Eyemind.psych.ac.cn

Psychonomic Bulletin & Review2005, 12 (6), 1089-1093 The effect of word predictability on the KEITH RAYNER, XINGSHAN LI, and BARBARA J. JUHASZ University of Massachusetts, Amherst, Massachusetts Tianjin Normal University, Tianjin, China Eye movements of Chinese readers were monitored as they read sentences containing target words whose predictability from the preceding context was high, medium, or low. Readers fixated for less time on high- and medium-predictable target words than on low-predictable target words. They were also more likely to fixate on low-predictable target words than on high- or medium-predictable target words. The results were highly similar to those of a study by Rayner and Well (1996) with English read-ers and demonstrate that Chinese readers, like readers of English, exploit target word predictability during reading.
Although a large proportion of the world’s population most words are made up of two characters, although some consists of readers of Chinese, much less is known about words consist of only one character and some consist of eye movements that occur during reading of this logo- three or more characters. Finally, Chinese words, like graphic script than about those that occur during reading English words, presumably vary in terms of how predict- of alphabetic writing systems (particularly English). Text able they are from the preceding context. To what extent written in Chinese is formed by strings of equally spaced do character complexity, character frequency, word fre- boxlike symbols called characters. Historically, Chinese quency, and word predictability influence eye movements text was printed from top to bottom (with the columns in the reading of Chinese? This question is interesting in printed from right to left). However, like English, Chi- the context of recent models of eye movement control, nese is now most typically printed horizontally from left such as E-Z Reader (Reichle, Pollatsek, Fisher, & Rayner, to right. Unlike English (and other alphabetic writing sys- 1998; Reichle, Rayner, & Pollatsek, 2003) and SWIFT tems), Chinese is written without spaces between succes- (Engbert, Longtin, & Kliegl, 2002; Kliegl & Engbert, sive characters and words. Furthermore, individual char- 2003), which do a good job of simulating the eye move- acters differ greatly in terms of complexity because they ment behavior of readers of alphabetic writing systems. vary in (1) the number of strokes per character, (2) the These models typically take as input information about number of radicals (or certain combinations of strokes that the frequency and predictability of the words in a text to denote semantic or phonological information), and (3) the simulate eye movement behavior in reading. In the case of manner of construction (i.e., radicals can be combined in Chinese, character complexity, character frequency, word different ways to form compound words). Basically, there frequency, and word predictability could all be important are many visual details packed into a constant, box-shaped factors influencing eye movements. Thus, it is important to understand the eye movement characteristics of Chinese Characters differ in complexity, as well as in the fre- readers, in order not only to understand Chinese reading quency with which they are seen. And although the con- per se, but also to determine the extent to which the mod- cept of a word is not so clearly defined in Chinese as it is els that have been developed might be able to account for in English (so that Chinese readers will disagree some- what concerning where word boundaries are located), it What is known about the eye movements of Chinese is the case that Chinese words also differ in frequency. readers? First, average fixation durations tend to be very The Chinese characters are more like morphemes, and similar (about 225–250 msec) for readers of Chinese and English (Chen, Song, Lau, Wong, & Tang, 2003; Rayner, 1998; Sun & Feng, 1999). Second, not surprisingly, av-erage saccades are much shorter in Chinese (about 2.6 This research was supported by Grant HD26765 from NIH and a Grant characters) than in English (about 7–8 letters), since the from Microsoft Corporation. We thank David Balota, Max Coltheart, information is more densely packed in Chinese (Chen Reinhold Kliegl, and an anonymous reviewer for helpful comments on an et al., 2003). Third, regression rate appears to be slightly earlier draft. Correspondence concerning this article should be addressed higher in Chinese (about 15%) than in English (about to K. Rayner, Department of Psychology, University of Massachusetts, Amherst, MA 01003 (e-mail: rayner@psych.umass.edu).
10%) skilled readers (Chen et al., 2003; Rayner, 1998). Copyright 2005 Psychonomic Society, Inc.
Fourth, the probability of skipping a word tends to be have weaker effects for Chinese readers than for English higher in Chinese than in English (42% vs. 20%), accord- readers. Finally, if reading processes are similar for Chi- ing to Chen et al., although a recent study by Tsai, Lee, nese and English readers, we would expect that predict-Tzeng, Hung, and Yen (2004) reported skipping rates of ability would have similar effects across languages.
around 10% for Chinese readers. Finally, the perceptual span of Chinese readers extends 1 character to the left of fixation to 2–3 characters to the right when they are read-ing from left to right (Inhoff & Liu, 1997, 1998; see also Participants
Chen & Tang, 1998);2 in contrast, for English, the span Sixteen native Chinese speakers who were students at the Uni- versity of Massachusetts were paid to participate in the experiment. extends 3–4 letters to the left of fixation to about 14–15 All of them had normal or corrected-to-normal vision, and all were letters to the right of fixation (Rayner, 1998).
naive regarding the purpose of the experiment.
With respect to the four different character/word vari- ables that could influence reading behavior, there is only Apparatus
scant data on Chinese. For English, it is well known that Eye movements were recorded by an SR EyeLink II eyetracker, word frequency and word predictability have strong in- which has a resolution of approximately 30′ of arc. The participants fluences on the fixation time on a word (Rayner, 1998), read the target sentences (which were printed horizontally from left to right) on a 19-in. NEC Trinitron monitor connected to a 166-MHz and word predictability influences word skipping (Brys- Pentium PC. The participants wore a lightweight helmet that is part baert & Vitu, 1998; Rayner, 1998).3 For Chinese, it has of the eye-tracking system. The eye-tracking system samples at the been demonstrated that character complexity and word rate of 500 Hz and provides eye movement data for further analysis frequency influence fixation time on a word and also skip- via another 166-MHz Pentium PC. Although the Eyelink II system ping (Yang & McConkie, 1999). Chen et al. (2003) noted is able to compensate for head movements, the participants did rest that regression analyses that they carried out on Chinese their heads in a chinrest to minimize head movements during experi-mental trials. Viewing was binocular, but eye movement data were adults’ eye movement data indicated that character com- collected only from the right eye. The participants were seated 75 cm plexity and frequency were more important than word from the video monitor; at this distance, one character subtended frequency. They also reported regression analyses with .45º of visual angle.
children (second, fourth, and sixth graders) that showed
basically the same thing. Clearly, more work is needed Materials
to examine more precisely the extent to which character
Two native Chinese-speaking readers first translated the Rayner and complexity, character frequency, and word frequency in- Well (1996) materials into Chinese. Those that were not easily trans- fluence the eye movements of Chinese readers.
lated were dropped, and a number of other sentences were developed. A total of 50 sentences were then rated by 24 native Chinese-speaking In the present study, we focus on the effect of word students at the University of Massachusetts (none of whom partici- predictability on Chinese readers’ eye movements. Sur- pated in the main experiment), by 16 Chinese students at Tianjin prisingly, there appears to be no prior work on this vari- Normal University, and by an additional 14 native Chinese speakers able, which, as was noted above, has been shown to have from mainland China. The rating procedure provided the sentence a robust effect on the eye movements of readers of En- up to where the target word should be, and the participants filled in glish (Ehrlich & Rayner, 1981; Rayner & Well, 1996). the next word that they thought would come in the sentence. Using the resulting norming data, we ended up with 36 sentences, and each The study we report here is quite similar to that of Rayner sentence had two possible target words: A given sentence could and Well. We constructed sentences that contained target have a high- or a medium-predictable target word, a high- or a low- words that varied in terms of their having high, medium, predictable target word, or a medium- or a low-predictable target or low predictability from the prior context. Readers’ eye word. Via this procedure, there were 24 high-, 24 medium-, and 24 movements were then recorded as they read these sen- low-predictable target words. The mean predictability values (and tences. Would the same pattern that was obtained in Eng- the ranges) are quite comparable to those used by Rayner and Well lish emerge in Chinese, or would the differences in writing (see Table 1). Indeed, 16 of the 36 sentences used in the present study were direct translations of the sentences used by Rayner and systems render the effect of predictability more potent or Well. Counterbalancing procedures ensured that each participant saw an equal number of words in the three conditions, and no sen- Theoretically, it seems a priori possible that the reading tences were repeated for a given participant.
process could be different for readers of Chinese than for Because we were interested in examining the effect of predict- readers of alphabetic writing systems such as English. On ability independently of word frequency, character frequency, and the one hand, the lack of spacing between words (and the character complexity, we controlled these three variables as much as possible. Of the 72 target words, 2 were 1 character, 67 were 2 fact that Chinese readers often disagree on where word characters, and 3 were 3 characters. The frequency of the first char- boundaries are) might lead them to rely more heavily on acter (based on a dictionary count of 1,568,608 characters; National contextual information than do readers of English. In this Languages Committee, 1997) averaged 1,328 (range, 4–6,066; SD = case, we would expect contextual constraint/predictability 1,463) and did not vary across predictability conditions ( ps > .25); to have stronger effects in Chinese than is the case in English. likewise, the frequency of the second character averaged 1,451 On the other hand, because more information (relatively (range, 2–13,167; SD = 2,066) and did not vary across predictabil- ity conditions ( ps > .27). Character complexity, as defined by the speaking) falls within the foveal region in Chinese than in number of strokes per character, averaged 7.74 (range, 3–16; SD = English, Chinese readers might not exploit word predict- 3.2) for the first character and 7.23 (range, 2–17; SD = 3.1) for the ability to the extent that readers of English do. In this case, second character and did not differ across predictability conditions we would expect contextual constraint/predictability to ( ps > .38). Finally, the overall word frequency of the target words Fixation Time Measures (in Milliseconds) and
Fixation Probability on the Target Word as a Function of Predictability
Note—The values in parentheses are the data when five high-frequency words were eliminated from the analyses. The values aligned with R&W represent the means from Rayner and Well (1996). For the Rayner and Well data, there was more variabil-ity in the word frequency of the target words, but they averaged 58 per million in the Francis and Kuˇcera (1982) norms. FFD, first-fixation duration; Gaze, gaze duration; TFD, total fixation duration; PF, probability of first-pass fixation; PV, probability value (values in parentheses represent the range).
averaged 59 (range, 1–874; SD = 146)4 and did not differ across the ence between high- and low-predictable targets was sig- three levels of predictability ( ps > .09). The mean frequencies of nificant [t (15) = 2.35, p < .05; t (46) = 3.03, p < .01], as the target words were comparable to those used by Rayner and Well was the difference between medium- and low-predictable targets [t (15) = 2.41, p < .05; t (46) = 2.24, p < .05]. For total time, the predictability effect was again significant Procedure
When a participant arrived for the experiment, the eye-tracking [F (2,30) = 5.0, p < .05; F (2,69) = 3.79, p < .05]. In the system was calibrated. The calibration generally lasted less than t tests, there were significant differences for high versus 5 min. After the calibration had been completed, the participant medium predictability [t (15) = 2.08, p = .055; t (46) = read the 36 sentences in a different random order, but with appropri- 1.74, p = .089] and for high versus low predictability ate counterbalancing procedures to ensure that an equal number of [t (15) = 3.07, p < .01; t (46) = 1.94, p < .01]. The com- each type of target word was read. The participant was told that the parisons for medium versus low predictability were not purpose of the experiment was to determine where people look as they read. He or she was also told that he or she would periodically significant [t (15) = 1.1, p = .29; t (46) = 1.02, p = .30].
be asked to answer comprehension questions about the sentences. These questions were asked after 25% of the 36 sentences that were Fixation Probability
read; the participants were correct over 90% of the time.
The probability of a first-pass fixation on the target word indicated that the readers were more likely to skip high- and medium-predictable target words than to skip low-predictable target words. An ANOVA yielded a sig- A number of eye movement measures were examined nificant effect of predictability [F (2,30) = 5.4, p < .05; with respect to the target word. Specifically, the measures F (2,69) = 4.6, p < .05]. In the t tests, there were sig- were (1) first-fixation duration (the duration of the first nificant differences for high versus low predictability fixation on a word independent of the number of fixa- [t (15) = 3.42, p < .01; t (46) = 2.21, p < .01] and for tions on the word), (2) gaze duration (the sum of all fixa- medium versus low predictability [t (15) = 2.79, p < .05; tions on a word prior to the reader’s moving to another t (46) = 2.38, p < .05].
word), (3) total fixation time (the sum of all fixations on a word, including regressions), and (4) the probability that Additional Analyses
the reader would fixate on the target word. Table 1 shows Although there were no significant differences across the three fixation time measures. An ANOVA was carried the three predictability conditions in terms of word fre- out on each of the sets of data, using participants (F ) and quency, it was the case that there were numerical differ- items (F ) as random effects, and was followed up with ences, owing to the fact that there were some very high frequency words in the high and medium conditions. We therefore excluded five target words (three from the Fixation Time
high and two from the medium conditions) with very For first-fixation duration, the predictability effect high frequency counts and redid the ANOVAs. Exclud- was not significant by participants [F (2,30) = 1.26, p > ing these high-frequency words resulted in mean frequen- .29] but was marginally significant by items [F (2,69) = cies of 27, 26, and 20 per million for the high-, medium-, 3.064, p = .053]. In the t tests, by items, the difference and low-predictable conditions, respectively ( ps > .50). between the high- and the low-predictable target words The means using these more restricted frequencies are was significant [t (46) = 2.29, p < .05]. For gaze duration, shown in Table 1; obviously, the exclusion of these high- the predictability effect was significant [F (2,30) = 4.79, frequency items did not change the pattern of results at all. p < .05; F (2,69) = 5.55, p < .01]. In the t tests, the differ- Indeed, none of the mean values for first-fixation dura- tion or gaze duration changed by more than 4 msec as a But this is clearly due to the fact that informational density result of this reanalysis. Exactly the same pattern of sta- is much higher in Chinese than in English. Indeed, when tistical results was present in this analysis as in the origi- reading rate is computed so that words per minute is the nal analysis [gaze duration, F (2,30) = 3.67, p < .05, and measure (rather than characters per minute), the reading F (2,64) = 4.24, p < .05; total time, F (2,30) = 5.5, p < rates of Chinese and English readers are quite comparable .01, and F (2,64) = 3.02, p = .056; fixation probability, (Rayner & Pollatsek, 1989; Sun, Morita, & Stark, 1985).
F (2,30) = 3.6, p < .05, and F (2,64) = 3.51, p < .05].
The present study clearly shows that Chinese read- ers exploit contextual constraint/predictability factors in Comparison With Rayner and Well (1996)
much the same way as English readers do. Other recent Table 1 also shows the comparable data from Rayner studies (Liu, Inhoff, Ye, & Wu, 2002; Pollatsek, Tan, & and Well (1996). Whereas the fixation time measures Rayner, 2000; Tsai et al., 2004) have demonstrated that line up reasonably well across the two studies (although orthographic and phonological codes are used by Chinese the fixation times are consistently longer for the Chinese readers to integrate information across saccades, just as readers), so that the data patterns are similar, the fixation they are in English (Pollatsek, Lesch, Morris, & Rayner, probability data are slightly different. That is, for the En- 1992). As we noted above, we observed a difference in glish readers, there was no difference in fixation probabil- the skipping patterns between the two groups of readers, ity between medium- and low-predictable target words; but this may be explainable in terms of the fact that the English readers were much more likely to skip over a high- characters are closer to fixation prior to a skip in Chinese predictable word than over a medium- or low-predictable than in English. Yang and McConkie (1999) reported that word. The Chinese readers, on the other hand, were more unlike readers of English, who show a clear landing posi- likely to skip a high- or medium-predictable word than tion effect, wherein the eyes tend to land about halfway to skip a low-predictable word. Thus, the data pattern for between the beginning and the middle of a word (the pre- the Chinese readers was the same for the fixation time ferred viewing location; Rayner, 1979), Chinese readers measures and the fixation probability measures. Why this do not show such an effect. Both the difference in skipping difference in pattern between the Chinese and the English and the lack of a preferred viewing position effect may readers emerged is not at all apparent, but it may have be due to the fact that the next character to be fixated is to do with the fact that words that are skipped are much much closer to the current eye position in Chinese than closer to the fixation point in Chinese. Finally, the total in English.
fixation durations were markedly longer for the Chinese In summary, although there are some clear differences readers. This reflects the fact, noted earlier, that Chinese between Chinese and English readers in terms of initial readers tend to regress more frequently than do readers encoding of print (due to the nature of the logographic of English.5 vs. alphabetic writing systems), once the material is en- Although it is difficult to make cross-experiment com- coded, reading processes appear to be more similar than parisons, it is the case that the predictability values of dissimilar for the two groups of readers. We noted at the the target words were very similar across the studies, the outset that one goal in determining how word predictabil- frequencies of the target words were quite similar, and ity influences the eye movements of Chinese readers was the reading skill of the participants was quite similar. to determine whether models such as E-Z Reader (Reichle Although it remains quite likely that there are important et al., 1998; Reichle et al., 2003) could read Chinese. cross-cultural differences, the extent to which predictabil- Clearly, more data need to be collected regarding charac- ity is exploited in reading seems similar across the two ter complexity, character frequency, and word frequency languages.
effects in Chinese. However, the present results (as well as the results that do exist on these other three variables) DISCUSSION
suggest that it might well be the case that the model would be effective with Chinese.
As we noted at the outset, data on the eye movements of Chinese readers is rather scant. It is sometimes suggested REFERENCES
that there are differences in the way Chinese and English readers read (due to major differences in the nature of the theories of eye movement control in reading. In G. Underwood (Ed.), orthographies). For example, there is some controversy Eye guidance in reading and scene perception (pp. 125-147). Amster- regarding the relative importance of orthographic versus phonological information in the initial access of Chinese Chen, H.-C., Song, H., Lau, W. Y., Wong, K. F. E., & Tang, S. L. versus English words (Feng, Miller, Shu, & Zhang, 2001; (2003). Developmental characteristics of eye movements in reading Chinese. In C. McBride-Chang & H.-C. Chen (Eds.), Reading devel- Rayner, Pollatsek, & Binder, 1998; Wong & Chen, 1999). opment in Chinese children (pp. 157-169). Westport, CT: Praeger.
Yet the more we learn about the eye movements of Chinese Chen, H.-C., & Tang, C.-K. (1998). The effective visual field in Chi- readers, the more apparent it becomes that there are more similarities than differences. Obviously, Chinese readers Dictionary of Chinese character information (1988). Beijing: Science make shorter saccades than do English readers, and the Ehrlich, S. F., & Rayner, K. (1981). Contextual effects on word per- perceptual span (or area of effective vision in reading) of Chinese readers is smaller than that of English readers. Engbert, R., Longtin, A., & Kliegl, R. (2002). A dynamical model ing Chinese script: A cognitive analysis (pp. 189-206). Mahwah, NJ: of saccade generation in reading based on spatially distributed lexical Sun, F., Morita, M., & Stark, L. W. (1985). Comparative patterns of Feng, G., Miller, K., Shu, H., & Zhang, H. (2001). Rowed to recov- reading eye movements in Chinese and English. ery: The use of phonological and orthographic information in reading Tsai, J.-L., Lee, C.-Y., Tzeng, O. J. L., Hung, D. L., & Yen, N.-S. (2004). Use of phonological codes from Chinese characters: Evidence Francis, W. N., & Kuˇcera, H. (1982). Frequency analysis of English from processing of parafoveal preview when reading sentences. usage: Lexicon and grammar. Boston: Houghton Mifflin.
Inhoff, A. W., & Liu, W. (1997). The perceptual span during the read- Vitu, F. (1991). The influence of parafoveal processing and linguistic ing of Chinese text. In H.-C. Chen (Ed.), The cognitive processing of context on the optimal landing position effect. Chinese and related Asian languages. Hong Kong: Chinese Univer- Wong, K. F. E., & Chen, H.-C. (1999). Orthographic and phonologi- Inhoff, A. W., & Liu, W. (1998). The perceptual span and oculomotor cal processing in reading Chinese text: Evidence from eye fixations. activity during the reading of Chinese sentences. Yang, H.-M., & McConkie, G. W. (1999). Reading Chinese: Some Kliegl, R., & Engbert, R. (2003). SWIFT explorations. In J. Hyönä, basic eye-movement characteristics. In J. Wang, A. W. Inhoff, & H.-C. R. Radach, & H. Deubel (Eds.), The mind’s eye: Cognitive and applied Chen (Eds.), Reading Chinese script: A cognitive analysis (pp. 207- aspects of eye movement research (pp. 391-411). Oxford: Elsevier.
Kliegl, R., Grabner, E., Rolfs, M., & Engbert, R. (2004). Length, frequency, and predictability effects of words on eye movements in Liu, W., Inhoff, A.W., Ye, Y., & Wu, C. (2002). Use of parafoveally 1. There are actually two Chinese scripts: the simplified Chinese script used in mainland China and Singapore and the traditional script used in Taiwan and Hong Kong and by Chinese minorities in other countries (such as the U.S. and Canada). The experiment reported here used the National Languages Committee (1997). Chinese Dictionary: Char- simplified script, which is visually less complex and written from left to acter and word frequency statistics report. Taipei: Author. (In Chinese) right and from top to bottom. All of our participants were from mainland Osaka, N. (1993). Asymmetry of the effective visual field in vertical reading as measured with a moving window. In G. d’Ydewalle & 2. Although studies on the perceptual span in reading Chinese have not J. van Rensbergen (Eds.), Perception and cognition: Advances in eye examined vertical reading, studies with Japanese readers (Osaka, 1993) movement research (pp. 275-283). Amsterdam: North-Holland.
have demonstrated that the perceptual span is asymmetric in the direc- Pollatsek, A., Bolozky, S., Well, A. D., & Rayner, K. (1981). tion of reading. Likewise, Pollatsek, Bolozky, Well, and Rayner (1981) Asymmetries in the perceptual span for Israeli readers. demonstrated that the perceptual span for Hebrew readers is also asym- metric to the left of fixation. Together, these studies strongly indicate that Pollatsek, A., Lesch, M., Morris, R. K., & Rayner, K. (1992). Pho- the perceptual span is primarily due to attentional factors (with the span nological codes are used in integrating information across saccades in asymmetric in the direction that the eyes will move next).
3. In other alphabetic languages, such as Dutch, French, and German, word frequency and predictability influence fixation times and skipping Pollatsek, A., Tan, L. H., & Rayner, K. (2000). The role of phonolog- (Brysbaert & Vitu, 1998; Kliegl, Grabner, Rolfs, & Engbert, 2004; Vitu, ical codes in integrating information across saccadic eye movements 4. Our primary source for word frequency was the Chinese Dictionary (National Languages Committee, 1997), in which the word frequency Rayner, K. (1979). Eye guidance in reading: Fixation locations within count is based on a corpus of 1,116,417 words. There were a few words that were not available in this frequency count, but they were available in Rayner, K. (1998). Eye movements in reading and information process- the Dictionary of Chinese Character Information (1988), which is based on a corpus of 1,310,000 words. Because the two counts are based on Rayner, K., & Pollatsek, A. (1989). The psychology of reading. En- slightly different numbers of words, we mathematically equated the two sources to the former frequency count.
Rayner, K., Pollatsek, A., & Binder, K. S. (1998). Phonological 5. Indeed, the overall regression rate to the target word in this study codes and eye movements in reading. Journal of Experimental Psy- averaged 16.3%. Readers regressed more to the medium- (20%) and chology: Human Perception & Performance, 24, 476-497.
low-predictable (17%) target words than to the high-predictable (12%) Rayner, K., & Well, A. D. (1996). Effects of contextual constraint on word [t(15) = 2.51, p < .05, and t(15) = 1.84, p = .086, respectively]. eye movements in reading: A further examination. There was no difference between the medium and the low conditions (t < 1). Although the differences were only marginally significant ( ps < Reichle, E. D., Pollatsek, A., Fisher, D. L., & Rayner, K. (1998). .11), due to high reader variability, regressions out of the target word Toward a model of eye movement control in reading. were more frequent for the low-predictable target word (23%) than for the high or the medium words (17%) in both cases. Both of these factors Reichle, E. D., Rayner, K., & Pollatsek, A. (2003). The E-Z Reader account for the increased total fixation durations.
model of eye-movement control in reading: Comparisons to other Sun, F., & Feng, D. (1999). Eye movements in reading Chinese and English text. In J. Wang, A. W. Inhoff, & H.-C. Chen (Eds.), Read- revision accepted for publication March 9, 2005.)

Source: http://eyemind.psych.ac.cn/PublishedPaper/RaynerLiJuhaszYan2005.pdf