview is available beforehand. In the current experiment, we would predict on these assumptions that reading would be disrupted because if stage L2 is completed when both words disappear, there is nowhere for the attention to shift to in the time before the saccade would naturally occur, having been programmed. This would mean also that no preview is available if there is no time for the attention shift and so frequency effects would be lessened.
Sixteen undergraduate students from the University of Durham took part in the experiments, with eight experiencing experiment 1, with masked text, and eight experiment 2, with disappearing text. They were all native English speakers with either normal vision or vision corrected by soft contact lenses, aged between 18 and 22.
40 experimental one-line sentences were used which were taken from the previous disappearing text experiments (Rayner et al., 2003, in press; Liversedge et al., 2003, under revision). Duplicating the earlier experiments, there were two versions of each sentence, with half containing a high frequency six letter word and half containing a low frequency six letter word in the same location. Frequencies were controlled using the CELEX English word form corpus (Baayen, Piepenbrock & Gulikers, 1995). A second manipulation meant that half the sentences in each set contained a short (four letter) critical word while the other half replaced this word with a long (ten letter) word which fit the same context. Words following critical words were always the same length, within one character difference. For example:
Jan began the weekly expedition/trip across the remote/craggy mountains this morning.
He liked the creepy dank/mysterious house that the stupid/ uppity teenagers avoided.
John used the little instrument/tool after the heavy/creaky mechanism broke down.
See Appendix A for a full list of sentences.
These manipulations resulted in 2 lists of sentences, each containing 40 sentences. 4 participants were randomly assigned to each list for each of the two experiments. The 40 experimental sentences in the list were presented in two blocks each containing 20 sentences in a fixed random order counterbalanced for high/low frequency and long/short words. One of the two blocks was presented normally and the other either in the disappearing or masked condition. The order of blocks was also counterbalanced, so an equal number of participants saw the normal condition first as saw the disappearing/masked condition first. In the masked or disappearing conditions the word being fixated disappeared or was masked after it had been fixated for 60ms. Refixating the word without moving off it did not result in the word returning; the word reappeared when the eyes saccaded to another word. At the beginning of each block five filler practice questions were presented and comprehension questions were interspersed throughout fillers and trials, with eight in total in each block. Response to these visually presented questions was by a single yes/no button press.
All sentences were displayed as white lower case text on a black background on a Phillips 21B582BH, 21 inch monitor (with a P22 phosphor with a decay to zero in less than 1.7 milliseconds) linked with a Phillips Pentium III computer, which was interfaced with a Fourward Technologies Dual-Purkinje Generation 5.5 eye-tracker (with spatial resolution of less than 10 min of arc). Five characters subtended one degree of visual angle. Eye position was sampled every millisecond with only movements of the right eye being recorded, although viewing was binocular. The mask consisted of a line of upper case x’s that replaced the screen area which the single word had previously covered.
Participants were seated one metre from the screen in an adjustable-height chair and were given printed instructions directing them to read the sentences to the best of their ability and press a button once they had understood each sentence. A bite bar and head restraint were employed to restrict head movements and then a calibration procedure was initially carried out before the sentence presentation began. Each sentence was presented followed by a calibration accuracy check screen then the next trial sentence or a comprehension question appeared. There was a break in between the two blocks of trials with the total procedure lasting approximately forty five minutes.
The data were analysed for global measures as well as local measures related to the critical words. Global measures used were sentence reading time, number of fixations, fixation duration and number of regressions. For both length and frequency critical words we considered first fixation duration, gaze duration, total fixation time and probabilities of skipping or refixating the words. For all measures we carried out two way ANOVA’s or t-tests both for participant variability (F1 and t1) and for item variability (F2 and t2). Following the earlier disappearing text experiments (Liversedge et al, 2003, under revision; Rayner et al, 2003, in press), we eliminated values for first fixation durations, gaze durations and total fixation times which were below 80ms or over 1200ms. Degrees of freedom vary due to instances of skipping by some participants or tracker loss. All participants gave reasonable numbers of correct answers to comprehension questions so no data was eliminated for this reason.
Results Experiment 1:
The mean sentence reading time for masked text (3436ms) was not significantly different from sentence reading time for normal text (3296ms), t1<1; t2(39) = 1.187, p=0.242.. Readers were not slowed in reading when the mask was presented after 60ms. This result matches that found by Liversedge et al. (2003, under revision) concerning disappearing text without a mask. There was also no difference in total number of fixations between masked and normal text, with an average of 13.9 for masked and 14.0 for normal, t1; t2<1. This differs from results discussed in the original disappearing text experiment, as does the related measure of average fixation duration, which also showed no significant difference, with fixations on masked text lasting 245ms and on normal text lasting 239ms (t1(7) = 1.290, p=0.238; t2(39) = 1.419, p=0.164). This indicates that, unlike in the earlier experiment where the text disappeared rather than being masked, subjects tended to still refixate the words after the display had changed as they do in normal reading rather than changing strategy as seemed evident in the disappearing experiment. The final global measure showed that the total number of sentence regressions was not significantly different between masked and normally presented text, t1; t2<1, suggesting that reading was not disturbed by the mask so that subjects had to return to words to complete understanding. Global summary data can be found in Appendix B.
Results from critical words in the sentences indicated that the frequency effect found in the previous disappearing text experiment by Liversedge et al. (2003, under revision) was retained for the masked text in relation to gaze duration and total fixation time. Initially we considered measures for first fixation duration but all effects were insignificant, although an insignificant frequency effect trend was noted (F1(1,7) = 4.10, p=0.083; F2(1,38) = 2.82, p=0.102). On further investigation, looking at gaze duration and total fixation time measures, it was found that low frequency words were fixated for longer than high frequency words despite the presentation of the mask. See table 1 for mean values. A 2 (frequency: high vs. low) X 2 (condition: masked vs. normal) ANOVA on gaze duration means gave a significant effect of frequency, F1(1,7) = 81.52, p<0.001; F2(1,38) = 9.58, p<0.005 but showed no effect of text condition (F1(1,7) = 2.00, p=0.2; F2(1,38) = 3.37, p=0.074) and no significant interaction (F1(1,7) = 1.21, p=0.31; F2<1). Similarly, an identical ANOVA for total fixation time located a significant effect of frequency, F1(1,7) = 11.11, p<0.05; F2(1,39) = 9.95, p<0.005, but no effect of text condition (F1<1; F2(1,39) = 1.04, p=0.31) and no interaction (F1 and F2<1).
Table 1: Mean gaze duration and total fixation times for high and low frequency words in masked and normally presented text conditions.
Total fixation times exclude cases of word skipping where values for total time are zero.
Examining skipping probabilities yielded no significant effects, although there was a marginal effect of mask. Seemingly words may be more likely to be skipped in the normal text than in the masking condition, F1(1,7) = 4.36, p=0.075; F2(1,39) = 5.35, p<0.05. Similarly no significant results were found for probabilities of refixations, however there were very nearly significant effects of frequency (F1(1,7) = 4.80, p=0.065; F2(1,38) = 8.85, p=0.005) and interaction (F1(1,7) = 8.71, p<0.05; F2(1,38) = 3.48, p=0.070). These indicated that maybe subjects were more likely to refixate infrequent words (probability 0.28) than frequent words (0.10) in the normal condition and that this differed from the trend in the masked condition. However this suggestion should be viewed with caution due to the significance values. Overall, as there is no significantly greater probability of refixating for normal text presentation, the difference between first fixation duration and gaze duration must be due to subjects fixating for longer when they do refixate, although the actual occurrence of refixations is no greater.
Other local analyses results were from those carried out on the words critical for length. Similar to the case of frequency, there were no significant effects noted for first fixation duration measures, either for text condition (F1<1; F2(1,32) = 1.85, p=0.184), for length (F1(1,7) = 1.64, p=0.241), or for interaction (F1; F2<1). However there was a significant effect found for length on gaze duration (F1(1,7) = 12.22, p=0.01; F2(1,33) = 13.61, p=0.001), and total fixation time (F1(1,7) = 59.06, p<0.001; F2(1,37) = 20.74, p<0.001), with participants fixating longer on long words than on short words. Mean values are shown in table 2. The 2 (length: long vs short) X 2 (condition: normal vs masked) ANOVA for gaze duration gave no significant effect of text condition (F1<1; F2(1,33) = 1.45, p=0.237) and no reliable interaction (F1(1,7) = 8.66, p<0.05; F2(1,33) = 1.01, p=0.322).
Table 2: Mean gaze duration and total fixation times for long and short words in masked and normally presented text conditions
In the case of the length analyses there were significant effects of length on skipping and refixation probabilities between text conditions. Short words were much more likely to be skipped in both the normal (probability 0.30) and masked (0.37) conditions than long words (0.05 / 0.06), F1(1,7) = 37.21, p<0.001; F2(1,39) = 56.18, p<0.001. There was no significant effect of text condition (F1: F2<1) or interaction (F1; F2<1) on skipping probability. Long words were more likely to be refixated in both the normal (probability 0.29) and masked (0.22) conditions than short words (0.05 / 0.10), F1(1,7) = 5.99, p<0.05; F2(1,39) = 14.87, p=0.001. There was again no significant effect of text condition (F1; F2<1) or interaction (F1(1,7) = 1.21, p=0.307; F2<1).
Discussion Experiment 1:
In summary, the results for Experiment 1 found that readers were not slowed in reading time and there was equally little disruption to reading processes as there was in the earlier, single-word, disappearing text experiment. There was no difference in fixation patterns from normal reading, showing that in this experiment subjects were unlikely to alter their reading strategy, and still refixated on the masked words after the mask was presented. In Starr & Rayner (2001), it is commented that when all text is replaced with z’s, characteristics of eye movements in reading are preserved, so this result is not surprising, if there is a visual stimulus it will be examined even if it does not form words. Other experiments using a masking type paradigm have found that saccades within strings of hashes resemble those for normal word reading (Ducrot & Pynte, 2002).
This result differs slightly from that observed in the single-word disappearing text experiment (Liversedge et al, 2003, under revision), which found decreased numbers of fixations and longer fixation durations when reading disappearing text in comparison to normally presented text. It seems that readers used minimum numbers of fixations in the previous experiments, possibly relying more on parafoveal preview (as will be discussed in relation to Experiment 2). To compensate for this, fixation durations were longer. In the case of the masked text there were no fewer fixations and also no difference in regressions and fixation durations were similarly unaffected. This difference has to be allocated to the fact that refixations on words which had been masked still occurred, while in the previous experiments subjects did not refixate on the blank spaces. This is primarily due to the nature of the mask and would not affect the presence of iconic memory, if anything in fact suggesting even more strongly that iconic memory was not in use; it was blocked by more access of information from the masked region.
Considering the local effects found in Experiment 1, frequency and length were still observed to maintain effects on fixation durations. Low frequency words were fixated for longer than higher frequency words despite the presence of the mask, indicating that linguistic processing continued after the 60ms depending on the difficulty of the word. There was no effect of the masking condition and no interaction so frequency effects remained the same size for masked as for normal text. This reinforces the finding of Liversedge et al that frequency effects continue having an effect after the 60ms presentation and drive eyes through the text.
Effects on skipping and refixation probabilities are more difficult to interpret but none of them reached reliable significance. It did however seem that subjects were more likely to skip words in the normal presentation condition than when the mask was present. It also appeared that subjects were more likely to refixate infrequent words than frequent words, as would be anticipated, but that this was less likely to show a difference in the masking condition. This seems easily explainable by the fact that readers learned that there was no more information to be gained by refixating an infrequent word but refixations still occurred, as seen in the global measures, possibly just to check that the strings were all x’s and did not contain any other characters yielding useful information. However, as previously mentioned, these effects for skipping and refixating were not entirely significant and a repeat experiment should be carried out involving more subjects as the result was not significant across subjects and could be due to difference in one subject’s data only.
Length effects were as strong for masked as for normally presented text conditions, with longer words being fixated for longer than short words and long words being more subject to refixations and short words being skipped more frequently. The mask did not affect length effects whatsoever, while in the previous investigation the disappearing condition produced an interaction whereby length effects on gaze duration were larger in normal presentation. There was also an effect of disappearing text on first fixation duration that was not found in the present experiment concerning masking. First fixations were longer for disappearing text, possibly because subjects anticipated that the word would disappear and so accessed as much information as possible before moving. This may not have been found for masked text again due to the nature of still having a visual stimulus to access after 60ms. These longer fixations occurred while processing of a word continued and in the masked condition it might be accepted that processing on the word could still continue during a refixation on the mask, although this seems to go against assumptions of the E-Z reader model. Long words may ideally need more than one fixation for all information to be attained and integration of information across fixations may be needed for lexical identification. When the word disappears the reader has to use what information was already available to identify the word using context and lower level information such as orthography. The fact that length effects for gaze duration were truncated for disappearing text in the earlier experiment and yet not in the masking experiment is also hard to explain. However, it seems to suggest that somehow the masking paradigm was slightly easier than the text disappearing so that length effects were not at all disturbed. This in any case goes against what would be predicted if iconic memory were being used in the previous experiments and disrupted by the mask. This enhances the evidence that iconic memory was not being used in the earlier experiment by Liversedge et al.
As the results found in this masking paradigm were similar to those found by Liversedge et al, our results appear to show that there was not a different strategy being utilised and iconic memory was not being used in the single-word disappearing study. Subjects reported not being disrupted by the presence of the delayed mask and did not recall having to change strategies. Overall, it appears that indeed 60ms is enough time for foveal word information to be accessed and that eye movements are then driven by linguistic factors.
Results Experiment 2:
The same measures were employed for the two-word disappearing experiment as for the masking experiment, with the same external values being eliminated and with the same problems for degrees of freedom limitation due to tracker loss and missing data.
Total sentence reading time did differ significantly between the disappearing and the normal text presentation conditions. Mean sentence reading time for normal text was 3208ms while for disappearing text it was 4361ms, t1(1,7) = 6.27, p<0.001; t2(1,39) = 10.94, p<0.001. This indicates that reading was disrupted by the two words disappearing, causing longer overall reading times for the stimuli. Investigating this further, it was shown that the total number of fixations in a sentence also differed as a function of text presentation condition. There were a greater mean number of fixations for disappearing (17.3) than normal text (13.0) which was significant: t1(1,7) = 9.19, p<0.001; t2(1,39) = 11.37, p<0.001. Similarly the number of regressions was also much higher for disappearing text presentation, t1(1,7) = 10.81, p<0.001; t2(1,39) = 18.40, p<0.001, although average fixation durations across the sentence showed no reliably significant difference: t1<1; t2(1,39) = 2.766, p<0.01. This final result seems to suggest that subjects were adopting different strategies to each other for fixation durations. However, overall, it appears that two words disappearing together after 60ms causes disruption to reading whereby readers have to make more fixations throughout the sentence and particularly make more regressions. Global summary data can be found in Appendix B.
When initially considering first fixation durations for the words critical for frequency, there appears to be a frequency effect trend with longer first fixations on infrequent words but this turned out to not be reliably significant (F1(1,7) = 4.53, p=0.071; F2(1,32) = 9.73, p<0.005). There was no effect of text presentation condition and no reliably significant interaction. There were also no reliably significant results regarding gaze duration for either text presentation (F1<1; F2(1,32) = 3.63, p=0.066), frequency (F1(1,7) = 3.30, p=0.112; F2(1,32) = 11.24, p<0.005) or interaction (F1<1; F2(1,32) = 4.91, p<0.05). However, there was a significant effect of text presentation condition on total fixation time, F1(1,7) = 6.54, p<0.05); F2(1,39) =18.74, p<0.001. This means that total fixation times were longer for words in the two-word disappearing condition. As this differs from the result for gaze duration we can assume the effect is mainly due to the presence of more regressions to the word in the disappearing condition. There is still no frequency effect or interaction for total time, indicating that reading may be so disrupted that frequency effects no longer occur. However, the frequency trend for total fixation time is so close to reliability (F1(1,7) = 5.41, p=0.053; F2(1,39) = 21.93, p<0.001) that the difference is possibly only due to one subject not following the normal trend. This supposition must be still treated as unreliable, but it does seem that frequency is eventually having an effect at the total fixation time level. Table 3 compares values of first fixation, gaze duration and total fixation time for frequent and infrequent words across text presentation conditions.
Table 3: Mean first fixation duration, gaze duration and total fixation times for frequent and infrequent words in normally presented and two-word disappearing text conditions.
The lack of any effect of frequency carried over into the analysis of probabilities of skipping and refixating, with no significant results being shown. However, there was an effect of text presentation condition on the probability of skipping. It seems that participants were more likely to skip words in the two-word disappearing text condition (0.22 for infrequent words) than in the normal presentation (0.05 for infrequent words), F1(1,7) = 8.36, p<0.05; F2(1,39) = 27.86, p<0.001. This indicates that reading in the disappearing condition was either altered so that the barest minimum of words were fixated or else, as is more likely, participants tended to skip the second word that had disappeared when attempting to saccade out of the blank region. The fact that there was no frequency effect observed indicates that very little preview was obtained in the visible 60ms and the skipping of the second word was regardless of its frequency type. There were no effects on refixation probability for text condition, frequency or interaction (all F1; F2<1). This result contrasts with that found in the earlier disappearing text experiments. From looking at the means, it does seem that infrequent words were more likely to be refixated (probability 0.20) in the normal condition than in the disappearing condition (0.15), but there was no difference in probability of refixating frequent words between conditions (0.16). However as there was no significance, we must assume that the effect of the two words disappearing was to cause refixations to be programmed relatively randomly, independent of word frequency, with the overall results indicating that the disappearing condition did not overall increase or decrease numbers of refixations from the normal condition.
For length analyses, there were no significant effects on first fixation durations, however, there was a significant effect of length on gaze duration (F1(1,7) = 9.16, p<0.05; F2(1,26) = 0.038, p<0.05) while there was no effect of text condition and no interaction. Due to skipping of words on first pass, as can be seen by the reduced degrees of freedom, this analysis did not show the full effect; therefore looking at total fixation times reveals effects of all three measures. There was a significant effect of the two-word disappearing text condition, F1(1,7) = 15.71, p=0.005; F2(1,37) = 20.52, p<0.001, indicating that both long and short words were fixated for longer overall in the disappearing condition than when text was presented normally. There was a significant effect of length, F1(1,7) = 32.66, p=0.001; F2(1,37) = 27.65, p<0.001, indicating that also long words were fixated longer than short words in both text conditions. The interaction was significantly unreliable (F1(1,7) = 5.01, p=0.060; F2(1,37) = 11.91, p=0.001) although the variability of subjects was so close to reliability that it seems possible there was an interaction but this suggestion should only be considered as such. See table 4 for mean gaze durations and total fixation times for the length variable.
Table 4: Mean gaze duration and total fixation times for long and short words in two-word disappearing and normally presented text conditions.
Data on skipping and refixation probabilities showed a significant effect of length for both but no effect of text condition. The critical word was more likely to be skipped if it was short rather than long, F1(1,7) = 17.48, p<0.005; F2(1,39) = 32.77, p<0.001, and more likely to be refixated if it was a long word, F1(1,7) = 14.08, p<0.01; F2(1,39) = 7.69, p=0.01. There was also no interaction between length and text condition. This indicates that the effects on total time were due to greater numbers of regressions back to long words in the two-word disappearing condition, as there was no effect of the disappearing condition on any other variable considered.
Discussion Experiment 2:
To summarize the findings of the second experiment, the fact that two words disappeared after 60ms caused general disruption to sentence reading, increasing sentence reading times and producing greater numbers of fixations across the sentence, particularly regressions. Despite these effects, average fixation durations did not differ. Within this, however, there were effects observed in both frequency and length analyses of total fixation time, showing that words were generally fixated for longer overall when text disappeared. This indicates that larger numbers of regressions on certain words were mainly responsible for the increase in fixation numbers. Length effects were preserved from normal reading, with long words fixated for more time and more likely to be refixated and short words more likely to be skipped. On the other hand, reversely, frequency effects were disturbed. There was no difference in the number of refixations for frequent or infrequent words and frequent words were no more likely to be skipped. The only hint of a frequency effect remaining was seen in total fixation durations, although unreliable, seeming that frequent words may receive less regressions.
A more detailed consideration of the local analyses shows that there was a very close to significance effect of interaction for length measures of total fixation time. There is slightly more of a length effect in two-word disappearing text (difference of 340ms) than in normally presented text (difference of 122ms). This would be assumed to be due to regressions to long words lasting longer in disappearing text than any regressions in normal text. Less information would have been collated from previous fixations in the disappearing condition than in the normal conditions so longer regressions would be required to complete lexical identification. However, this difference was not reliably significant for variability of subjects, indicating that participants may have used different strategies.
The fact that short words were still more likely to be skipped in the two-word disappearing condition could indicate one of two occurrences. Either the 60ms was long enough for parafoveal information to allow identification of the next short word, or alternatively the greater numbers of words skipped was due to saccades overshooting when precise location information was not available for the word to the right of fixation. If saccades were being programmed to a location estimated to be correct for where the second word along was positioned, short words are more likely to have been skipped in error. The comment could be made that according to the E-Z reader model, the saccade would have been programmed when stage L1 was completed. This presumably would have been slightly before or at about the 60ms cut off point, using evidence from Liversedge et al (2003, under revision) and Rayner et al (1981), which would suggest that the saccade target would have already been chosen. However, if the saccade target is determined in terms of character spaces or is aimed at a particularly identified letter shape, this information would be removed.
In a similar way refixation likelihoods were not disturbed by the two-word disappearing paradigm, with long words still being refixated much more. This follows from the findings of the earlier disappearing experiments which found increased chance of refixation on longer words as well. These refixations may have been programmed before the text disappeared and the program could have reached the non-labile stage, according to the E-Z reader model and been executed even though the word was no longer available. This is likely to lead to the observed increase in regressions to long words in the disappearing condition when sufficient information had not been accessed from the word previously.
In the previous experiment by Liversedge et al (2003, under revision), we might assume that reading was relatively unaffected because subjects adapted to the text disappearance by relying more on parafoveal information. This would imply that lexical information can be available from parafoveal words. Removing this information as well as foveal information was expected to disrupt reading. Very little work can be done to consider how long is needed for available parafoveal information to be accessed as well as foveal information. Presumably this time would be longer than 60ms, which has been shown to be minimal time for visual information to be encoded. According to the E-Z reader model, whatever time is left after processing the foveal word, before a saccade is initiated, is allocated to the parafovea. Assuming that a saccade has already been programmed when the text disappears in the one-word disappearing investigation, the remaining time until a saccadic movement occurs would have been spent processing the parafoveal word. In the two-word disappearing manipulation, this would not be able to occur and either the saccade would land in the blank space left by the parafoveal word or a new saccade would be programmed to land in the still-visible region of text.
Frequency effects, as mentioned, were much reduced or removed in two-word disappearing conditions. This must be explained by the lack of parafoveal preview. In the one-word disappearing experiments frequency effects were maintained but the removal of the information to the right of fixation decreases the availability of frequency or processing difficulty information that is available before direct fixation. The E-Z reader model presumes that some type of information is accessed from the word to the right of fixation, sometimes leading to its early identification, after the shift of attention at the end of the L2 stage of lexical identification. It seems that this stage ends at, or just before 60ms generally, allowing parafoveal access in the one-word disappearing condition but not in the two-word disappearing condition. The lack of this information means that extra time spent in the blank region before a previously programmed saccade is not used to gain preview of the next word and so a new fixation is literally the first time that information about that word can enter the processor. Frequency effects would not be as evident if this were the case. There would be an increase in the number of fixations required, possibly, because regressions are needed more than they would be in the one-word disappearing condition. More time would be required overall on direct fixations, which would possibly increase time needed for stages L1 and L2 to over 60ms, after which a saccade and then a regression is needed to regain the foveal word to complete lexical identification. This would have knock-on effects throughout a sentence so that even less preview would be possible. Frequency would cease to have an effect on skipping because this decision is based on preview information.
In these ways, it seems as though the E-Z reader model can explain the effects that were obtained. Because there is no time for an attention shift to occur, parafoveal preview is prevented and this is shown to have a large disadvantage for reading. It seems that the information acquired from the right of fixation is very important for reading processes, as the only difference between this study and the earlier studies by Rayner et al (2003, in press) and Liversedge et al (2003, under revision) is the fact that the word to the right is not available during each fixation after 60ms.
It is worth considering how the parallel processing advocates might attempt to explain these effects. The gradient shift model proposed by Inhoff et al (2000) would claim that all words within ‘effective range of vision’ are attended at once, with resources allocated according to visual resolution (the foveal word) and harder to process, less frequent or longer, words. Therefore, in the two-word disappearing experiment information would be available from the parafovea throughout the whole 60ms, depending on the frequency and length of the foveal word, the parafoveal word would have received some processing and over the 60ms the gradient would change to direct more attention to the word to the right. It would seem that frequency effects should still remain from this kind of system even when both words are removed. The model could also not explain saccades that land in the blank region, for example, when long words were refixated even after the word had disappeared, because saccades are only initiated towards the locus of attention when the gradient has moved to the right. The gradient would presumably alter as soon as the words disappeared and a large saccade would have to be made to find visible information. This would explain an increase in skipping in the two-word disappearing experiment because there seems to be no account for how a saccade could be made into a blank region, so the second of the two disappearing words would be likely to be skipped. However it seems as though this model could not account for the lack of frequency effects that were still observed, as some early parafoveal information should still have been available.
The SWIFT model of Engbert et al (2002) faces similar problems, having been based on the gradient model. Again lexical processing is distributed over an attentional window, with the window having lexical activities associated with words within it. When processing is in progress lexical activity increases; so at the end of the 60ms, we can assume the second word has a high lexical activity. This is because it is likely the first word has already reached the second stage and lexical activity is decreasing for that word. This makes the second word a likely saccade target. However, this model has a similar saccade mechanism to the E-Z reader model, in that it introduces a labile program, but it also posits that during this stage target location can be modified. This is rather than having to program another saccade and cancel an old one. If the saccade target disappears, presumably the target can be modified, but it is difficult to see how a new target would be chosen as no other words would have any activation attributed to them. If, on the other hand, activation can remain present after a word has disappeared, the saccade would still be executed to the space that was occupied by the second word. There would then be preview available for this word and processing of it would presumably continue as usual, this does not seem to suggest reading would be disrupted as much as was observed. If this was the case predicted by this model it is also hard to account for the findings of higher probability of skipping in two-word disappearing text. Because the model relies on random generation of saccades only modified by processing difficulty feedback from foveal words, it also appears that fixations would have to remain on the blank space until a saccade happened to be generated. This would be a similar amount of time for each word, independent of length or frequency, which does not seem to fit the pattern of results.
Having considered these three types of models it seems as though the sequential shift assumptions can best explain the results shown from this two-word disappearing experiment. Our results, to some extent, contradict predictions of parallel processing, as it appeared that very little information about the word to the right of fixation had been accessed within the 60ms in which the word was visible. Our results have reinforced the importance of parafoveal preview, however, having shown that reading is vastly disrupted when it is removed in this way.
What could be said overall about the study of the disappearing text is that it gives subjects a situation that is very unlikely to occur in natural reading and that processes when faced with this strange pattern of text disappearance are likely to behave differently to normal text reading. Indeed, this seems likely and subjects reported finding it very difficult and having to resort to regressing to almost every word in the two-word disappearing condition. However, it was shown in the initial experiments that a one-word disappearing condition allows very normal processes to occur. We can assume that the removal of the text forces the processing system to use its most efficient mechanism to carry out processing to the best of its ability. This can give useful information as to the backup systems that the processor might possess. Even the authors of the E-Z reader model comment that the model is designed to explain reading under normal, most often experienced, conditions and that it does not account for higher level linguistic processing because these effects ‘typically occur when the reader is having difficulty understanding the text’; the model is referred to as regarding the ‘default’ processes only (Reichle et al, 2003, in press).
From Experiment 1 we have provided evidence that iconic memory was not responsible for the effects found in the disappearing text experiments conducted earlier by Rayner et al and Liversedge et al. If anything, it seemed as though the masking paradigm had less detrimental effect than the single word disappearing text condition. This refutes the suggestion that visual characteristics of words were held in a pre-categorical store which enabled continued processing after the disappearance of the stimuli. From the second of the current experiments we merely aimed to further investigate the effects of disappearing text, with specific interest in parafoveal access of information and looking to further support the claim of the earlier investigation that lexical processes drive the eyes sequentially through the text rather than oculomotor factors.
Results enhanced the importance of parafoveal preview in reading and suggested severe disruption due to a limit placed on the time available after a sequential shift of attention, according to the model of Reichle et al (2003, in press). A lack of frequency effects suggests that parafoveal information integrated with later foveal information is important in influencing the effect of frequency, although length effects were relatively spared, presumably as length information is accessed before frequency information. What was not examined in the present experiment, which could have provided more interesting information, was where saccades landed within words that had disappeared in the prior fixation. It would be interesting to note whether the preferred viewing position effects were kept intact. These have been extensively investigated in text reading by Vitu (Vitu, 1991; Vitu, McConkie, Kerr & O’Regan, 2001) in relation to the influence of availability of parafoveal information.
As previously referenced, Rayner et al (1982) showed that reading could continue if only part of a parafoveal word was left available in a masking paradigm, specifically the first three letters, claiming that word integrity does not need to be maintained for preview to have beneficial effects. It would be interesting for further research to investigate whether making the foveal word plus the first three letters of the next word disappear would show exactly the same effects as making two whole words disappear or if the partial information would be sufficient for processing to be less disturbed. The same study also indicated that when processing to the right is disrupted, as it was in the second of the present experiments, more information is used from the left of fixation. Removing the word to the left as well as the right would most probably interrupt the reading process more, however it would be interesting to see if sentences could still be understood in this situation, as they were still comprehensible in the current experiment, merely slowed down and read using different strategies.
Baayen, R.H., Piepenbrock, R. & Gulikers, L. (1995). The CELEX lexical database. (CD-ROM). Philadelphia: Linguistic Data Consortium, University of Pennsylvania.
Clark, J.J. (1999). Spatial attention and latencies of saccadic eye movements. Vision Research, 39, 585-602.
Coltheart, M. (1980). Iconic memory and visible persistence. Perception and Psychophysics, 27, 183-228.
Coltheart, M. (1983). Iconic Memory. Philosophical transactions of the Royal Society of London: Series B – Biological Science, 302, 283-294.
Ducrot, S. & Pynte, J. (2002). What determines the eye’s landing position in words? Perception and Psychophysics, 64, 1130-1144.
Ehrlich, S.F. & Rayner, K. (1981). Contextual effects on word perception and eye movements during reading. Journal of Verbal Learning and Verbal Behaviour, 20, 641-655
Engbert, R., Longtin, A. & Kliegl, R. (2002). A dynamical model of saccade generation in reading based on spatially distributed lexical processing. Vision Research, 42, 621-636.
Hyönä, J., Niemi, P. and Underwood, G. (1989). Reading long words embedded in sentences: informativeness of word halves affects eye movements. Journal of Experimental Psychology: Human Perception and Performance, 15, 142 – 152.
Inhoff, A.W., Radach, R., Starr, M. & Greenberg, S. (2000). Allocation of visuo-spatial attention and saccade programming during reading. In Kennedy, A., Radach, R., Heller, D. & Pynte, J. (Eds.). Reading as a perceptual process (pg 221-246). Elsevier. Oxford, UK.
Inhoff, A.W. & Rayner, K. (1980). Parafoveal word perception: A case against semantic pre-processing. Perception and Psychophysics, 27, 457-464.
Ishida, T. & Ikeda, M. (1989). Temporal properties of information extraction in reading studied by a text-mask replacement technique. Journal of the Optical Society of America: Optics and Image Science, 6, 1624-1632.
Kennedy, A. (2000 a). Parafoveal processing in word recognition. The Quarterly Journal of Experimental Psychology, 53A, 429-455.
Kennedy, A. (2000 b). Attention allocation in reading: sequential or parallel? In Kennedy, A., Radach, R., Heller, D. & Pynte, J. (Eds.). Reading as a perceptual process (pg 221-246). Elsevier. Oxford, UK.
Kennedy, A., Ducrot, S. & Pynte, J. (2002). Parafoveal-on-foveal interactions in word recognition. The Quarterly Journal of Experimental Psychology, 55A, 1307-1337.
Krekelberg, B. (2001). The persistence of position. Vision Research, 41, 529-539.
Liversedge, S.P., Rayner, K., White, S. J., Vergilino-Perez, D., Findlay, J.M. & Kentridge, R.W. (2003, Under revision). Eye movements when reading disappearing text: Is there a gap effect when reading? Vision Research.
Loftus, G.R. & Irwin, D.E. (1998). On the relations among different measures of visible and informational persistence. Cognitive Psychology, 35, 135-199.
Rayner, K. (1998). Eye movements in reading and information processing: 20 years of research. Psychological Bulletin, 124, 372-422.
Rayner, K., Inhoff, A.W., Morrison, R.E., Slowiaczek, M.L. & Bertera, J.H. (1981). Masking of foveal and parafoveal vision during eye fixations in reading. Journal of Experimental Psychology: Human Perception and Performance, 7, 167-179.
Rayner, K., Liversedge, S.P., White, S.J. & Vergilino-Perez, D. (2003, In press). Reading Disappearing Text. Psychological Science.
Rayner, K. & McConkie, G.W. (1976). What guides a reader’s eye movements? Vision Research, 16, 829-837.
Rayner, K. & Morris, R.K. (1992). Eye movement control in reading: Evidence against semantic pre-processing. Journal of Experimental Psychology: Human Perception and Performance, 18, 163-172.
Rayner, K., Well, A.D., Pollatsek, A. & Bertera, J.H. (1982). The availability of useful information to the right of fixation in reading. Perception and Psychophysics, 31, 537-550.
Reichle, E.D., Rayner, K. & Pollatsek, A. (1999). Eye movement control in reading: accounting for initial fixation locations and refixations within the E-Z reader model. Vision Research, 39, 4403-4411.
Reichle, E.D., Rayner, K. & Pollatsek, A. (2003, in press). The E-Z Reader model of Eye Movement Control in reading: Comparisons to other models. Behavioral and Brain Sciences.
Reuter-Lorenz, P.A., Hughes, H.C. & Fendrich, R. (1991). The reduction of saccadic latency by prior offset of the fixation point: An analysis of the gap effect. Perception and Psychophysics, 49, 167-175.
Saslow, M.G. (1967). Effects of components of displacement-step stimuli upon latency for saccadic eye movement. Journal of the Optical Society of America, 57, 1024-1029.
Slowiaczek, M. L. & Rayner, K. (1987). Sequential masking during eye fixations in reading. Bulletin of the Psychonomic Society, 25, 175-178.
Starr, M.S. & Rayner, K. (2001). Eye movements during reading: Some current controversies. Trends In Cognitive Sciences, 5, 156-163.
Vitu, F. (1991). The influence of parafoveal preprocessing and linguistic context on the optimal landing position effect. Perception and Psychophysics, 50, 58-75.
Vitu, F., McConkie, G.W., Kerr, P. & O’Regan, J. K. (2001). Fixation location effects on fixation durations during reading: an inverted optimal viewing position effect. Vision Research, 41, 3513-3533.