LANGUAGE AND SPEECH, 1990,33(1),69 -81
69
EYE FIXATIONS PREDICT READING COMPREHENSION: THE RELATIONSHIPS BETWEEN READING SKILL, READING SPEED, AND VISUAL INSPECTION” GEOFFREYUNDERWOOD, ALISONHUBBARD, and HOWARD WILKINSON Utiiversity of Nottingliani. U.K. This experiment addressed the question of whether reading comprehension and speed could be predicted by eye futations. From a sample of university students who completed tests of reading comprehension and vocabulary, we selected a group of highly skilled readers and a group of less skilled readers. These two groups then read sentences as their eye movements were monitored, with fixation locations and durations recorded. A discriminant function analysis showed that futation duration was a successful predictor of reading comprehension, but that the number of Lxations, regressive fmations, reading speed, and vocabulary were not reliable predictors A multiple regression analysis revealed that reading speed was predicted by the number of fixations, the average fixation duration, and the duration of the final fixation upon the sentence, but there was no relationship with reading ability. Highly skilled readers are those who can extract information efficiently, but are not necessarily those who have fast overall reading rates.
Key words: reading skill, eye movements
INTRODUCTION Is there a relationship between the acquisition of information from printed text and the ability to use that information? A reader’s pattern of eye f i a t i o n s may be described in t e r m of the locations and durations of futations, the number of fixations, and.the number of regressive futations. A reader’s characteristic pattern of information extraction may be represented by any combination o f these measures. Alternatively, individual readers may have no characteristic pattern of eye movements and fiiations which can be related to their use of the information being extracted. The movements of a reader’s eyes may have little in common with the understanding of the written language being f i a t e d , and there is certainly more to reading than moving one’s eyes in a particular way. However, recognising individual word meanings, integrating syntactically related words, and making inferences about the relationships between events
*
We aregrateful to Alan Kennedy, Sandy Pollatsek, and Bruno Repp for their comments on this paper. Joanne Leach, Helen hlacleod, Frances Villain, and Belinda Winder provided valuable assistance during the administration and scoring of the reading and vocabulary tests. Send correspondence t o Geoffrey Underwood, Department of Psychology, University
of Nottingham, Nottingham NG7 2RD, U.K.
Downloaded from las.sagepub.com at CAMBRIDGE UNIV LIBRARY on August 22, 2015
70
Eye Fixations arid Reading Coniprehertsiort
and actors in the text are all time-consuming processes, and the time course of processing might be expected to be indicated by the time taken to extract the information from the display. The length of the terminating (“wrap-~p’~) futation also adds to the total reading time when sentences are displayed sequentially, and especially when the reading task demands a decision to be made on completion of the reading of each sentence. Just and Carpenter (1978) presented their sentences sequentially and found strong evidence for long wrap-up futations, and in reading studies which have presented texts word-by-word there is evidence of longer pauses at the ends of sentences (Aaronson and Scarborough, 1976). It is less clear whether the fmal futation on a sentence differs from other fiations in more natural reading exercises. However, in a study by Just and Carpenter (1980), in which short paragraphs were displayed and eye movements monitored, there was evidence of long wrap-up furations. This study developed a regression model in which wrap-up fiations made a’ reliable 71 msec per futation contribution to the equation for total reading time. When eye movements have been observed, the evidence suggests that there is a relationship between the difficulty of processing and the futation pattern upon the word. In the case of individual words, Rayner and Duffy (1986) found longer fuation durations upon low frequency words (e.g., goitdola) than upon high frequency words (e.g., vehicle), and also found longer futations upon ambiguous nouns with equally likely meanings (e.g., coach) than upon ambiguous nouns with a dominant meaning (e.g., perch). These demonstrations indicate that the difficulty of accessing a word’s meaning is indicated by the duration of the initial futation upon the word. Within a word there may be an uneven distribution of information, with the final letters being more or less useful in its identification. Compare, for example, the first five and last five letters of the two words nioralistic and supervisor. In the first example the beginning of the word is most informative (try deleting the first five letters and having someone guess the word, and then try it again, with the last five letters deleted instead), but for supervisor the ending is more informative. This uneven distribution has an effect upon futation durations upon each half of the word. O’Regan, Le’vy-Schoen, F‘ynte, and Brugalle’re (1984), Underwood, Bloomfield, and Clews (1988), and Hyona, Niemi, and Underwood (1989) found longer futation durations upon informative ends of long words. The number of futations upon word-halves also increased for halves containing more information. Underwood, Clews, and Everatt (1990), in extending this investigation of the effects of information distribution upon fiiation strategies,. have found a trade-off between measures of duration and measures of location. When only one fiation is made upon a long word, then it tends to last longer than when two futations have to be made. These results together suggest a sensitivity of fuation durations to the microprocessing of information within words. Words which would not by themselves be difficult to process can be placed in sentences in such a way as to produce local processing difficulties. The prior context in a sentence can reduce the duration of a fixation if the context and the word are congruous (Ehrlich and Rayner, 1981). In a related experiment, Kennedy (1978) reported that anaphors with high conjoint frequency gain shorter fuations than those
Downloaded from las.sagepub.com at CAMBRIDGE UNIV LIBRARY on August 22, 2015
G. Underwood, A. Hiibbard, arid H. lvilkirisoii
71
preceded by low conjoint associates. The presentation of the word vehicle in one sentence resulted in less visual attention being allocated to the word btrs in the following sentence, in comparison with paragraphs in which vehicle was followed by tank. This could be interpreted as a word priming effect, with a strong associate providing greater facilitation than a weaker associate. Difficult syntactic structures also result in longer f i a t i o n durations than simpler forms (Rayner, Carlson, and Frazier, 1983). These obscrvations of the effects of processing difficulties upon futations were made with homogeneous samples of readers, but a similar case can be made for samples of readers who find text comprehension difficult, in comparison with more skilled readers. Fixation patterns indicate a moment-to-moment sensitivity to comprehension difficulty, and so perhaps we should expect that individuals with less reading skill will also show their comprehension difficulty by producing furation profiles which differ from those of more skilled readers. Although there is some evidence of a decrease in futation duration with increasing age in children (see Rayner, 1978), there is little evidence of a relationship between reading ability and inspection duration when reading experience is controlled (Kennedy and Murray, 1987; hlurray and Kennedy, 1988). In the present study we compared inspection patterns during reading using two groups of adults who varied in their reading abilities. The two groups were highly educated, but differed in their performance on a standard test of reading comprehension. The purpose was to determine whether readers could be differentiated by the measures of eye fixations taken during a sentence comprehension task.
Stibjects
One hundred University of Nottingham students completed a reading comprehension exercise based on the Black (1962) Comprehension Test for College of Education Students, in pencil and paper format. All subjects were native English speakers. The reading test presents short passages from published works (fiction, biography, and newspaper articles), followed by comprehension questions. The passages continue to be available during completion of the questions. The test took 30 minutes and consisted of a number of types of items, most of which offered multi-choice answers (five choices). Some items required the interpretation of words in the context of the passage, other items required words missing from a paragraph to be selected from five alternatives (a “cloze” test), and others required textual interpretation. For example, in one paragraph which described the activities and life style of a primitive tribe, the words feud, vendetta, and warrior appeared in sentences preceding the tested sentence: “The iiiiiiieroiis tribes arid conrbiiiations of tribes all have their accoiiiits to settle with one another, and very few debts are left triipaid’: A test item then asked whether these debts were: (i) unpaid loans, . . . or (v) murders unavenged. Yet other items required assessment of the perspective of the writer. The maximum score on our version of the test was 25, and the subjects ranged from nine t o 22 answers correct. The same 100 students also completed a vocabulary test based on the more difficult items taken from the hlill-Hill
Downloaded from las.sagepub.com at CAMBRIDGE UNIV LIBRARY on August 22, 2015
72
Eye Fivatioris arid Readiiig Conipreheiuioii
Vocabulary Scale (Raven, 1977). These items required that low frequency target words be matched with one of six possible synonyms. The maximum score on our version of this test was 30, and subjects scored between four and 23 items correct. On the basis of the reading comprehension test, the 15 highest scoring and the 15 lowest scoring subjects were invited to participate in the experiment which recorded eye movements during reading. The mean reading scores for the highest scorers was 20.7 (sd = 0.79), and their mean vocabulary score was 13.2 (sd = 3.82). The mean reading score for the lowest scorers was 10.7 (sd = 1.33) and their mean vocabulary score was 10.47 (sd = 3.96). As these subjects were drawn from a university population, and could therefore all be considered to be good readers relative to the population as a whole, we shall refer to the two groups as consisting of highly skilled readers and less skilled readers. Stiriiiihis materials
Three sets of sentences formed the experimental materials, with each set containing 15 sentences which represented a brief story. In addition to these 45 test sentences, six practice sentences were prepared. Each sentence contained between six and 12 words (33-52 character spaces) with an average of 8.5 words per sentence, and was displayed as a single line of text. Each contained a target word near its centre. The target words varied in frequency (Kucera and Francis, 1967). High frequency targets had counts in excess of 150 (e.g., turn, brown, little), medium frequency words were between 30 and 70 (e.g., soft, trend, tuefiil), and low frequency words had counts below 10 (e.g., twin, srreak, uproar). These words varied between four and seven letters in length, with equal numbers of four, five, six and seven letter words in each of the three frequency categories. The target words were organised into triplets which were to be presented in identical contexts, to ensure that variations in eye futations upon the targets could not be attributed to variations in the preceding contexts. Within each triplet was a high, medium, and low frequency word. As an example, consider the following triplet of words in their sentences: High frequency:
There was a little noise from under the bushes.
Medium frequency:
There was a treiid to vote for someone uninvolved.
Low frequency:
There was a screech as the woman hit the brakes.
The target words are highlighted here but were not, of course, in the experimental presentations. Procediire
Each subject read six practice and 45 experimental sentences, with the instruction to read each paragraph for meaning and in preparation for a comprehension question. At the start of the experiment the eye movement tracker was calibrated and the head restraint set for each subject. The 45 sentences were presented in three blocks of 15, with the order of sentences within a block being maintained between subjects, and the order of the blocks counterbalanced. Each trial consisted of a fvtation mark first appearing in the centre of the screen, and
Downloaded from las.sagepub.com at CAMBRIDGE UNIV LIBRARY on August 22, 2015
G. Uitdenvood. A. Hiibbard, and H. IVilkiiisoii
73
then at a position on the screen corresponding t o the first letter of the sentence. The sentence appeared 0.5 seconds later, and the subjects indicated completion of reading by pressing a hand-held button. The sentence was then removed, and the next trial started. Between blocks of trials the equipment was re-calibrated, and between sentences there was an opportunity for an adjustment if necessary. At the end of the practice sentences, and at the end of the complete experiment, comprehension questions were asked about each block of sentences to check that understanding had been achieved. Subjects had no difficulty in answering the questions correctly.
Apparatus The sentences were displayed on a high resolution monitor (Taxan KX-12) controlled by a BBC Master microcomputer. At the viewing distance of 48 cm a ten-letter word subtended a horizontal angle of 3 degrees. hiaximum letter height was 0.6 degrees. The microcomputer also received inputs from the eye-movement recorder which sampled every 10 msec. The recorder detected the horizontal position of the eye by monitoring the reflection of an infra-red beam from the iris-sclera border of the left eye. This recorder has been described elsewhere, and gives an accuracy of plus or minus one character space (Wilkinson, 1976). The infra-red source and photoelectric detector were mounted on a headband, which was adjusted for each subject. This procedure locked the positions of the eye movement recorder and the subject’s head, and head movement was restricted by a chin rest and head frame. Displays were terminated with a button press by the subject. The output from the eye position recorder was interpreted with the aid of a program which converted sampled positions into eye fiations. The fEation detection algorithm was based on a running average technique, and can be summarised as follows. The difference between the first and second values of ‘the recorder’s output was computed and compared with a set predetermined value, called W. If this difference was not in excess of the value of W, an average of the two values was computed. The difference between the third reading from the recorder and this computed average was then compared with W, and a new average of the three readings was computed if the difference was again not in excess of W. This process was repeated until, after a number of readings, the comparison exceeded W whereupon a futation was deemed to have finished and a saccade started. Iteration of this algorithm identified the end of this saccade and hence the start of the next furation. The value of W was determined primarily by the variation of the output of the recorder when a typical subject attempted t o furate a single letter character. Once determined at the start of the investigation the value of W was held const ant.
RESULTS From each sentence a number of measures were taken in preparation for three sets of analyses. The measures were total reading time, total number of fixations, number of regressive fiiations, average fiation duration (excluding the first and last futations),
Downloaded from las.sagepub.com at CAMBRIDGE UNIV LIBRARY on August 22, 2015
74
Eye Fixatioris arid Readiiig Coniprelieiisiori TABLE1 Means for the two groups of readers for the measures taken from each sentence (Standard deviations in brackets) Highly skilled readers
Less skilled readers
Total sentence reading time (sec)
2.85 (1.08)
3.21
(0.76)
Number of futations per sentence
10.58 (3.63)
10.13
(2.16)
1.75 (1.68)
1.95
(1.21)
Number of regressive futations Overall futation.duration (msec) [excluding wrap-up f ~ a t i o n ]
249.1 (20.1)
284.7
Wrap-up futation duration (msec)
326.2 (90.2)
386.0 (144.6)
(33.8)
Fixation duration on target word (msec):
- high frequency words - medium frequency words
226.5 (30.3)
262.8
(46.0)
241.2 (32.6)
254.0
(36.6)
-low frequency words
237.1 (20.3)
288.6
(67.5)
duration of the final (wrap-up) futation, and duration of the fixation upon the target word. The means of these measures. are shown in Table 1. The analyses looked for differences between the two groups of readers on each measure, the discriminant function which best describes the characteristics of the two groups, and the regression function which best describes the time taken to read a sentence.
Differeiices betweeii the two groups of readers The measures taken from each sentence were averaged for each subject, and submitted to separate analyses of variance. Reading time per sentence did not differ reliably between the highly skilled and less skilled readers [F (1,28) = 1.113, neither did the number of futations per sentence [F (1,28) < 11, and neither did the number of regressive futations per sentence [F (1,28) < I]. Mean futation duration did vary between the groups [F (1,28) = 12.26, p < 0.0021, with highly skilled readers futating for approximately 35 msec less than the less skilled readers. The more specific measure of the final wrap-up futation duration did not differ reliably between groups [F (1,28) = 1.871. The final analysis concerned the futation durations upon target words, for the first futation only, and yielded a significant main effect of reading skill [F ( I , 27) = 6.77, p < 0.021, a marginal effect of frequency [F (2,54)= 3.10, p < 0.061 and a marginal interaction [F (2,54) = 3.08, p < 0.061. Analyses of simple main effects showed differences between the two groups
Downloaded from las.sagepub.com at CAMBRIDGE UNIV LIBRARY on August 22, 2015
G. Uttdenvood, A . Hiibbard, arid H. IViIkitisoii
75
of readers for high frequency words [F (1,27) = 6.08, p < 0.031 and for low frequency words [F (1,27) = 7.44, p < 0.021. There was no difference between groups for the durations of futations on medium frequency words [F (1,27) < I].
Discrimiiiatitig betweeit grotips of readers To describe the differences that best differentiate between highly skilled and less skilled readers, a discriminant function analysis was performed. Tatsuoka (1971) recommended the use of this procedure as a follow-up to analysis of variance, to determine the best linear combination of variables which will maximally differentiate the groups. The variables entered into the discriminant function analysis were those listed in Table 1 together with vocabulary score. This stepwise analysis first determined Wilks’ lambda ratio’ between groups for each discriminating variable, and entered the variable with the smallest ratio into the model. After entering this variable, the remaining variables were inspected, and the variable with the smallest ratio then entered. The process was continued until no remaining variable had an acceptable F-ratio associated with its Wilks’ lambda value. At the end of the process there were three variables in the model. These were, in order of entry, mean f i a t i o n duration (A = 0.695, equivalent F (1,28) = 12.26, p < 0.002), mean futation duration on medium frequency words (A = 0.668, equivalent F (2,27) = 6.72, p < 0.005), and mean fuation duration on low frequency words (A = 0.612, equivalent F (3,28) = 5.49, p < 0.005). The canonical correlation was 0.62, and the model correctly classified 11 of the 15 less skilled readers (73.3%), and 14 of the 15 highly skilled readers (93.3%). What is interesting here is that all of the useful measures involve fixation duration. At the first step, with mean fuation duration used as the only predictor, we correctly classified 10 less skilled readers and 12 highly skilled readers. Adding the predictor of futation duration upon medium frequency words improved these figures slightly to 11 and 12, respectively. As a predictor in isolation, the fuation duration upon medium frequency words was unreliable, but in combination with other measures it made a contribution. The absence from the final model of furation duration upon high frequency words is due to its strong correlation with mean fuation duration (see Table 2). Since mean fuation duration was entered into the stepwise analysis first, on the basis of the highest Wilks’ lambda, it pre-empted the contribution of the highly correlated predictor of futation duration on high frequency words. Cotiipottettts of readittg titlie
All of the measures taken from the eye movement records contributed to the total reading time, and so these measures were entered into a multiple linear regression analysis to determine the most reliable predictors of reading time. In addition, the reading comprehension scores and vocabulary scores were used as variables in this analysis.
’
Wilks’ lambda is the ratio of the within-group sum of squares for a single predictor to the total sum of squares.
Downloaded from las.sagepub.com at CAMBRIDGE UNIV LIBRARY on August 22, 2015
Eye Fixations and Readirlg Comprehensioti
76
TABLE2 Pearson product-moment correlations between the measures taken with pencil and paper tests and with the sentence reading task
VS
RT
Vocab.Read- No. fues score ing time RC Reading comprehension
.34
-.22
VS Vocabulary score
-
RT Total sentence reading time F
No.of fucations
RF
FD
No. regs
dur.
-.08
-.58
F
.07
WU
HF
MF
Fix. Wrap- Fix. Fix. up High Me-
Fix. LOW
dium
dur.
-.27
LF
-.46
-.24
-.44
.06
.15
.05
.03
.02
.07
.04
-
.89
.83
.42
.ll
.3 1
.47
.35
-
.9 1
.12
-.20
.05
.28
.14
-
.23
-.36
.12
.25
.14
-27
.8 1
.56
.63
-
.2 1
.13
.10
.61
-53
-
.60
RF No. of regressive fuations FD hlean fuation duration
WU Wrap-up fuation duration HF Fixation on high frequency words hlF Fixation on medium frequency words LF Fixation on low frequency words Correlations indicated in bold are reliable at p
-.14
-
< 0.05
The method used in the regression analysis was forwar.. entry, wi ... the first variable being that which had the largest correlation with reading time. A regression equation which was found t o be reliable then allowed the procedure to test a further variable for entry. The second variable to be entered was that with the largest partial correlation with reading time. Variables were entered until no further variables had individual reliability. Three variables were admitted, and they accounted for 93.5% of the variance
Downloaded from las.sagepub.com at CAMBRIDGE UNIV LIBRARY on August 22, 2015
G. Uiidenvood,A . Hiibbard, and H. IVilkiitsori
77
in reading time (the adjusted R2 was 0.93).2 These three variables were, in their order of entry, the number of fixations per sentence (adjusted R2 = 0.79 upon entry of the fxst variable), mean futation duration (adjusted R2 = 0.89 when entered), and wrap-up futation duration. The final regression equation was reliable [F (3,26) = 124.3, p < 0.0011.
DISCUSSION The analyses of variance revealed a number of differences between highly skilled and less skilled readers, and these were all associated with the durations of fiations. Skilled readers have futations which are approximately 35 msec briefer than those of less skilled readers. The means presented in Table 1 suggest that there are differences in a number of other measures taken, but these were not reliable. The discriminant analysis used to describe the variables which best classified the individual subjects tended to confirm the independent analyses of variance. In both of these analyses it was futation duration that was highlighted as the most reliable predictor of reading ability. The predictive use of f i a t i o n duration here may be an indication of a general relationship with reading ability, or it may be a consequence of using the particular test of comprehension which we selected. The test has surface validity, in that it appears to require skills which are required for text comprehension - word recognition and integration, and inference, for instance - but other tests of comprehension may possibly correlate less well with fiiation duration. A second caveat concerns the readers sampled in the study. It may be that if we had not tested undergraduate students then the relationship between ability and f i a t i o n duration would have been weaker or absent altogether. However, support for the generality of the relationship comes from a study of reading disability reported by Olson, Kliegl, and Davidson (1983). Their readers were high school students with above average ability and who were reading well below the level predicted on the basis of their general scholastic ability. Two groups of subjects had their eye movements monitored during a reading task. The disabled readers had futation durations which were 50 msec longer than those of non-disabled readers, again suggesting a relationship between reading ability and futation duration. In addition to identifying futation duration as the best predictor of reading comprehension in the undergraduate students sampled in the present experiment, the analyses revealed a number of variables t o be irrelevant. Overall reading time per sentence was not reliably related to reading ability even though there was a considerable difference between the means for the two groups. There was sufficient variability in reading styles to mask any real difference in reading time, and there was greater variance in a number of measures taken from the highly skilled readers. The standard deviations of sentence reading time, number of futations per sentence, and number of regressive fixations The adjusted R2,takes into account the number of predictors in the regression equation, and is more conservative than simple R2
Downloaded from las.sagepub.com at CAMBRIDGE UNIV LIBRARY on August 22, 2015
78
Eye Fixatioris arid Reading Coniprelie~isiort
(Table 1) indicate more variability in the reading styles of better readers. This raises the possibility that better readers are those who more readily adjust their inspection strategy according t o perceived task demands and according t o the interpretation of the material being read. Although observing the relationship between reading ability and futation duration reported by Olson et al. (1983), our study failed to confirm the relationship between ability and other measures. Olson er al. were among the first t o apply multiple regression techniques to the study of eye movement patterns in readers of different ability, and emphasised the importance of the regression f i a t i o n index (RFI) in predicting ability. RFI is the ratio of regressive furations to the number of fvations preceded and succeeded by a rightwards saccade. It correlates at 0.97 with the percentage of regressions, and successfully predicts reading ability. The relationship between the number of regressive futations and reading ability did not approach significance in the present study of highly skilled and less skilled readers, accounting for 14.2%and 16.1% of all futations in the two groups respectively. Whereas Olson et al. tested 13-17 year old disabled and normal readers, we sampled undergraduate students, and so it may be that the emphasis upon regressive futations is appropriate only in the prediction of ability in relatively unskilled readers. In the analyses reported b y Graesser, Hoffman and Clark (1980) a number of structural components of sentences (such as narrativity and number of propositions) were used in a regression analysis t o predict reading time per sentence. The readers in this study were college students, as in the present experiment. The analysis was performed for the fastest readers and for the slowest, and conclusions were drawn concerning differences in reading skill. The absence of a relationship between reading speed and reading comprehension in our experiment advocates caution in the use of an overall measure o f reading time as an estimate of reading ability. Another notable absence from the discriminant analysis was vocabulary score: the size of an individual’s vocabulary does not help us predict whether or not he or she is a good reader - a t least in the relatively skilled readers contributing to our data. It is entirely possible that, as the range of reading ability is extended to more fairly represent the range found outside of a undergraduate population, vocabulary may become a reliable predictor. Indeed, with a more widely based sample of readers, Baddeley, Logie, NimmoSmith, and Brereton (1985) found positive correlations between vocabulary and reading comprehension of 0.33 and 0.48 in two studies. The regression analysis in both of their studies showed vocabulary to be a reliable predictor of comprehension. Whereas our positive correlation of 0.34 between these two variables was similar to that in one of the Baddeley e f al. .studies, vocabulary did not prove to be a reliable predictor or discriminator. The absence of a convincing effect of word frequency upon f i a t i o n duration might also be a consequence of using skilled readers for whom the target words presented n o recognition difficulties. The second question addressed in this study concerned the relationship between variations in sentence reading time and variations in the measures which contribute towards this overall measure of performance. Total reading time constitutes the product of mean f i a t i o n duration and the number of futations, plus saccade durations, plus the time takin for processes involved in sentence comprehension which are assumed to be
Downloaded from las.sagepub.com at CAMBRIDGE UNIV LIBRARY on August 22, 2015
G. Utidenvood, A . Hiibbard, atid H. Wilkbisoti
79
reflected in the wrap-up fixation. The regression analysis which was used to determine the components of reading time revealed that the total number of fixations alone accounted for 79% of the variance in reading time. Adding the variable of fvtation duration took this figure to 89%, and the fmal model included the wrap-up futation and accounted for 93%of the variance. The inclusion of wrap-up futation duration in our regression model confirms the suggestions of Just and Carpenter (1980) concerning the significance of the final fvtation, possibly for inter-word and inter-sentence integration. Saccade duration was not entered into the model, but could be expected to contribute little to the final model. The discriminant analysis showed that skilled and less skilled readers were not differentiated by their numbers of futations per line or by their sentence reading times, however, and that variations in skill are associated with variations in fixation durations alone. By observing that each group makes approximately 10 fixations per line of text, and the highly skilled readers fixate for approximately 35 msec per fixation less than the less skilled readers, we have accounted for most of the difference in reading time between the two groups. The difference was, in fact, in the order of 350 msec per sentence (highly skilled readers: 2.85 seconds; less skilled readers: 3.21 seconds). Although the means for the two groups differ, there was high variance in the individual scores, particularly amongst the highly skilled readers, and the difference was not reliable. The cause of this variation appears to be the variable number of regressive fucations made by the better readers. Table 1 shows that the highly skilled readers make slightly fewer regressive movements (not statistically reliable), but that there was more variability in the number of regressives, in comparison with the poorer readers. Some of the highly skilled readers made very few regressives, while others in effect re-read each sentence before declaring that they had understood it. While overall reading time did not differ reliably between the two groups, the difference in fuation duration proved consistent. In general, these data are in accord with Jackson and hlcclelland’s (1975) view that fast readers pick up more information per fucation because their futations are shorter, but they also make fewer of these fuations. As reading ability was not a predictor of reading speed, we can also conclude that fast readers with shorter and fewer futations gain no advantage or disadvantage by their reduced total inspection of each sentence. Cumming (1977) concluded that reading speed is varied mainly by variations in saccade length. In our experiment this would correspond approximately to variations in the number of futations, with longer saccades allowing fewer fixations per line. The correspondence is not exact because an increased number of regressive fixations could be associated with a constant saccade length and an increased total number of fixations. The regression analysis which determined the components of reading time confirmed Cumling’s conclusion, with the number of fixations accounting for the greatest proportion of the variance in sentence reading time. However, variations in reading ability are not associated with increasing numbers of forward or regressive fixations so much as decreasing durations of those futations. The question of causality is not addressed by the present data. We have no indication of whether the highly skilled readers are more successful as a result of their faster extraction of information from each fuation, or whether efficient information extraction
Downloaded from las.sagepub.com at CAMBRIDGE UNIV LIBRARY on August 22, 2015
80
Eye Fixatioris arid Reading Coniprelierrsioit
has consequences for both f i a t i o n duration and reading comprehension. Masson (1985) reported variability in the amount o f visual attention to each word by individuals who had attended a commercial speed reading course (inspection durations of 220 msec in comparison with durations of 330 msec for normal readers), and considerable reductions in the number of words f i a t e d in each passage of text. These speed readers read at an average of 700 words per minute in comparison with an untrained reider’s average of 240 words per minute, but they were less successful in tests of comprehension and recall. This evidence points in favour of shorter f i a t i o n s being a product of more effective information extraction and integration, rather than shorter f i a t i o n s resulting in these improvements in reading ability. Training ourselves to make shorter fvtations does not result in improved reading comprehension, but training ourselves to comprehend better may have effects upon our inspection strategies and upon our f i a t i o n durations. Of the measures of eye movements which were recorded in the present experiment, the best predictor of reading speed was the number of f i a t i o n s per sentence, and the best predictor of reading comprehension was the average duration of a fixation. Reading ability is related to the extraction of information within a fixation, but what is the cognitive source of this variation in ability? Candidates here include faster propositional integration, faster word recognition, faster syntactic parsing, but all we can conclude from the present data is that whatever cognitive processes are performed faster by more skilled readers, they are reflected in shorter f i a t i o n s upon the text. (Received April 16, 1990; accepted June 27, I990)
REFERENCES AARONSON, D., and SCARBOROUGH, H.S. (1976). Performance theories for sentence coding: Some quantitative evidence. Journal of Experimental Psychology: Human Perception and Performance, 2.56 -70. BADDELEY, A., LOGIE, R., NIhlhlO-ShlITH, I., and BRERETON, N. (1985). Components of fluent reading. Journal of Verbal Learning and Verbal Behavior, 24.119- 13 1. BLACK, E.L. (1962). Compreliension Test f o r College o f Education Students. Slough: NFER P u b lishing Co. CUMhlING, G.D. (1977). Eye movements and visual perception. In E.C. Carterette and hl. Friedman (eds.), Handbook of Perceptioti. Volume 9 Cpp. 221-255). New York: Academic Press. EHRLICH, S.F., and RAYNER,K. (1981). Contextual effects o n word perception and eye movements during reading. Journal of Verbal Learning and Verbal Behavior, 20,64 1-655. GRAESSEK, A.C., HOFFMAN, N.L., and CLAKK, L.P. (1980). Structural components o f reading timc. Journal of Verbal Leariling and Verbal Behavior. 19,135-15 1. HYONA, J., NIEMI, P., and UNDERWOOD, G. (1989). Reading long words embedded in sentences: Informativeness of word halves affects eye movements. Journal of Experitnenful Psychology: Human Perception and Performance, 15, 142-152. JACKSON, hI.D., and hfCCLELLAND, J.L. (1975). Sensory and cognitive determinants of reading speed. Journal of Verbal Learning aitd Verbal Behavior. 14,565 -574. JUST, hI.A., and CARPENTER. P.A. (1978). Inference processes during reading: Reflections from ‘eye fiiations. In J.W. Senders, D.F. Fisher, and K.A. Monty (eds), Eyefifovemeiltsand rlze Higher Psychological Processes (pp. 157-174). Hillsdale, NJ: Erlbaum.
Downloaded from las.sagepub.com at CAMBRIDGE UNIV LIBRARY on August 22, 2015
G. Uiidenvood, A . Hubbard. arid H. IViIkiitson
81
JUST, hl.A., and CARPENTER, P.A. (1980). A theory of reading: From eye fiixations t o comprehension. Psychological Review, 87.329-354. KENNEDY, R.A. (1978). Eye movements and the integration o f semantic information during reading. In h1.M. Gruneberg, P.E. hforris, and R.N. Sykes (eds), Practical Aspects of Memory (pp. 484-490). London: Academic Press. KENNEDY, R.A., and hIURRAY, W.S. (1987). The components of reading time: Eye movement patterns of good and poor readers. In J.K. O'Regan and A. Lkvy-Schoen (eds.), Eye hlorenieiits: FfOIlZ Physiology t o Cognition (pp. 509-520). Amsterdam: North-Holland. KUCERA, H., and FRANCIS, W. (1967). Coniputatiorial Analysis of Present-day Anierican English. Providence, RI: Brown University Press. BIASSON, h1.E.J. (1985). Rapid reading processes and skills. In G.E. hIacKinnon and T.G. Waller (eds.), Reading Researcli: Advances in Theory and Practice, Voliiine 4 (pp. 183-230). New York: Academic Press. h1URRAY.W.S. and KENNEDY, A. (1988). Spatial coding in the processing of anaphor by good and poor readers: Evidence from eye movement analyses. Quarterly Journal of Experirizerztal PSyCliOlOg~,40A, 693-3 18. OLSON, R.K., KLIEGL, R., and DAVIDSON, B.J. (1983). Eye movements in reading disability. In K. Rayner (ed.), Eye Aloveaients in Reading: Perceptual and Language Processes (pp. 467479). New York: Academic Press. O'REGAN, J.K., LEVY-SCHOEN. A., PYNTE. J., and BRUGALLERE, B. (1984). Convenient fixation location with isolated words of different Iength and structure. Jotinial of Experinteiifa1Psycliology: I h i i i a n Perception and Performance, 10,250-257. RAVEN, J.C. (1977).The Mill-Hill Vocabtilary Scale. London: Lewis. RAY NER, K. (1978). Eye movements in reading and information processing. Psycliological Bulletin,
85,618-660.
RAYNER, K., CARLSON. hl., and FRAZIER, L. (1983). The interaction of syntax and semantics during sentence processing: Eye movements in the analysis of semantically biased sentences Jotirnal of Verbal Learning and Verbal Behavior, 22,358-374. RAYNER, K., and DUFFY, S.A. (1986).Lexical complexity and fixation times in reading: Effects of word frequency, verb complexity, and lexical ambiguity. Memory & Cognition, 14, 191-
201.
TATSUOKA, hl. (1971). hltiltivariate Analysis: Tecliniqiies for Ediictiorial and Psychological Research. New York: Wiley. UNDERWOOD, G., BLOOMFIELD, R., and CLEWS, S. (1988). Information influences the pattern of eye fixations during sentence comprehension. Perception. 17,267-278. UNDERWOOD. G., CLEWS, S., and EVERATT, J . (1990). How d o readers know where t o look next? Local information distributions influence eye fixations. Quarterly Joiirnal of Experiniental PSycltOlOgy. 421%.39-65. WILKINSON, H.P. (1976). \Vide range eye and head movement monitor. Qiiarterly Jorirnalof Experirneiital Psycliology, 28, 123-124.
Downloaded from las.sagepub.com at CAMBRIDGE UNIV LIBRARY on August 22, 2015
69
EYE FIXATIONS PREDICT READING COMPREHENSION: THE RELATIONSHIPS BETWEEN READING SKILL, READING SPEED, AND VISUAL INSPECTION” GEOFFREYUNDERWOOD, ALISONHUBBARD, and HOWARD WILKINSON Utiiversity of Nottingliani. U.K. This experiment addressed the question of whether reading comprehension and speed could be predicted by eye futations. From a sample of university students who completed tests of reading comprehension and vocabulary, we selected a group of highly skilled readers and a group of less skilled readers. These two groups then read sentences as their eye movements were monitored, with fixation locations and durations recorded. A discriminant function analysis showed that futation duration was a successful predictor of reading comprehension, but that the number of Lxations, regressive fmations, reading speed, and vocabulary were not reliable predictors A multiple regression analysis revealed that reading speed was predicted by the number of fixations, the average fixation duration, and the duration of the final fixation upon the sentence, but there was no relationship with reading ability. Highly skilled readers are those who can extract information efficiently, but are not necessarily those who have fast overall reading rates.
Key words: reading skill, eye movements
INTRODUCTION Is there a relationship between the acquisition of information from printed text and the ability to use that information? A reader’s pattern of eye f i a t i o n s may be described in t e r m of the locations and durations of futations, the number of fixations, and.the number of regressive futations. A reader’s characteristic pattern of information extraction may be represented by any combination o f these measures. Alternatively, individual readers may have no characteristic pattern of eye movements and fiiations which can be related to their use of the information being extracted. The movements of a reader’s eyes may have little in common with the understanding of the written language being f i a t e d , and there is certainly more to reading than moving one’s eyes in a particular way. However, recognising individual word meanings, integrating syntactically related words, and making inferences about the relationships between events
*
We aregrateful to Alan Kennedy, Sandy Pollatsek, and Bruno Repp for their comments on this paper. Joanne Leach, Helen hlacleod, Frances Villain, and Belinda Winder provided valuable assistance during the administration and scoring of the reading and vocabulary tests. Send correspondence t o Geoffrey Underwood, Department of Psychology, University
of Nottingham, Nottingham NG7 2RD, U.K.
Downloaded from las.sagepub.com at CAMBRIDGE UNIV LIBRARY on August 22, 2015
70
Eye Fixations arid Reading Coniprehertsiort
and actors in the text are all time-consuming processes, and the time course of processing might be expected to be indicated by the time taken to extract the information from the display. The length of the terminating (“wrap-~p’~) futation also adds to the total reading time when sentences are displayed sequentially, and especially when the reading task demands a decision to be made on completion of the reading of each sentence. Just and Carpenter (1978) presented their sentences sequentially and found strong evidence for long wrap-up futations, and in reading studies which have presented texts word-by-word there is evidence of longer pauses at the ends of sentences (Aaronson and Scarborough, 1976). It is less clear whether the fmal futation on a sentence differs from other fiations in more natural reading exercises. However, in a study by Just and Carpenter (1980), in which short paragraphs were displayed and eye movements monitored, there was evidence of long wrap-up furations. This study developed a regression model in which wrap-up fiations made a’ reliable 71 msec per futation contribution to the equation for total reading time. When eye movements have been observed, the evidence suggests that there is a relationship between the difficulty of processing and the futation pattern upon the word. In the case of individual words, Rayner and Duffy (1986) found longer fuation durations upon low frequency words (e.g., goitdola) than upon high frequency words (e.g., vehicle), and also found longer futations upon ambiguous nouns with equally likely meanings (e.g., coach) than upon ambiguous nouns with a dominant meaning (e.g., perch). These demonstrations indicate that the difficulty of accessing a word’s meaning is indicated by the duration of the initial futation upon the word. Within a word there may be an uneven distribution of information, with the final letters being more or less useful in its identification. Compare, for example, the first five and last five letters of the two words nioralistic and supervisor. In the first example the beginning of the word is most informative (try deleting the first five letters and having someone guess the word, and then try it again, with the last five letters deleted instead), but for supervisor the ending is more informative. This uneven distribution has an effect upon futation durations upon each half of the word. O’Regan, Le’vy-Schoen, F‘ynte, and Brugalle’re (1984), Underwood, Bloomfield, and Clews (1988), and Hyona, Niemi, and Underwood (1989) found longer futation durations upon informative ends of long words. The number of futations upon word-halves also increased for halves containing more information. Underwood, Clews, and Everatt (1990), in extending this investigation of the effects of information distribution upon fiiation strategies,. have found a trade-off between measures of duration and measures of location. When only one fiation is made upon a long word, then it tends to last longer than when two futations have to be made. These results together suggest a sensitivity of fuation durations to the microprocessing of information within words. Words which would not by themselves be difficult to process can be placed in sentences in such a way as to produce local processing difficulties. The prior context in a sentence can reduce the duration of a fixation if the context and the word are congruous (Ehrlich and Rayner, 1981). In a related experiment, Kennedy (1978) reported that anaphors with high conjoint frequency gain shorter fuations than those
Downloaded from las.sagepub.com at CAMBRIDGE UNIV LIBRARY on August 22, 2015
G. Underwood, A. Hiibbard, arid H. lvilkirisoii
71
preceded by low conjoint associates. The presentation of the word vehicle in one sentence resulted in less visual attention being allocated to the word btrs in the following sentence, in comparison with paragraphs in which vehicle was followed by tank. This could be interpreted as a word priming effect, with a strong associate providing greater facilitation than a weaker associate. Difficult syntactic structures also result in longer f i a t i o n durations than simpler forms (Rayner, Carlson, and Frazier, 1983). These obscrvations of the effects of processing difficulties upon futations were made with homogeneous samples of readers, but a similar case can be made for samples of readers who find text comprehension difficult, in comparison with more skilled readers. Fixation patterns indicate a moment-to-moment sensitivity to comprehension difficulty, and so perhaps we should expect that individuals with less reading skill will also show their comprehension difficulty by producing furation profiles which differ from those of more skilled readers. Although there is some evidence of a decrease in futation duration with increasing age in children (see Rayner, 1978), there is little evidence of a relationship between reading ability and inspection duration when reading experience is controlled (Kennedy and Murray, 1987; hlurray and Kennedy, 1988). In the present study we compared inspection patterns during reading using two groups of adults who varied in their reading abilities. The two groups were highly educated, but differed in their performance on a standard test of reading comprehension. The purpose was to determine whether readers could be differentiated by the measures of eye fixations taken during a sentence comprehension task.
Stibjects
One hundred University of Nottingham students completed a reading comprehension exercise based on the Black (1962) Comprehension Test for College of Education Students, in pencil and paper format. All subjects were native English speakers. The reading test presents short passages from published works (fiction, biography, and newspaper articles), followed by comprehension questions. The passages continue to be available during completion of the questions. The test took 30 minutes and consisted of a number of types of items, most of which offered multi-choice answers (five choices). Some items required the interpretation of words in the context of the passage, other items required words missing from a paragraph to be selected from five alternatives (a “cloze” test), and others required textual interpretation. For example, in one paragraph which described the activities and life style of a primitive tribe, the words feud, vendetta, and warrior appeared in sentences preceding the tested sentence: “The iiiiiiieroiis tribes arid conrbiiiations of tribes all have their accoiiiits to settle with one another, and very few debts are left triipaid’: A test item then asked whether these debts were: (i) unpaid loans, . . . or (v) murders unavenged. Yet other items required assessment of the perspective of the writer. The maximum score on our version of the test was 25, and the subjects ranged from nine t o 22 answers correct. The same 100 students also completed a vocabulary test based on the more difficult items taken from the hlill-Hill
Downloaded from las.sagepub.com at CAMBRIDGE UNIV LIBRARY on August 22, 2015
72
Eye Fivatioris arid Readiiig Conipreheiuioii
Vocabulary Scale (Raven, 1977). These items required that low frequency target words be matched with one of six possible synonyms. The maximum score on our version of this test was 30, and subjects scored between four and 23 items correct. On the basis of the reading comprehension test, the 15 highest scoring and the 15 lowest scoring subjects were invited to participate in the experiment which recorded eye movements during reading. The mean reading scores for the highest scorers was 20.7 (sd = 0.79), and their mean vocabulary score was 13.2 (sd = 3.82). The mean reading score for the lowest scorers was 10.7 (sd = 1.33) and their mean vocabulary score was 10.47 (sd = 3.96). As these subjects were drawn from a university population, and could therefore all be considered to be good readers relative to the population as a whole, we shall refer to the two groups as consisting of highly skilled readers and less skilled readers. Stiriiiihis materials
Three sets of sentences formed the experimental materials, with each set containing 15 sentences which represented a brief story. In addition to these 45 test sentences, six practice sentences were prepared. Each sentence contained between six and 12 words (33-52 character spaces) with an average of 8.5 words per sentence, and was displayed as a single line of text. Each contained a target word near its centre. The target words varied in frequency (Kucera and Francis, 1967). High frequency targets had counts in excess of 150 (e.g., turn, brown, little), medium frequency words were between 30 and 70 (e.g., soft, trend, tuefiil), and low frequency words had counts below 10 (e.g., twin, srreak, uproar). These words varied between four and seven letters in length, with equal numbers of four, five, six and seven letter words in each of the three frequency categories. The target words were organised into triplets which were to be presented in identical contexts, to ensure that variations in eye futations upon the targets could not be attributed to variations in the preceding contexts. Within each triplet was a high, medium, and low frequency word. As an example, consider the following triplet of words in their sentences: High frequency:
There was a little noise from under the bushes.
Medium frequency:
There was a treiid to vote for someone uninvolved.
Low frequency:
There was a screech as the woman hit the brakes.
The target words are highlighted here but were not, of course, in the experimental presentations. Procediire
Each subject read six practice and 45 experimental sentences, with the instruction to read each paragraph for meaning and in preparation for a comprehension question. At the start of the experiment the eye movement tracker was calibrated and the head restraint set for each subject. The 45 sentences were presented in three blocks of 15, with the order of sentences within a block being maintained between subjects, and the order of the blocks counterbalanced. Each trial consisted of a fvtation mark first appearing in the centre of the screen, and
Downloaded from las.sagepub.com at CAMBRIDGE UNIV LIBRARY on August 22, 2015
G. Uitdenvood. A. Hiibbard, and H. IVilkiiisoii
73
then at a position on the screen corresponding t o the first letter of the sentence. The sentence appeared 0.5 seconds later, and the subjects indicated completion of reading by pressing a hand-held button. The sentence was then removed, and the next trial started. Between blocks of trials the equipment was re-calibrated, and between sentences there was an opportunity for an adjustment if necessary. At the end of the practice sentences, and at the end of the complete experiment, comprehension questions were asked about each block of sentences to check that understanding had been achieved. Subjects had no difficulty in answering the questions correctly.
Apparatus The sentences were displayed on a high resolution monitor (Taxan KX-12) controlled by a BBC Master microcomputer. At the viewing distance of 48 cm a ten-letter word subtended a horizontal angle of 3 degrees. hiaximum letter height was 0.6 degrees. The microcomputer also received inputs from the eye-movement recorder which sampled every 10 msec. The recorder detected the horizontal position of the eye by monitoring the reflection of an infra-red beam from the iris-sclera border of the left eye. This recorder has been described elsewhere, and gives an accuracy of plus or minus one character space (Wilkinson, 1976). The infra-red source and photoelectric detector were mounted on a headband, which was adjusted for each subject. This procedure locked the positions of the eye movement recorder and the subject’s head, and head movement was restricted by a chin rest and head frame. Displays were terminated with a button press by the subject. The output from the eye position recorder was interpreted with the aid of a program which converted sampled positions into eye fiations. The fEation detection algorithm was based on a running average technique, and can be summarised as follows. The difference between the first and second values of ‘the recorder’s output was computed and compared with a set predetermined value, called W. If this difference was not in excess of the value of W, an average of the two values was computed. The difference between the third reading from the recorder and this computed average was then compared with W, and a new average of the three readings was computed if the difference was again not in excess of W. This process was repeated until, after a number of readings, the comparison exceeded W whereupon a futation was deemed to have finished and a saccade started. Iteration of this algorithm identified the end of this saccade and hence the start of the next furation. The value of W was determined primarily by the variation of the output of the recorder when a typical subject attempted t o furate a single letter character. Once determined at the start of the investigation the value of W was held const ant.
RESULTS From each sentence a number of measures were taken in preparation for three sets of analyses. The measures were total reading time, total number of fixations, number of regressive fiiations, average fiation duration (excluding the first and last futations),
Downloaded from las.sagepub.com at CAMBRIDGE UNIV LIBRARY on August 22, 2015
74
Eye Fixatioris arid Readiiig Coniprelieiisiori TABLE1 Means for the two groups of readers for the measures taken from each sentence (Standard deviations in brackets) Highly skilled readers
Less skilled readers
Total sentence reading time (sec)
2.85 (1.08)
3.21
(0.76)
Number of futations per sentence
10.58 (3.63)
10.13
(2.16)
1.75 (1.68)
1.95
(1.21)
Number of regressive futations Overall futation.duration (msec) [excluding wrap-up f ~ a t i o n ]
249.1 (20.1)
284.7
Wrap-up futation duration (msec)
326.2 (90.2)
386.0 (144.6)
(33.8)
Fixation duration on target word (msec):
- high frequency words - medium frequency words
226.5 (30.3)
262.8
(46.0)
241.2 (32.6)
254.0
(36.6)
-low frequency words
237.1 (20.3)
288.6
(67.5)
duration of the final (wrap-up) futation, and duration of the fixation upon the target word. The means of these measures. are shown in Table 1. The analyses looked for differences between the two groups of readers on each measure, the discriminant function which best describes the characteristics of the two groups, and the regression function which best describes the time taken to read a sentence.
Differeiices betweeii the two groups of readers The measures taken from each sentence were averaged for each subject, and submitted to separate analyses of variance. Reading time per sentence did not differ reliably between the highly skilled and less skilled readers [F (1,28) = 1.113, neither did the number of futations per sentence [F (1,28) < 11, and neither did the number of regressive futations per sentence [F (1,28) < I]. Mean futation duration did vary between the groups [F (1,28) = 12.26, p < 0.0021, with highly skilled readers futating for approximately 35 msec less than the less skilled readers. The more specific measure of the final wrap-up futation duration did not differ reliably between groups [F (1,28) = 1.871. The final analysis concerned the futation durations upon target words, for the first futation only, and yielded a significant main effect of reading skill [F ( I , 27) = 6.77, p < 0.021, a marginal effect of frequency [F (2,54)= 3.10, p < 0.061 and a marginal interaction [F (2,54) = 3.08, p < 0.061. Analyses of simple main effects showed differences between the two groups
Downloaded from las.sagepub.com at CAMBRIDGE UNIV LIBRARY on August 22, 2015
G. Uttdenvood, A . Hiibbard, arid H. IViIkitisoii
75
of readers for high frequency words [F (1,27) = 6.08, p < 0.031 and for low frequency words [F (1,27) = 7.44, p < 0.021. There was no difference between groups for the durations of futations on medium frequency words [F (1,27) < I].
Discrimiiiatitig betweeit grotips of readers To describe the differences that best differentiate between highly skilled and less skilled readers, a discriminant function analysis was performed. Tatsuoka (1971) recommended the use of this procedure as a follow-up to analysis of variance, to determine the best linear combination of variables which will maximally differentiate the groups. The variables entered into the discriminant function analysis were those listed in Table 1 together with vocabulary score. This stepwise analysis first determined Wilks’ lambda ratio’ between groups for each discriminating variable, and entered the variable with the smallest ratio into the model. After entering this variable, the remaining variables were inspected, and the variable with the smallest ratio then entered. The process was continued until no remaining variable had an acceptable F-ratio associated with its Wilks’ lambda value. At the end of the process there were three variables in the model. These were, in order of entry, mean f i a t i o n duration (A = 0.695, equivalent F (1,28) = 12.26, p < 0.002), mean futation duration on medium frequency words (A = 0.668, equivalent F (2,27) = 6.72, p < 0.005), and mean fuation duration on low frequency words (A = 0.612, equivalent F (3,28) = 5.49, p < 0.005). The canonical correlation was 0.62, and the model correctly classified 11 of the 15 less skilled readers (73.3%), and 14 of the 15 highly skilled readers (93.3%). What is interesting here is that all of the useful measures involve fixation duration. At the first step, with mean fuation duration used as the only predictor, we correctly classified 10 less skilled readers and 12 highly skilled readers. Adding the predictor of futation duration upon medium frequency words improved these figures slightly to 11 and 12, respectively. As a predictor in isolation, the fuation duration upon medium frequency words was unreliable, but in combination with other measures it made a contribution. The absence from the final model of furation duration upon high frequency words is due to its strong correlation with mean fuation duration (see Table 2). Since mean fuation duration was entered into the stepwise analysis first, on the basis of the highest Wilks’ lambda, it pre-empted the contribution of the highly correlated predictor of futation duration on high frequency words. Cotiipottettts of readittg titlie
All of the measures taken from the eye movement records contributed to the total reading time, and so these measures were entered into a multiple linear regression analysis to determine the most reliable predictors of reading time. In addition, the reading comprehension scores and vocabulary scores were used as variables in this analysis.
’
Wilks’ lambda is the ratio of the within-group sum of squares for a single predictor to the total sum of squares.
Downloaded from las.sagepub.com at CAMBRIDGE UNIV LIBRARY on August 22, 2015
Eye Fixations and Readirlg Comprehensioti
76
TABLE2 Pearson product-moment correlations between the measures taken with pencil and paper tests and with the sentence reading task
VS
RT
Vocab.Read- No. fues score ing time RC Reading comprehension
.34
-.22
VS Vocabulary score
-
RT Total sentence reading time F
No.of fucations
RF
FD
No. regs
dur.
-.08
-.58
F
.07
WU
HF
MF
Fix. Wrap- Fix. Fix. up High Me-
Fix. LOW
dium
dur.
-.27
LF
-.46
-.24
-.44
.06
.15
.05
.03
.02
.07
.04
-
.89
.83
.42
.ll
.3 1
.47
.35
-
.9 1
.12
-.20
.05
.28
.14
-
.23
-.36
.12
.25
.14
-27
.8 1
.56
.63
-
.2 1
.13
.10
.61
-53
-
.60
RF No. of regressive fuations FD hlean fuation duration
WU Wrap-up fuation duration HF Fixation on high frequency words hlF Fixation on medium frequency words LF Fixation on low frequency words Correlations indicated in bold are reliable at p
-.14
-
< 0.05
The method used in the regression analysis was forwar.. entry, wi ... the first variable being that which had the largest correlation with reading time. A regression equation which was found t o be reliable then allowed the procedure to test a further variable for entry. The second variable to be entered was that with the largest partial correlation with reading time. Variables were entered until no further variables had individual reliability. Three variables were admitted, and they accounted for 93.5% of the variance
Downloaded from las.sagepub.com at CAMBRIDGE UNIV LIBRARY on August 22, 2015
G. Uiidenvood,A . Hiibbard, and H. IVilkiitsori
77
in reading time (the adjusted R2 was 0.93).2 These three variables were, in their order of entry, the number of fixations per sentence (adjusted R2 = 0.79 upon entry of the fxst variable), mean futation duration (adjusted R2 = 0.89 when entered), and wrap-up futation duration. The final regression equation was reliable [F (3,26) = 124.3, p < 0.0011.
DISCUSSION The analyses of variance revealed a number of differences between highly skilled and less skilled readers, and these were all associated with the durations of fiations. Skilled readers have futations which are approximately 35 msec briefer than those of less skilled readers. The means presented in Table 1 suggest that there are differences in a number of other measures taken, but these were not reliable. The discriminant analysis used to describe the variables which best classified the individual subjects tended to confirm the independent analyses of variance. In both of these analyses it was futation duration that was highlighted as the most reliable predictor of reading ability. The predictive use of f i a t i o n duration here may be an indication of a general relationship with reading ability, or it may be a consequence of using the particular test of comprehension which we selected. The test has surface validity, in that it appears to require skills which are required for text comprehension - word recognition and integration, and inference, for instance - but other tests of comprehension may possibly correlate less well with fiiation duration. A second caveat concerns the readers sampled in the study. It may be that if we had not tested undergraduate students then the relationship between ability and f i a t i o n duration would have been weaker or absent altogether. However, support for the generality of the relationship comes from a study of reading disability reported by Olson, Kliegl, and Davidson (1983). Their readers were high school students with above average ability and who were reading well below the level predicted on the basis of their general scholastic ability. Two groups of subjects had their eye movements monitored during a reading task. The disabled readers had futation durations which were 50 msec longer than those of non-disabled readers, again suggesting a relationship between reading ability and futation duration. In addition to identifying futation duration as the best predictor of reading comprehension in the undergraduate students sampled in the present experiment, the analyses revealed a number of variables t o be irrelevant. Overall reading time per sentence was not reliably related to reading ability even though there was a considerable difference between the means for the two groups. There was sufficient variability in reading styles to mask any real difference in reading time, and there was greater variance in a number of measures taken from the highly skilled readers. The standard deviations of sentence reading time, number of futations per sentence, and number of regressive fixations The adjusted R2,takes into account the number of predictors in the regression equation, and is more conservative than simple R2
Downloaded from las.sagepub.com at CAMBRIDGE UNIV LIBRARY on August 22, 2015
78
Eye Fixatioris arid Reading Coniprelie~isiort
(Table 1) indicate more variability in the reading styles of better readers. This raises the possibility that better readers are those who more readily adjust their inspection strategy according t o perceived task demands and according t o the interpretation of the material being read. Although observing the relationship between reading ability and futation duration reported by Olson et al. (1983), our study failed to confirm the relationship between ability and other measures. Olson er al. were among the first t o apply multiple regression techniques to the study of eye movement patterns in readers of different ability, and emphasised the importance of the regression f i a t i o n index (RFI) in predicting ability. RFI is the ratio of regressive furations to the number of fvations preceded and succeeded by a rightwards saccade. It correlates at 0.97 with the percentage of regressions, and successfully predicts reading ability. The relationship between the number of regressive futations and reading ability did not approach significance in the present study of highly skilled and less skilled readers, accounting for 14.2%and 16.1% of all futations in the two groups respectively. Whereas Olson et al. tested 13-17 year old disabled and normal readers, we sampled undergraduate students, and so it may be that the emphasis upon regressive futations is appropriate only in the prediction of ability in relatively unskilled readers. In the analyses reported b y Graesser, Hoffman and Clark (1980) a number of structural components of sentences (such as narrativity and number of propositions) were used in a regression analysis t o predict reading time per sentence. The readers in this study were college students, as in the present experiment. The analysis was performed for the fastest readers and for the slowest, and conclusions were drawn concerning differences in reading skill. The absence of a relationship between reading speed and reading comprehension in our experiment advocates caution in the use of an overall measure o f reading time as an estimate of reading ability. Another notable absence from the discriminant analysis was vocabulary score: the size of an individual’s vocabulary does not help us predict whether or not he or she is a good reader - a t least in the relatively skilled readers contributing to our data. It is entirely possible that, as the range of reading ability is extended to more fairly represent the range found outside of a undergraduate population, vocabulary may become a reliable predictor. Indeed, with a more widely based sample of readers, Baddeley, Logie, NimmoSmith, and Brereton (1985) found positive correlations between vocabulary and reading comprehension of 0.33 and 0.48 in two studies. The regression analysis in both of their studies showed vocabulary to be a reliable predictor of comprehension. Whereas our positive correlation of 0.34 between these two variables was similar to that in one of the Baddeley e f al. .studies, vocabulary did not prove to be a reliable predictor or discriminator. The absence of a convincing effect of word frequency upon f i a t i o n duration might also be a consequence of using skilled readers for whom the target words presented n o recognition difficulties. The second question addressed in this study concerned the relationship between variations in sentence reading time and variations in the measures which contribute towards this overall measure of performance. Total reading time constitutes the product of mean f i a t i o n duration and the number of futations, plus saccade durations, plus the time takin for processes involved in sentence comprehension which are assumed to be
Downloaded from las.sagepub.com at CAMBRIDGE UNIV LIBRARY on August 22, 2015
G. Utidenvood, A . Hiibbard, atid H. Wilkbisoti
79
reflected in the wrap-up fixation. The regression analysis which was used to determine the components of reading time revealed that the total number of fixations alone accounted for 79% of the variance in reading time. Adding the variable of fvtation duration took this figure to 89%, and the fmal model included the wrap-up futation and accounted for 93%of the variance. The inclusion of wrap-up futation duration in our regression model confirms the suggestions of Just and Carpenter (1980) concerning the significance of the final fvtation, possibly for inter-word and inter-sentence integration. Saccade duration was not entered into the model, but could be expected to contribute little to the final model. The discriminant analysis showed that skilled and less skilled readers were not differentiated by their numbers of futations per line or by their sentence reading times, however, and that variations in skill are associated with variations in fixation durations alone. By observing that each group makes approximately 10 fixations per line of text, and the highly skilled readers fixate for approximately 35 msec per fixation less than the less skilled readers, we have accounted for most of the difference in reading time between the two groups. The difference was, in fact, in the order of 350 msec per sentence (highly skilled readers: 2.85 seconds; less skilled readers: 3.21 seconds). Although the means for the two groups differ, there was high variance in the individual scores, particularly amongst the highly skilled readers, and the difference was not reliable. The cause of this variation appears to be the variable number of regressive fucations made by the better readers. Table 1 shows that the highly skilled readers make slightly fewer regressive movements (not statistically reliable), but that there was more variability in the number of regressives, in comparison with the poorer readers. Some of the highly skilled readers made very few regressives, while others in effect re-read each sentence before declaring that they had understood it. While overall reading time did not differ reliably between the two groups, the difference in fuation duration proved consistent. In general, these data are in accord with Jackson and hlcclelland’s (1975) view that fast readers pick up more information per fucation because their futations are shorter, but they also make fewer of these fuations. As reading ability was not a predictor of reading speed, we can also conclude that fast readers with shorter and fewer futations gain no advantage or disadvantage by their reduced total inspection of each sentence. Cumming (1977) concluded that reading speed is varied mainly by variations in saccade length. In our experiment this would correspond approximately to variations in the number of futations, with longer saccades allowing fewer fixations per line. The correspondence is not exact because an increased number of regressive fixations could be associated with a constant saccade length and an increased total number of fixations. The regression analysis which determined the components of reading time confirmed Cumling’s conclusion, with the number of fixations accounting for the greatest proportion of the variance in sentence reading time. However, variations in reading ability are not associated with increasing numbers of forward or regressive fixations so much as decreasing durations of those futations. The question of causality is not addressed by the present data. We have no indication of whether the highly skilled readers are more successful as a result of their faster extraction of information from each fuation, or whether efficient information extraction
Downloaded from las.sagepub.com at CAMBRIDGE UNIV LIBRARY on August 22, 2015
80
Eye Fixatioris arid Reading Coniprelierrsioit
has consequences for both f i a t i o n duration and reading comprehension. Masson (1985) reported variability in the amount o f visual attention to each word by individuals who had attended a commercial speed reading course (inspection durations of 220 msec in comparison with durations of 330 msec for normal readers), and considerable reductions in the number of words f i a t e d in each passage of text. These speed readers read at an average of 700 words per minute in comparison with an untrained reider’s average of 240 words per minute, but they were less successful in tests of comprehension and recall. This evidence points in favour of shorter f i a t i o n s being a product of more effective information extraction and integration, rather than shorter f i a t i o n s resulting in these improvements in reading ability. Training ourselves to make shorter fvtations does not result in improved reading comprehension, but training ourselves to comprehend better may have effects upon our inspection strategies and upon our f i a t i o n durations. Of the measures of eye movements which were recorded in the present experiment, the best predictor of reading speed was the number of f i a t i o n s per sentence, and the best predictor of reading comprehension was the average duration of a fixation. Reading ability is related to the extraction of information within a fixation, but what is the cognitive source of this variation in ability? Candidates here include faster propositional integration, faster word recognition, faster syntactic parsing, but all we can conclude from the present data is that whatever cognitive processes are performed faster by more skilled readers, they are reflected in shorter f i a t i o n s upon the text. (Received April 16, 1990; accepted June 27, I990)
REFERENCES AARONSON, D., and SCARBOROUGH, H.S. (1976). Performance theories for sentence coding: Some quantitative evidence. Journal of Experimental Psychology: Human Perception and Performance, 2.56 -70. BADDELEY, A., LOGIE, R., NIhlhlO-ShlITH, I., and BRERETON, N. (1985). Components of fluent reading. Journal of Verbal Learning and Verbal Behavior, 24.119- 13 1. BLACK, E.L. (1962). Compreliension Test f o r College o f Education Students. Slough: NFER P u b lishing Co. CUMhlING, G.D. (1977). Eye movements and visual perception. In E.C. Carterette and hl. Friedman (eds.), Handbook of Perceptioti. Volume 9 Cpp. 221-255). New York: Academic Press. EHRLICH, S.F., and RAYNER,K. (1981). Contextual effects o n word perception and eye movements during reading. Journal of Verbal Learning and Verbal Behavior, 20,64 1-655. GRAESSEK, A.C., HOFFMAN, N.L., and CLAKK, L.P. (1980). Structural components o f reading timc. Journal of Verbal Leariling and Verbal Behavior. 19,135-15 1. HYONA, J., NIEMI, P., and UNDERWOOD, G. (1989). Reading long words embedded in sentences: Informativeness of word halves affects eye movements. Journal of Experitnenful Psychology: Human Perception and Performance, 15, 142-152. JACKSON, hI.D., and hfCCLELLAND, J.L. (1975). Sensory and cognitive determinants of reading speed. Journal of Verbal Learning aitd Verbal Behavior. 14,565 -574. JUST, hI.A., and CARPENTER. P.A. (1978). Inference processes during reading: Reflections from ‘eye fiiations. In J.W. Senders, D.F. Fisher, and K.A. Monty (eds), Eyefifovemeiltsand rlze Higher Psychological Processes (pp. 157-174). Hillsdale, NJ: Erlbaum.
Downloaded from las.sagepub.com at CAMBRIDGE UNIV LIBRARY on August 22, 2015
G. Uiidenvood, A . Hubbard. arid H. IViIkiitson
81
JUST, hl.A., and CARPENTER, P.A. (1980). A theory of reading: From eye fiixations t o comprehension. Psychological Review, 87.329-354. KENNEDY, R.A. (1978). Eye movements and the integration o f semantic information during reading. In h1.M. Gruneberg, P.E. hforris, and R.N. Sykes (eds), Practical Aspects of Memory (pp. 484-490). London: Academic Press. KENNEDY, R.A., and hIURRAY, W.S. (1987). The components of reading time: Eye movement patterns of good and poor readers. In J.K. O'Regan and A. Lkvy-Schoen (eds.), Eye hlorenieiits: FfOIlZ Physiology t o Cognition (pp. 509-520). Amsterdam: North-Holland. KUCERA, H., and FRANCIS, W. (1967). Coniputatiorial Analysis of Present-day Anierican English. Providence, RI: Brown University Press. BIASSON, h1.E.J. (1985). Rapid reading processes and skills. In G.E. hIacKinnon and T.G. Waller (eds.), Reading Researcli: Advances in Theory and Practice, Voliiine 4 (pp. 183-230). New York: Academic Press. h1URRAY.W.S. and KENNEDY, A. (1988). Spatial coding in the processing of anaphor by good and poor readers: Evidence from eye movement analyses. Quarterly Journal of Experirizerztal PSyCliOlOg~,40A, 693-3 18. OLSON, R.K., KLIEGL, R., and DAVIDSON, B.J. (1983). Eye movements in reading disability. In K. Rayner (ed.), Eye Aloveaients in Reading: Perceptual and Language Processes (pp. 467479). New York: Academic Press. O'REGAN, J.K., LEVY-SCHOEN. A., PYNTE. J., and BRUGALLERE, B. (1984). Convenient fixation location with isolated words of different Iength and structure. Jotinial of Experinteiifa1Psycliology: I h i i i a n Perception and Performance, 10,250-257. RAVEN, J.C. (1977).The Mill-Hill Vocabtilary Scale. London: Lewis. RAY NER, K. (1978). Eye movements in reading and information processing. Psycliological Bulletin,
85,618-660.
RAYNER, K., CARLSON. hl., and FRAZIER, L. (1983). The interaction of syntax and semantics during sentence processing: Eye movements in the analysis of semantically biased sentences Jotirnal of Verbal Learning and Verbal Behavior, 22,358-374. RAYNER, K., and DUFFY, S.A. (1986).Lexical complexity and fixation times in reading: Effects of word frequency, verb complexity, and lexical ambiguity. Memory & Cognition, 14, 191-
201.
TATSUOKA, hl. (1971). hltiltivariate Analysis: Tecliniqiies for Ediictiorial and Psychological Research. New York: Wiley. UNDERWOOD, G., BLOOMFIELD, R., and CLEWS, S. (1988). Information influences the pattern of eye fixations during sentence comprehension. Perception. 17,267-278. UNDERWOOD. G., CLEWS, S., and EVERATT, J . (1990). How d o readers know where t o look next? Local information distributions influence eye fixations. Quarterly Joiirnal of Experiniental PSycltOlOgy. 421%.39-65. WILKINSON, H.P. (1976). \Vide range eye and head movement monitor. Qiiarterly Jorirnalof Experirneiital Psycliology, 28, 123-124.
Downloaded from las.sagepub.com at CAMBRIDGE UNIV LIBRARY on August 22, 2015
