Chapter 2 Literature Review
3.1 Method
3.1.7 Eye movement recording procedure in reading experiments
國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
correctly, the procedure stopped and a score was give by subtracted 0.5 from the number of sentences at the size. For example, if the participant correctly answered two sets of 3 sentences, he/she gained a 2.5 (3 minus 0.5) score. If less than two sets of the same size was correct, the procedure stopped and a score was given by the number of the previous size. For instance, if the participant answered only one or zero set of 3-sentence size, he/she scored 2. We used these scores as estimate of working memory capacity during silent reading. This procedure was to ensure that the participants were in common in terms of the working memory capacity, not from extreme groups of people who had especially good or poor working memory capacity.
3.1.7 Eye movement recording procedure in reading experiments
The participants were tested individually. They were seated in a dimly lit room in front of a PC monitor. The eye movement recording procedure is summarized in Figure 2. Prior to the experiment, participants were tested for their dominant eye. In the beginning of the experiment, the instruction was presented on the screen. The nine-point calibration and validation procedure was followed to determine the correspondence between pupil position and gaze position (Tsai, Lee, Tzeng, Hung, &
Yen, 2004). At the beginning of each trial, participants were asked to fixation on a
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
cross where the sentence would begin. If the eye fixation was on the area of the cross to an acceptable degree, the cross vanished and the sentence was shown. Once participants failed to accurately fixate on the cross, the calibration and validation procedure was performed again. Eight practice trails preceded the 96 experimental sentences. Participants were instructed to read the sentence at their own pace and pressed any button on the response box when they understood the meaning of the sentence. Half of the sentences (half of the practice trials and half of the experimental trials) were followed by a comprehension question. The participants had to judge whether the statement of the question was consistent with the sentence they read. A yes (
是 shi) and a no ( 否 fou) were displayed on the screen’s left and right bottom
corner, each of which corresponds to the left and right button respectively. Feedback indicating whether their answer was correct or wrong was then presented. The whole experiment took about one hour to complete.‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
Figure 2. Eye-tracking experiment procedure
3.2 Data analysis
In order to examine eye fixations in particular meaningful zones, the regions of interest (ROIs) were defined as the regions that will be subjected to analyses. The texts before and after the target sentence which contains the relative clause were excluded from analyses. The target of sentence was divided into Word 1 region (W1:
‧
relativizer (REL) region, head noun (N2) region, main V (V2) region, and main-clause object (N3) region. An additional pre-relativizer (Pre-REL) region enclosed V1 and N1 in one analytic unit. Take SRC (20a) and ORC (20b) for example (with sentences reprinted and modified here), the regions are summarized as (23a) and (23b). Each region of interest included the character space immediately to the left of the first character in the region.(23) Regions of Interest for SRC and ORC:
a. Subject relative clause with reversible argument-verb relation 抱住
b. Object relative clause with reversible argument-verb relation 牧師
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
Fixations durations on the regions of interest were analyzed according to both duration and probability measures. These eye movement indices are divided into 3 groups based on the cognitive processes they reveal: (a) first-pass reading time and probability, (b) go-past reading time and probability, and (c) total viewing time and re-reading rate. Group (a) reveals the initial processing of each regions of interest.
Group (b) manifests the detection of ambiguities, representing readers’ efforts to solve ambiguities in order to comprehend the materials read at the point. Group (c) demonstrates readers’ efforts to integrate information needed for solving the ambiguities. The definition of each measure is listed in (24).
(24) Three groups of eye movement indices:
a. First-pass reading time and probability:
i.
First fixation duration (FFD): the duration of the first fixation in a
region.ii.
First-pass time (FPT): the sum of all the fixations beginning with the
first fixation in a region until the gaze leaves the region, either to the right or left of the boundary. For single-word regions, the first-pass time equals to the commonly addressed gaze duration measure.iii.
Re-fixation rate (RFR): the probability of eye movements that re-fixate
within a region during the first-pass gaze.b. Go-past reading time and probability:
i.
Go-past time (GPT): the sum of all the fixations on a region before the
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
gaze crosses the right of the region. It is also called right-bounded time or quasi-first pass time in previous research. This measure suggests the time required to process the present and the previous material before the reader feels ready to process further materials.
ii.
Regressions-out rate (ROR): the probability of eye movements that
cross a region’s left boundary immediately following a first-pass gaze on the region. First-pass regressions-out rate denotes a storage want or a need for integration in a later stage of sentence processing, suggesting a desire for reanalysis.c. Total viewing time and re-reading rate
i. Total viewing time (TVT): the sum of all the fixation duration in a
region, regardless of the direction of eye movement in the previous or afterwards period.ii.
Rereading rate (RRR): the probability of reading a region again after
the first-pass reading, regardless of the direction of the eye-movement to the region, given that the region is not skipped in the first-pass reading.The subject’s data was excluded if the comprehension accuracy was below 75%.
In all of the first-pass duration measures, any duration time less than 80 milliseconds (msec) was excluded from analyses. Except for the pre-relativizer region, any duration time in a region lasting more than 800 msec was also excluded. If a TVT of a
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
trial in a region fell behind 80 msec or exceeded 2000 msec, the data was eliminated.
If a GPT of a trial in a region was less than 80 msec or more than 3000 msec, it was also excluded from analyses. Moreover, if a blink occurs before a fixation in an ROI, the data of the trial was also excluded from calculation of that ROI. Both of the beginning and the ending fixation of each trial were also eliminated. Any fixation located outside the ROIs was subtracted from analyses. Less than 2 percent of trials were excluded because of blinking. An additional 0.77% of trials were eliminated because they involved the fixation beginning or ending the trial or the fixation out of the ROI range. In total, 5.37% of the data were removed because blinking occurred, the data occurred at the beginning or end of the trial, or the fixation was out of the ROI range.
Each of the dependence measures in each ROIs were analyzed in 2 (plausibility:
reversible vs. irreversible) × 2 (RC type: SRC vs. ORC) repeated measures analyses of variance (ANOVAs). Two analyses of variance were computed, one with participants (F1) and the other with items (F2) as a random variable.
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
3.3 Results
3.3.1 Reading span score and comprehension accuracy
The performance of the participants’ reading span test averaged 2.63 (SD = 0.63), which was a medium working memory span. The results of eye movement dependent measures were displayed in the sequence of first-pass reading time and probability, go-past reading time and probability, general index, and regress-in indices. The participants made correct answer to 90.76% (SD = 5.79) of the comprehension questions. The accuracy of answers to comprehension questions based on fillers was up to 94.87% (SD = 4.7), which is higher than the accuracy for each condition. With the reversible argument-verb relation, the accuracy of SRCs and ORCs is 84.38% (SD
= 21.88) and 82.29% (SD = 20.16), respectively. With the irreversible argument-verb relation, the accuracy of SRCs and ORCs are both higher than the former two, 93.75% (SD = 11.06) and 85.42% (SD = 22.56), respectively.
The accuracy for the four conditions were subjected to repeated measure two-way ANOVAs. The analyses produced only a marginal effect of plausibility [F (1, 23) = 3.96, p = .059], suggesting that the irreversible relation tends to be less difficult than the reversible relation. There were no RC type effect [F (1, 23) = 2.145, p = .157],
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
though numerically more SRCs were answered accurately than ORCs with both reversible and irreversible relations. The interaction between plausibility and RC type was not found either (F<1).
3.3.2 Eye movement indices
3.3.2.1 First-pass reading time and probability
The mean and standard errors of first-pass reading time and probability are shown in Table 5—Table 7 and Figure 3—Figure 5. The statistical results are summarized in Table 8. All first-pass indices showed a main effect of RC type at the W2 region, indicating greater processing effort in the first-pass reading for ORCs than SRCs on the second word of the relative clause. However, at the W1 region was a RC type main effect denoting greater efforts for SRCs rather than ORCs on the FPT and RFR measures. At the pre-relativizer, there were main effects of not only RC type but also plausibility in the by-participant analysis, but not in the by-item analysis. The results indicated that FPTs were longer for SRCs than ORCs and also longer for irreversible conditions than reversible ones. Moreover, a marginal plausibility effect was found at W2 in the by-item analysis of FFD. This result also suggested the
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
disadvantage of processing irreversible relations before meeting the relativizer.
Furthermore, there were marginal interactions between plausibility and RC type at W2 in the by-participant analyses of FFD and FPT. Tests of simple main effects showed that there was a robust simple main effect showing that ORCs with irreversible argument-verb relation require much longer first-pass time than the corresponding SRCs [FFD: F(1,23)=9.013, p <.01**; FPT: F(1,23)=14.785, p
<.001***]; however, the SRCs and ORCs with the reversible relation did not differ from each other [FFD: F<1; FPT: F(1,23) = 2.636, p = .118]. These interactions were consistent with the main effect of plausibility at W2 on FFD and the one at the pre-relativizer on FPT. All of them suggested the verb denoting irreversible argument-verb relation calls for longer first-pass time than the verb involved in a reversible relation.
At the relativizer region, no effect or interaction was found. At the head noun, only an effect of plausibility was found, but only in the by-participant analysis of RFR.
It suggested more re-fixations on the head noun for reversible pairs than irreversible ones. This result was expected since both arguments in a reversible relation can be a plausible agent of the verb and this confusable relation consumes greater processing effort most possibly when readers encounter the head noun.
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
At the main verb, there was an RC effect denoting longer first-pass processing time for ORCs than SRCs merely in the by-item analysis of FFD and FPT, indicating ORC difficulty in the first-pass processing of the main verb. At the N3 region, only a very marginal effect of plausibility was found in the by-participant analysis of FPT, also indicating the disadvantage of reversible argument-verb relation.
Table 5. Experiment 1: Mean and standard error (in parenthesis) of first fixation duration (FFD) for the four conditions on all the ROIs (msec)
Figure 3. Experiment 1: First fixation duration (FFD) by regions for the four conditions (msec) 200
210 220 230 240 250 260 270 280 290 300
W1 W2 DE head N main V N3
rev-SRC rev-ORC irrev-SRC irrev-ORC
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
Table 6. Experiment 1: Mean and standard error (in parenthesis) of first first-pass time (FPT) for the four conditions on all the ROIs (msec)
Figure 4. Experiment 1: First-pass time (FPT) by regions for the four conditions (msec)
Table 7. Experiment 1: Mean and standard error (in parenthesis) of re-fixation rate (RFR) for the four conditions on all the ROIs (%)
200 220 240 260 280 300 320 340 360 380
W1 W2 DE head N main V N3
rev-SRC rev-ORC irrev-SRC irrev-ORC
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
Figure 5. Experiment 1: Re-fixation rate (RFR) by regions for the four conditions (%)
Table 8. Experiment 1: Statistic results of first-pass reading time and probability (FFD: first-fixation duration, FPT: first-pass time, RFR: re-fixation rate) on each ROIs in the by-participant analyses (F1) and by-item analyses (F2) with repeated measures two-way ANOVAs testing the effect of plausibility (plau), RC type (RC), and their interaction (plau*RC) (p < .05*, p < .01**, p < .001***)
3.3.2.2 Go-past reading time and probability
As for the go-past reading time and probability, the results were more stable and
0%
5%
10%
15%
20%
25%
30%
35%
40%
W1 W2 DE head N main V N3
rev-SRC rev-ORC irrev-SRC irrev-ORC
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
reliable than those of the first-pass reading time and probability. Table 9—Table 10 and Figure 6—Figure 8 display their means and standard errors. Table 11 shows the statistic results of the go-past reading time and probability. This group of indices reveals the possible disambiguating position in a sentence.
At the W2 region, there was a robust effect of RC type in both analyses of ROR, indicating that readers much more frequently regress immediately after the first-pass gaze when reading SRCs than ORCs. A marginal effect of RC type on ROR also occurred at the relativizer DE, also showing that readers regress right after they meet the relativizer more often for SRCs than ORCs. The GPT analyses also showed a significant effect of RC type at W2, suggesting that readers require longer processing time before they get ready to read further materials following W2.
A marginal effect of RC type that denotes SRC difficulty was found in the earlier region, W1, as well in both analyses of GPT. When W1 and W2 were combined in analyses, the effect of RC type indicating SRC difficulty was also found on GPT, but only in the by-participant analysis. In addition, a less significant effect of SRC difficulty was found at the relativizer in the by-item analysis. To sum up the findings above, it seems that the initiating V in the SRC may confuse the readers, and later the V plus N sequence disambiguated this clause by alerting readers that there was a
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
missing agent, or a gap. Therefore, readers regressed frequently, as frequently as 41%
of the time right after they encountered the second word, the N, in the SRCs. They also spent longer time lingering in the contents before the relativizer, especially before the head noun. Later at the single-word relativizer DE, readers reduced the frequency of regression in general, but the regressions made for SRCs were still numerically more than for ORCs.
At the head noun, RC type effect was missing. Instead, the effect of plausibility was robust in both F1 and F2 analyses of GPT and ROR. This plausibility effect denotes greater integration effort for the reversible relation than the irreversible relation. This effect was even more robust at the main verb, where the relative clause is to be integrated into the main clause. Besides, the RC type effect was found at the main verb with an opposite pattern. The RORs for ORCs increased a great deal at the main verb while the RORs for SRCs decreased at the main verb. As a result, readers regressed much more often for ORCs than SRCs at the main verb. Readers also needed more GPTs for the integration of the main verb with the preceding materials when reading ORCs, compared with SRCs. At the N3 region, no effect or interaction was found, which suggested that there was no difference among conditions. The RORs for all the four conditions dropped radically at the N3, indicating that the
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
possible ambiguities had been extensively solved before the N3.
Table 9. Experiment 1: Mean and standard error (in parenthesis) of go-past time (GPT) for the four conditions on all the ROIs (msec)
Figure 6. Experiment 1: Go-past time (GPT) by regions for the four conditions (msec) 200
250 300 350 400 450 500 550 600 650 700
W1 W2 DE head N main V N3
rev-SRC rev-ORC irrev-SRC irrev-ORC
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
Table 10. Experiment 1: Mean and standard error (in parenthesis) of regressions-out rate (ROR) for the four conditions on all the ROIs (%)
Figure 7. Experiment 1: Regressions-out rate (ROR) by regions for the four conditions (%)
Table 11. Experiment 1: Statistic results of go-past reading time and probability (GPT: go-past time, ROR: regressions-out rate) on each ROIs in the by-participant analyses (F1) and by-item analyses (F2) with repeated measures two-way ANOVAs testing the effect of plausibility (plau), RC type (RC), and their interaction (plau*RC) (p < .05*, p < .01**, p < .001***)
0%
5%
10%
15%
20%
25%
30%
35%
40%
45%
50%
W1 W2 DE head N main V N3
rev-SRC rev-ORC irrev-SRC irrev-ORC
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
3.3.2.3 Total viewing time and re-reading rate
The means and standard errors of total viewing time (TVT) are presented in Table 12 and Figure 8. The statistic results of this measure are shown in Table 14.
The analyses of TVT displayed a pervasive benefit of the irreversible argument-verb relation. Among these plausibility effects, the ones at the head noun and the relativizer were robust. Readers seem to enjoy a greater ease from the irreversible relation than the reversible one.
The plausibility effect at the W1 was significant in the by-participant analysis, showing a need for greater efforts for reversible conditions about the information of W1. However, the effects of RC type found at W1 and the main verb indicated different patterns. The one at W1 denoted SRC difficulty while the one at the main verb indicated ORC difficulty. This pattern echoed the results of the go-past reading time and probability, further supporting the hypothesis that readers face SRC difficulty earlier before the relativizer and then suffer from ORC difficulty later at the head noun or the main verb.
The last measure is re-reading rate (RRR). Table 13 and Figure 9 exhibit the means and standard errors of each measure by region and Table 14 summarize the statistic results of these measures. This measure unveils the processes of re-analyses
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
by showing which region(s) is re-read again.
There were robust effects of RC type at W1 in the analyses of RRR, indicating that the W1 in SRCs was more frequently revisited than ORCs. The RC type effects were found at the relativizer in both analyses of RRR and at the main verb in the by-participant analysis of RRR. When reading ORCs, readers seem to re-read the relativizer and the main verb more frequently than reading SRCs. The effects of plausibility showing more re-processing were prevalently found from W1 to the main verb. At W1 and W2, the plausibility effects were significant in the by-participant analysis of RRR.
At the pre-relativizer, where W1 and W2 were combined, the effect was significant in both analyses of RRR. At the relativizer, the plausibility effect was significant on RRR. At the head noun, the analyses of RRR showed robust significance of plausibility effects. At the main verb, the plausibility effect was significant only in the analyses of RRR. These effects of plausibility suggested that the reversible argument-verb relation elicit more re-reading in most of the ROIs.
The interaction between plausibility and RC type was marginally significant (p
=.07) at the head noun in the by-participant analysis of RRR. The simple main effect
tests showed that the frequency of re-reading ORCs significantly exceeded that of
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
SRCs in the reversible condition [F(1,24) = 5.932, p = .023], while the irreversible pairs did not differ significantly [F<1]. At this region, there was also a robust effect of plausibility, suggesting processing difficulty of the reversible argument-verb relation.
This interaction suggested the semantic plausibility of argument-verb relation modulates the processing of RCs. The irreversible relation eases the processing difficulty of ORCs and reduced the re-reading rate of ORCs with an irreversible relation at the head noun.
Table 12. Experiment 1: Mean and standard error (in parenthesis) of total viewing time (TVT) for the four conditions on all the ROIs (msec)
‧
Figure 8. Experiment 1: Total viewing time (TVT) by regions for the four conditions (msec)
Table 13. Experiment 1: Mean and standard error (in parenthesis) of re-reading rate (RRR) for the four conditions on all the ROIs (%)
Figure 9. Experiment 1: Re-reading rate (RRR) by regions for the four conditions (%) 200
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
Table 14. Experiment 1: Statistic results of total viewing time (TVT) and re-reading rate (RRR) on each ROIs in the by-participant analyses (F1) and by-item analyses (F2) with repeated measures two-way ANOVAs testing the effect of plausibility (plau), RC type (RC), and their interaction (plau*RC) (p
< .05*, p < .01**, p < .001***)
3.3.2.4 Supplementary analyses of V1 and N1
In the previous analyses, the W1 and W2 regions were defined by pooling V1 and N1 together in a way to show the significances resulted from incremental processing. However, this definition of ROIs confounded the word class factor with the RC type factor, which co-varies with the sequence of V1 and N1. Therefore, the data were further grouped into V1 and N1 for inspection of the possible influence of word class, namely, the difference between verbs versus noun. Table 15—Table 18 displays the means and standard errors on each measures in terms of three factors:
word class (V1 vs. N1), plausibility (reversible vs. irreversible) and RC type (SRC vs.
ORC). The data was subjected to the analyses with repeated measures three-way
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
ANOVAs according to these three factors. The findings indicated that there was no three-way interaction (word class × plausibility × RC type) on all the measures. The two-way interaction of word class × RC type were significant on every measure except for RFR. All these interactions simply revealed the effect of sequence; that is,
ANOVAs according to these three factors. The findings indicated that there was no three-way interaction (word class × plausibility × RC type) on all the measures. The two-way interaction of word class × RC type were significant on every measure except for RFR. All these interactions simply revealed the effect of sequence; that is,