4. RESULTADOS Y DISCUSIÓN
4.7. PRUEBA DE HIPÓTESIS
In an attempt to optimize the effect of visual degradation, a pilot study was run. Bearing in mind the findings regarding the type of visual degradation, care was taken to ensure that the quality of individual letters was affected. Thus, degradation was created by superimposing over the target a random pattern of white moving dots, which created a ’snowy’ effect. The amount of occlusion through the dots was about 40%. The stimulus set was the same as in Experiment 2, but it was reduced by taking out all filler trials (sentences with adjectives). There were 128 trials (plus ten start-up trials), 50% of which had non word targets and 50% word targets. Again, 30% of the word trials were in the positive fusion condition, 30% in the violated fusion condition and 30% in the neutral condition. Since three parallel lists were constructed, each subject saw a particular target word only once.
Results
Results were obtained from 9 subjects. The overall mean RT in the correct word trials was 894 ms (range=384 to 6464). After a cut-off was set at 1433 ms, the mean RT was 835 ms, which is about 200 ms longer than in Experiment 2. The overall error rate, however, was not higher than in Experiment 2: 7.6% out of all trials had an incorrect response (8.1% out of all word trials).
Mean and median RTs based on individual responses rather than subject- or itemmeans are presented below:
pos neutr viol
mean 764 809 807
median 732 772 762
Thus, compared to E2, the difference between the positive and the neutral condition was increased (from 20 ms in E2, to 45 ms in E3), but the neutral and violated
conditions now did not differ any more. Comparing the median RTs of each individual subject at each level of Fusion, there were only two subjects for whom the predicted sequence (pos < neutr < viol) was true. For three more subjects, the median for positive was smaller than that for violated.
It might be that the mask was too effective, requiring subjects to check the visual input very closely letter by letter. Note that half of the nonwords, i.e. 32, were
pseudohomophones. Most of them, 23, differed from real words by only one letter (the other 9 differed by two letters from real words). Even though all 32
nonpseudohomophonic nonwords differed from real words by at least two letters, and thus seemed to be sufficiently distinct from real words, subjects might have adopted a very close spelling checking procedure. One subject reported that he checked every letter of the target before finally making a response. Such a strategy might override the effect of context: if there is detailed visual checking, context cannot exert much of an impact. The finding that the violated condition did not lead to longer responses than the neutral condition might be due to a close visual scanning strategy: if there is no facilitating context, i.e. either a neutral or a violated context, subjects might entirely rely on a close visual scanning strategy, and make a lexical decision based on the output of the visual processor; if that output identifies a target as a word, the fact that it might not fit the context will not be able to overrule the results of the visual
6.3.3. Method
Design, task, procedure and apparatus were the same as in Experiment 2. The only change consisted in presenting the target in visual noise.
Experiment 3 was run with slightly less masking than had been used in the pilot study. The target was again presented in visual noise (superimposed random dot pattern creating snowy effect), but this time the occlusion was around 30%. There were 42 subjects and 118 trials (10 buffer trials, 54 word trials and 54 nonword trials). A subset of the sentences of Experiment 2 were used, again representing the three levels of the factor Fusion (positive, neutral and violated). Three parallel stimulus lists were created, using a Latin Square, to ensure that each subject saw only one member of each sentence triplet.
6.3.4. Results
The overall error rate (false positives and false negatives) was 10%. The percentage of false negatives on the word trials per condition is listed in table 6.4. below:
Table 6.4.: Experiment 3: Percentages of false negative responses per condition positive neutral violated
6.7% 8.7% 11.9%
13.3% of the pseudohomophones and 8.4% of the nonpseudohomophones had false positive responses.
Eleven subjects and eleven items were excluded because more than 20% of the responses (per subject, and per item) were false decisions, or extremely long correct decisions (this was calculated for word trials only). A further three items had to be
dropped because of an error during stimulus presentation. Thus, 11 subjects and 40 items were included in the following analyses. The cut-off for extremely long correct word decisions was set at 1468 ms (i.e. two standard deviations above the mean for all correct word decisions, collapsing over Fusion; one response fell below the lower cut off of 260 ms). The percentages of extremely long correct decision times to word targets is given for each condition in table 6.5. below:
Table 6.5.: Experiment 3: Percentages of extremely long correct decision times per condition
positive neutral violated
3.6% 4.7% 4.6%
The analyses reported below include only those subjects and items which did not have to be excluded, and only times for correct decisions to word trials which fell below the cut-off. Because the number of included subjects differed per List, and the number of included items differed per Itemgroup, data were collapsed. Thus, in both the by subjects and the by-items design there was one repeated measures factor. Fusion, with three levels (positive, neutral, violated). Subject-and itemmeans for each condition were calculated for the logtransformed data and entered into repeated-measures analyses of variance.
The mean RTs for subjects, based on antilogs, are reported below.
Table 6.6.: Experiment 3: Mean response times for correct decisions on word trials per condition
positive neutral violated
735 762 786
analysis [Fl(2,60)=8.60, p=001; F2(2,78)=8.11, p=006; MinF’(2,137)=4.17, p<.025]. Planned comparisons indicated that the difference between the positive and the neutral conditions was significant by-subjects and by-items, but not significant according to M inF [Fl(l,30)=4.33, p=.046; F2(l,39)=8.55, p= 006; M inF(l,60)=2.87, p<.10]. The difference between the violated and the neutral conditions was significant only in the by-subjects analysis [Fl(l,30)=4.78, p=.037], but not in the by-items analysis. The overall priming effect, i.e. the difference between the positive and the violated conditions was also significant [Fl(l,30)=15.85, p<.0005; F2(l,39)=13.76, p=.001; M inF(l,69)=7.37, p<.01].
6.3.5. Discussion
Experiment 3 replicated the results of Experiment 2, except that in this experiment, the difference between the violated and the neutral condition was less reliable (significant only in tlie by-subjects analysis). Visual degradation in this experiment increased the overall correct decision times to word trials by about 200 ms. The number of false negative decisions (i.e. making a No-decision to a target word) is comparable in the two experiments. However, if number of errors and extremely long responses are taken together as a combined error score, more subjects had to be excluded in Experiment 3 than Experiment 2 (26% as compared to 14%). The finding that subjects did not make more false negative decisions in Experiment 3, where the target was visually degraded, indicates that they adopted a very close visual checking strategy. This might also account for the finding that the violated condition differed from the neutral condition only in the by-subjects analysis: as discussed above, violations might be disregarded if close visual analysis led to the identification of a word.
6.4. Discussion of Experim ents 2 and 3
In the following discussion I will first briefly sum up my assumptions regarding the representation of different types of information affecting sentence priming, and the
kinds of processing systems involved. I will then discuss existing models of sentence priming vis-a-vis the findings of Experiments 2 and 3, and will finally present my own model to account for the pattern of results found in Experiment 2 and 3.
6.4.1. Assumptions about types of information and processing systems involved in