CHAPTER 4 RESULTS AND DISCUSSIONS
4.2. Phonological Effect on Lexical Encoding
4.2.3 Generation of Speech Errors and RT in Stroop’s Tasks
國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
87
became less than (a). Homophonous naming test reduced the strength in semantic layer because the literal term and visual color did not belong to the same semantic field. However, the strength of activation in phonological layer retained strong because of high phonological similarity between the dual concepts. The strength in phonological layer as well as in semantic layer might get declined when the dual concepts showed little phonological similarity, as depicted in process (d).
According to above statistic results, process (a) and (b) did not show significant difference in network strength, and process (c) and (d) did not, either. It seems that reduced strength of phonological layer would not be significant, so there was nearly little phonological effect within each visual task. However, if we compared the network in naming or reading task to Stroop naming or homophonous naming task, the strength in network appears significant difference. The effect of phonological strength might get greater when visual competition came up, as in (a), (b), (c), and (d).
The dual activations through visual channel at the same time caused phonological effect to be greater than task 1 and 2, and it also led subjects to produce rather more speech errors in task 3 and 4. It could help us explain why phonological effect induced significant difference in Trial F, Error N, and RT between tasks. The extra strength from more than one input could cause entire activation strength to be more enhanced and confused for subjects than the strength from single input.
4.2.3 Generation of Speech Errors and RT in Stroop’s Tasks
Connectionist model provides a probability for us to explain how retrieving errors came out by means of Stroop technique. Dell & O’Seaghdha’s model (1991) depicted the way which external signals affect the lexical process, as in Figure 2-6. In their model, external signals could come in from any layer of lexical process, such as from semantic, word, and phonological layers at the same time, and relevant nodes in
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
88
each linguist representation were ready to be activated. Even though subjects were asked to focus on specific visual target (color), the literal term also sent external signals into linguistic layers with the visual color information meanwhile. The mental process is supposed to be “busier” because subjects should resist the strength of activation from color term, which is called the “noise” of activation in the lexical process. The coming out of Stroop errors might be the wrongly activated process and encoded to the motor generator. The busyness of the lexical network could also explain why it took much more time for subjects to process trials in task 3 and 4. The external signals from more than one source coming at the same time could explain the longer time that subjects spent in processing Stroop technique.
4.3. Linguistic Effects and Speech Errors: Test 1 ~ Test 4
In this section, we will display and discuss the linguistic effects among the color naming, color reading, Stroop naming, and homophonous naming tests. These effects include initialness effect, rhyme effect, tone effect, phonotactic regularity effect, and Stroop effect. Table 4-10 shows the structure of these linguistic effects, and then the following sections will discuss these effects during lexical process. The total number of speech errors here is 680 (N=680). The following table displays the phonological distribution of these linguistic effects among all target-error pairs we collected in the four tasks. Table 4-11 is the computed result of one-way categorical ANOVA, and its post-hoc analysis will also be shown in table 4.12.
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
89
Table 4-10. The Counts of Linguistic Effects among All Errors (N=680)
Table 4-10 provides the result of counts after we analyzed and categorized for all of the target-error pairs. The data shows the numbers and proportions of these units that the errors share with their targets. Phonotactic effect is an overwhelming effect that all the errors follow in all tests, which means subjects never produced a word whose phonological structure did not exist in their language. Stroop effect seems to be the secondary effect to affect when there were dual visual representations coming up at the same time. We need to put above data into statistic examination for further observation and discussion, as shown in the following.
Linguistic Effects
Task 1 Task 2 Task 3 Task 4 Total
Count % Count % Count % Count % Count %
Initial 38 39.6% 42 53.9% 94 36.6% 75 30.1% 249 40.0%
Rhyme 8 8.3% 14 18.0% 25 9.7% 17 6.8% 64 10.7%
Prenuclear 5 5.2% 9 11.5% 52 20.2% 28 11.2% 94 12.1%
Vowel 0 0.0% 2 2.6% 5 2.0% 22 8.8% 29 3.3%
Coda 14 14.9% 10 12.8% 41 16.0% 31 12.5% 96 14.0%
Structure 36 37.5% 35 44.9% 137 53.3% 110 44.2% 318 45.0%
Tone 43 44.8% 33 42.3% 113 44.0% 103 41.4% 292 43.1%
Phonotactic 96 100.0% 78 100.0% 257 100.0% 249 100.0% 680 100.0%
Stroop No Data No Data 221 86.0% 189 75.9% 410 81.0%
‧
Table 4-11. One-way ANOVA Results of Speech Errors
One-way
Prenuclear 23.50 21.49
Vowel 7.25 10.05
Coda 24.00 14.54
Structure 79.50 51.99
Tone 73.00 40.83
Phonotactic 170.00 96.18
Stroop 205.00 22.63
Total 65.65 12.20
Table 4-12. Post-hoc Analysis for Table 4-11 (Scfeffe)
Post-hoc
Pairs Rhyme Prenuclear Vowel Coda Structure Tone Phonotactic Stroop
Initial .96 .99 .90 .99 1.00 1.00 .18 .12
Note: *, ** are significant at the .05 and .01 levels respectively.
According to the computed result in table 4-11, because the F-ratio (8.22) exceeds the value of F (2.34), we could accept the scientific hypothesis that the counts of these linguistic effects show significant difference [F(8,25)=8.22, p<.01]. It appears that these linguistic effects were found to impact error generation differently. The total mean value among the four tests which one-way ANOVA generated is 65.65, which provides a basic level to justify what kinds of effects influenced and dominated the generation of speech errors. It seems that the mean values of tone, structure,
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
91
phonotactic, and Stroop effects exceed the total mean value. These effects are more prominent than the other linguistic effects, such as the effects of vowel, rhyme, prenuclear glide, coda, and initial. However, the mean value could not help declare certain of these effects exist in generation of speech errors or in lexical encoding.
Since we know these effects appear difference among errors, we need to know how each effect have contract with all the other effects. These effects were compared to see their difference by means of Scfeffe post-hoc test in table 4-12, and we will go over all of these linguistic effects in the following and have discussion on their individual effects in lexical network.
4.3.1. Initialness Effect
According to table 4-10, the total amount of the errors sharing the initial part with the target is 249, and the occurring frequency among speech errors is 40.03%.
Color reading task induced the most errors which share initial part, whose frequency reaches to 53.85%. There were about half of errors in reading test which tended to preserve the initial part, or to retrieve a lexicon which shares the initial. Naming test, Stroop test, and homophonous test shows that the rate of onset-sharing is between 30% and 40%. As to proportion, it seems to be consistent with the findings of Dell (1986) that initial is always detectable and salient in phonological structure. With the account of interactive process model, the salient structure which is already activated in phonological layer also sends feedback to the nodes of lemma layers, and sometimes retrieves the inaccurate lexicon with the same initial, as well as the relevant lexical meaning. However, according to the result of one-way ANOVA in table 4-11 and 4-12, we noticed that not only that the mean of initialness (mean=62.25) did not exceed the total mean 65.65, but it did not show any significant difference with the other effects in the post-hoc result. It means that the errors sharing initial
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
92
with their targets did not achieve to a statistically salient amount. The effect of initialness might not be as apparent as Dell (1986) mentioned in this study.
As to the phonological similarity, we see that speech errors with high phonological similarity still rely greatly on the information of syllable initial. The counts of initialness override the amount of rhyme, vowel, prenuclear glide, and coda individually. It implies that syllable initial is a rather salient structure among all phonological structures in lexical process, which not only helps us retrieve target lexicons correctly, but also retrieve the wrong lexicons with the same onset backwards from phonological layer. However, the issue of whether initial is a facilitative or an interfering effect in lexical will be examined and discussed in the section of advance planning unit. At this phase, we could only assume and infer that initialness would not induce a salient effect, but induce certain amount, when speech errors are generated.
4.3.2. Rhyme Effect
In the four tasks, the proportion of the errors preserving rhyme comes to 10.71%
(N=680, Rhyme Preserving=64). Comparing rhyme effect to initialness effect, the percentage of speech errors with rhyme sharing (10.71%) is far less than the percentage of sharing syllable onset (40.03%). According to the post-hoc result, we found that the count of rhyme only contrast with the amounts in phonotactics (p=.01) and Stroop effect (p=.01), but there is no significant difference between rhyme and any other sub-syllable units, such as initial, vowel, prenuclear glide and tone (p>.05).
Apparently, rhyme effect seems not to impose influence as greatly as initialness.
Rhyme could not be a salient phonological organization which could lead to retrieve a lexicon sharing the same content of rhyme, especially in such a strong lexical network of colors.
The issue of content (segments) in rhyme structure seems not to induce abundant
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
93
rhyme sharing errors. The subordinate contents, prenuclear glide, vowel, and coda, did not show any significant difference in their counts with other possible effects, but they contrasted with the numbers of phonotactics and Stroop effect significantly, as rhyme did. Therefore, we could claim that rhyme, including its subordinate contents, did not show any prominent effect in error generation or lexical encoding. On the other hand, the effect of structure in rhyme will be involved in the section of phonological structure.
4.3.3. Tone Effect
The number of the errors sharing tone with targets is 292 in total, which occupies 43.11% of all. The mean value of tone is 73.00, which exceed the value of total mean 65.65. Tone seems to act more prominent than initial with regard to their numbers and mean values. According to the post-hoc result, the number of tone did not show apparent difference with other linguistic effects, as the same with initialness. It still hang us a vague area for us to judge whether effect play a dominant role in lexical encoding. One thing for certain is that tone effect weighed over the initialness effect in this study. If the tones in Chinese have equal chance to substitute for each other, the chance estimate of tone replacement is supposed to be 25%. Then the percentage of 43.11% seems to imply that tone effect affects and dominates the generation of lexical substitution. We could not deny that this effect exists, but we noticed that tone effect is rather more significant than initial effect and rhyme effect.
The distribution of error counts seems to tell us that tone might be different from the structure of phonemes, such as initial and rhyme. The proportion tells us that initial (53.9%) and rhyme (18%) appear to affect greater on autonomous lexical process (reading task) than the other tasks, but tone effect appears nearly fair among the four tasks, within 41% to 45%. It seems that tone is more than a pure phonological
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
94
structure, and it is supposed to be attributed to a larger framework, such as a phonological organization in lexical structure.
Among the four tests, we could not see any lexical substitution with pure tone substituting, omitting or addition. All of the substitutions mapped to colors, except for one tone-sharing case of “xuang [xuɑŋ35] “ (yellow) which is replaced for “ba [pa35]
(pull)” and could not map to any colors. The exception could not be the evidence of independent status of tone in Chinese. Since the fact that target tone tends to affect the lexical retrieval within the same semantic domain, it seems to support the viewpoint of tone status in the works of Wan & Jaeger (1998) and Wan (2007) that tones are represented lexical underlyingly and ought to be part of the phonological organization of the lexicon, rather than the view of phonological frame acted like stress in English, which was proposed by Chen (1999).
4.3.4. Syllable Structure Effect
The amount of errors sharing phonological structure with targets is 318, which occupies 45% among all errors. The proportion within individual task is from 37.5%
to 53.3%. The mean value of stricture is 79.5, which exceeds the total mean value 65.65. According to the post-hoc result, there is no significant difference in number with other possible effects. It means that syllable structure is not necessary information in lexical encoding, but a salient effect. From the results of proportion and mean value in one-way ANOVA, it still appears that syllable structure effect affects and dominates the error generation and lexical encoding.
In order to explain the generation of errors with the same phonological with targets, as discussed in former section, the selected phonological frame and relevant nodes would send activating weight backwards to the lexicon layer.
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
95
When a color is processed, the lexical nodes with the same phonological frame as well as relevant meaning might get ready to be activated at the same time. Then, the imprecise feedback from phonological layer might have higher probability to retrieve the lexical nodes with the same phonological frame, which causes the lexical substitution to occur more frequently with the same phonological structure.
When we look into the percentage of syllable sharing of these speech errors, we found that the effect of syllable structure affected differently among the four tests. In naming task, 96 errors were generated, there were 47 cases occurred with syllable sharing, which occupied 48.96% in task 1. As to reading task, there were 78 errors in total, and 42 cases were produced to share syllable structure with targets, which took 53.85%. With regard to the Stroop naming task, there were 257 errors in total, and 171 errors, which occupied 66.54%, shared phonological similarity with target color.
In homophonous naming task, there were 249 errors in total, and 143 errors shared phonological structured, which occupied 57.43%. Except for task 1, the other tasks induced more than half errors with the same syllable structure of targets. The computed statistic result of Chi-square test shows that the cases of syllable structure sharing among the four tasks appears significant difference (x2=11.15, df=3, p<.05).
Therefore, we could claim that Stroop naming task seems to induce a stronger effect on syllable structure than all the other tasks, and homophonous naming task follows.
The possible reason might be that Stroop naming involves processing visual color and color term at the same time, and the dual input might lead to a better preparation for phonological framework in this task before lexical activation. Therefore, the error would be easily to be encoded with the same syllable structure. Even though there is no semantic relation between the dual inputs in task 4, phonological structure still affected the generation of speech errors. In addition, in comparison of task 3 and 4, we noticed that the effect was quite stronger when dual inputs were in semantic
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
96
relation, as in task 3.
Generally speaking, syllable structure effect apparently exists in lexical process, especially in the case of Stroop technique.
4.3.5. Phonotactic Regularity Effect
According to the result in table 4-10, none of phonotactic errors were found in this study. It shows that phonotactic regularity is an absolute effect which affects and dominates the lexical process. There were no speech errors which violate the phonological constraint in Mandarin. Dell (1993) proposed that there should be phonological frame for legitimate sound sequence in phonological representation to be checked and encoded. To explain this phenomenon, phonological frame might act like a syllable frame (schema), and only the legal sequence of sounds could be encoded in an available frame. Since there are no illegitimate frames to map, little phonological violation could be encoded to phonetic level and produced. After a lemma is activated in lexical layer, certain phonological frame and relevant phonological nodes in the phonological layer could be activated meanwhile. That could be the reason why there were not any phonotactic errors to be generated in these experiments. A lexical network which is prepared to be activated might let the relevant lexical and phonological nodes get ready before activation. High phonological similarity might lead to a neighboring lexical node to be selected (lexical substitution), but phonological frame could help prevent the activation at lemma level from forming phonotactic errors during phonological encoding.
4.3.6. Stroop Effect
By means of Stroop technique, we could induce the errors involving visual representation. We, tentatively, name this type of error after the pioneer’s
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
97
name—“Stroop error” in this study. In table 4-10, there were 86% of speech errors produced by reading the term rather than naming the color in task 3, and there were 75.9% of errors by the same way in task 4. The mean value of this effect is 205, which far exceeds the total mean value 65.65, so this effect seems to be obvious and dominant in lexical encoding. According to the post-hoc result, we found that number of Stroop type contrasted significantly with the numbers of rhyme, prenuclear glide, vowel, and coda.
In order to account for the Stroop errors, we would like to analyze the lexical process in terms of connectionist model in figure 4-1. In task 3, when the trials presented, the dual inputs caused the relevant nodes in lexical and phonological layers to be prepared for activation. It seems to be reasonable that errors came from reading color terms by accident during naming task, because the wrong visual representation induced relevant nodes being activated and encoded to the motor program. These nodes would receive more strength than the nodes beyond the dual inputs. That could serve to account for why the number of Stroop errors always exceeded producing other error naming. If the dual inputs were both attributed to color lexicon, as in task 3, the chance to retrieve the literal term might get a little higher. If there were no semantic relation, as in task 4, the chance might get decreased slightly. The strength of the whole network still played a crucial part in generating speech errors and the chance of Stroop error.
In current study, the classical Stroop technique opens up another window for lexical processing and linguistics issue. Stroop errors provide a new direction to explain the generation of speech errors resulting from visual representation. We always focus on one visual task during lexical processing, but other inputs which is not intended also come into our sensory channel and make relevant nodes in lexical network ready to be activated. If there are more similarities in their linguistic
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
98
characteristics, and the strength of lexical network for dual inputs would be stronger.
Therefore, the chance of retrieving the wrong lexicon might get higher.
To sum up, table 4-13 is shown to conclude the above linguistic effects we have discussed so far.
Table 4-13. Summary of Linguistic Effects
Effects Saliency
Initialness Salient in number
Rhyme No
Vowel No
Tone Salient
Phonotactic Regularity Salient Syllable Structure Salient
Stroop Effect Salient
4.4. The Structure of Speech error and Reaction Time: Task 5
In order to testify the possible units in lexical process, we conducted shared unit test and observed their processing speed and error amounts. In this experiment, we recruited six sets of carriers which share different phonological units with the target colors, and we asked subjects to name the visual colors instead of reading the characters. Table 4-14 and figure 4-2 show the results of speech errors and response time in shared unit test.
‧
Table 4-14. The Structure of RT and Errors N in Test 5 (N=376)
Shared Units Onset Vowel Rhyme Syllable Tone Syl.+tone Average
Shared Units Onset Vowel Rhyme Syllable Tone Syl.+tone Average