Chapter 5 Experiment Two: The time course of lexical ambiguity resolution
5.1.3 Norming studies
5.1.3.3 Norming study three: Semantic Relatedness Rating
國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
and graduate students (8 males, 12 females) participated in this norming survey. The average rating for homograph- and monograph-embedded sentences was 5.56 and 5.66 respectively. Paired t test showed that they did not differ significantly [t(23) = -.628, p = .536].
5.1.3.3 Norming study three: Semantic Relatedness Rating
This norming study is to assess semantic relatedness for spoken homograph target and its dominant- and subordinate-related associative words, and the unrelated words, also for monograph target and its semantic associative words.
Forty participants of undergraduate and graduate students took part in this norming study and they were in the age range of 18 to 26 years. They were given the definition of dominant or subordinate meaning and their semantic associative words and were asked to rate the definition–target pair on a 7-point scale ranging from 1(very unrelated) to 7 (very related). Ninety-six definition–target pairs were given to forty subjects in total. Materials were divided into two versions such that each subject saw an equal number of semantic related and unrelated words. Only semantic associative words that were considered highly-related (4-7 point at 7-point scale) were included in the main experiment. The unrelated words were excluded if the rating value was above 3. Twenty-seven items were dropped and replaced with new
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
qualified items. The dominant meaning–DR pairs received a mean rating of 5.75 and subordinate meaning–SR pairs received a mean rating of 5.27. In addition, the related meaning–MSR pairs had a mean rating of 5.16. Independent t-test was performed for dominant meaning–DR and subordinate meaning–SR, dominant-meaning–DR and related meaning–MSR. The results showed that DR and SR did not differ significantly [t(46) = 1.691, p = .098]. However, there was a marginal significance between DR and MSR [t(46) = 1.965, p = .055]. Paired t-test was performed for subordinate meaning–SR and related meaning–MSR. No significance was found [t(23) = 0.419, p
= 0.679]. Furthermore, the semantic associative words with each meaning (e.g., dominant meaning-DR, subordinate meaning-SR and related meaning–MSR) were judged to be significantly semantically related to the target words than the other three stimuli (all ps < .000) and there was no difference among the other three. The results are summarized in Table 11.
‧
Table 11. Results of the semantic relatedness norming for Experiment 2. Mean (with standard deviations in parentheses) semantic relatedness between the target words and each type of printed word
Target word Type of word in visual display
DR SR UR1 UR2
1. Homograph
Dominant meaning 5.75 (1.00) 1.40 (0.41) 1.12 (0.16) 1.20 (0.33) Subordinate meaning 1.56 (0.40) 5.27 (0.98) 1.17 (0.19) 1.20 (0.35)
2. Monograph MSR UR1 UR2 UR3
Related meaning 5.16 (1.07) 1.37 (0.39) 1.13 (0.15) 1.28 (0.40) Note. DR= dominant-related associative words; SR= subordinate-related associative words; MSR= semantically-related associative words in monograph condition; UR=
unrelated words
5.1.4 Procedure
Eye movements were monitored and measured with an SR Research EyeLink 1000 Desktop Mount eye tracker, sampling at 1000 Hz (the eye-movement recording procedure was identical to Experiment 1). The character size for visual display was 42x42 pixels. One character on the screen corresponded approximately to 3.4° of visual arc. Spoken sentences were presented to the participants through headphones.
Prior to the experiment, the instruction and 6 practices were given. The structure of each trial was as follows (see Figure 6). First, a central fixation cross appeared on the screen as the auditory presentation of a sentence was initiated. Until 1200 ms before the acoustic onset of the target word, the cross disappeared and was replaced by a blank screen for 500 ms. Then, a cross appeared again for 500ms and subsequently
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
the display of four words appeared on the time 200 ms before the acoustic onset of the target word. The trial was terminated as the sentence utterance ended. Participants were instructed to listen to the sentence carefully and look whatever they want except taking their eyes off the screen throughout a trial. One-third of the trials were followed by true-and-false comprehension questions to ensure that they understand the sentences. The entire experiment lasted less than 35 minutes.
Figure 6. Experimental procedure of Experiment 2.
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
5.2 Results
The number of fixations on all types of the printed words was obtained every millisecond for the duration of 4000 msec. Fixation proportions were computed across trials. Fixation proportions of competitors represented the ratio between the number of fixations to a particular semantic competitor (DR, SR or MSR) and the sum of all fixations; in addition, fixation proportions of competitors represented the ratio between the average fixations of the distractors and the sum of all fixations Figure 7 plots the fixation proportions over time in all conditions. For the statistical analyses, we computed mean fixation proportions in 100-msec intervals. The means and standard errors of fixation proportions are shown in Table 12. The increase in looks to the printed target identical to the spoken target words in filler trials began less than 500 ms after target’s acoustic onset and was earlier than SR and MSR in homograph and monograph conditions (Figure 8). We performed the analyses of variance (ANOVAs) by participants (F1) and items (F2) for each time period during 501-1300 ms (Table 13).
There were statistically significant differences among DR, SR, and UR both by participants and by items from 501 ms to 1300 ms. The statistical results were reported here, taking the 501-600 bin as an example. At 501 ms, F1(2, 62) = 5.50, p
= .006, F2(2, 46) = 4.59, p = .02. The remaining results in different time bins can refer
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
to Table 13. At the acoustic offset of the initial character of target word (approximately 400 ms), the fixation proportions to SR in the homograph condition started to diverge. Starting at 501 ms, the subordinate-biased context increased looks to SR compared with unrelated distractors. This divergence was significant both by participants and by items [501 ms, z1 = -3.088, p = .006, z2 = -2.836, p = .01]. The large difference in fixating on the SR compared to that on the distractors continued at the later time points. It was assumed that any difference before the word was fully specified (i.e. acoustic offset of the target word) can be taken as the contextual influence. Fixation proportions to the DR did not differ significantly from those to the distracters from 501 ms to 900 ms by participants [all ps > 0.1] and 501 ms to 1000 ms by items [all ps > 0.1]. However, there were more fixations towards the DR, as compared with the distractors from 901 ms to 1300 ms, that was significant from 901 ms to 1300 ms [901 ms, z1 = -3.13, p = .003] and marginal significant from 1001 ms to 1300 ms [1001 ms, z1 = -2.375, p = .05].
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
Figure 7. Fixation proportions to all types of printed words across two experimental conditions and one target (filler) condition in experiment 2. The x-axis shows time in milliseconds from the display onset, for 4000 ms
In the monograph condition, more fixations were directed to MSR compared to those to the unrelated distractors, starting at 501 ms, significant both by participants and by items [501 ms, z1 = -2.671, p = .007, z2 = -2.645, p = .008] and was maintained
Homo (dom/sub meaning) related) Mono (meaning related)
Mono (target)
Word offset Word offset
Word offset
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
at the later time points. This indicates that the strength of context for homograph and monograph is fairly consistent. In the filler condition, participants strongly shifted their eye gaze towards the matching printed words relative to the distractors at 401 ms, significant both by participants and by items [401 ms, z1 =-4.714, p = .000, z2 = -3.694, p = .000].
Table 12. Means and standard errors (in parentheses) of fixation proportions on three types of visual words in two experimental conditions from 1 ms to 1300 ms in Experiment 2
Homograph Monograph
target onset SR DR UR MSR UR
1-100 0.07(0.01) 0.09(0.02) 0.11(0.01) 0.10(0.02) 0.10(0.01) 101-200 0.15(0.02) 0.17(0.02) 0.19(0.01) 0.14(0.02) 0.18(0.01) 201-300 0.18(0.02) 0.19(0.02) 0.21(0.01) 0.17(0.02) 0.21(0.01) 301-400 0.20(0.01) 0.21(0.02) 0.22(0.01) 0.20(0.02) 0.22(0.01) 401-500 0.24(0.02) 0.22(0.02) 0.21(0.01) 0.22(0.02) 0.22(0.01) 501-600 0.29(0.02) 0.21(0.02) 0.20(0.01) 0.28(0.02) 0.21(0.01) 601-700 0.29(0.02) 0.21(0.01) 0.20(0.01) 0.31(0.02) 0.20(0.01) 701-800 0.31(0.02) 0.21(0.01) 0.20(0.01) 0.33(0.02) 0.19(0.01) 801-900 0.32(0.02) 0.23(0.02) 0.19(0.01) 0.35(0.02) 0.19(0.01) 901-1000 0.32(0.02) 0.26(0.02) 0.18(0.01) 0.36(0.02) 0.19(0.01) 1001-1100 0.35(0.02) 0.27(0.02) 0.16(0.01) 0.34(0.02) 0.19(0.01) 1101-1200 0.35(0.02) 0.26(0.02) 0.17(0.01) 0.35(0.03) 0.19(0.01) 1201-1300 0.31(0.02) 0.26(0.02) 0.17(0.01) 0.36(0.03) 0.19(0.01)
‧
Figure 8. Time course of fixation proportions on different word types in two experimental conditions from 1 ms to 1300 ms
Note. F1/z1 = 32 participants. F2/z2 = 24 items.
1-100 101-200 201-300 301-400 401-500 501-600 601-700 701-800 801-900 901-1000 1001-1100 1101-1200 1201-1300 1301-1400 1401-1500 1501-1600 1601-1700 1701-1800 1801-1900 1901-2000 2001-2100 2101-2200 2201-2300
SR DR UR
MSR UR
Table 13. Analyses of variance by participant and item comparing mean fixation proportions to competitors with those to the average of the distractors in 100-ms bins, from 501 ms to 1300 ms after acoustic target word onset in homograph condition in experiment 2
Time bin
Condition Test 501-600 601-700 701-800 801-900 901-1000 1001-1100 1101-1200 1201-1300
Homograph,
-2.712 -3.053 -3.895 -3.267 -2.361 -2.758 -3.127 -1.691 p = .02 p = .007 p < .001 p = .003 p = .05 p = .02 p = .005 p = 0.27 z2
-2.49 -2.444 2.861 -2.299 -1.453 -1.727 -1.886 -1.21 p = .04 p = .05 p = .01 p = .06 p = 0.4 p = 0.3 p = 0.2 p = 0.6
-0.367 -0.319 -0.422 -1.442 -3.13 -3.791 -3.599 -3.009 p = 1 p = 1 p = 1 p = .45 p = .003 p < .001 p < .001 p = .008
z2
-0.346 -0.256 -0.31 -1.015 -2.037 -2.375 -2.171 -2.153 p = 1 p = 1 p = 1 p = 0.9 p = 0.1 p = .05 p = .09 p = .09
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
5.3. Discussion
The key finding in Experiment 2 was that a greater fixation proportions were towards contextually appropriate semantic associates (SR) when participants heard the first character of ambiguous words. However, fixations on the contextually inappropriate semantic associative words (DR) increased as identifying information of lexical semantics was perceived. Table 14 summarizes the fixation proportions of competitors, filler-targets and distrators across three conditions. Statistically greater fixation proportions of SR and MSR comparing with their UR in homograph and monograph condition respectively were firstly found at the target offset, while a greater fixation proportions of DR comparing with UR was found at 200 ms after target offset. Contextual information influenced a relatively greater fixation proportions on the SR and MSR comparing with those on their respective distractors, starting from 500 ms after the target onset. This may suggest the successful manipulation of the contextual strength. Furthermore, the DR varied greatly as compared with the distractors from 901 ms to 1300 ms entailed that context influence emerged 400 ms earlier for contextually-selected semantic associative words than they did for the contextually-unselected semantic associative words. This is consistent with all accounts of the subordinate bias effect, which reflected the rapid use of context. The intriguing findings were that first, DR attracted more fixations than its
‧
relative distractors even in the strongly subordinate-biased context. Second, the time course of the activation of dominant meaning revealed that the associated semantic representation of the context-unselected meaning was fairly weak since the context-selected meaning arrived early prior to the target-offset. Together, these finding showed the subordinate bias effect and thus supported the competition account and reordered access model, but it is not consistent with frequency account and selective access model.
Table 14. The average fixation proportions to each type of words at acoustic target offset and 200 ms after the target offset across three conditions in Experiment 2
Time Region (milliseconds) Type of word
DR SR UR
1.homograph condition
p(fix) at offset 0.21 0.31*** 0.20
p(fix) at offset+200 ms 0.26*** 0.32*** 0.18
2.monograph condition MSR UR
p(fix) at offset p(fix) at offset+200 ms
0.34*** 0.19
0.35*** 0.19
3.target condition TAR UR
p(fix) at offset 0.49*** 0.15 p(fix) at offset+200 ms 0.56*** 0.11 Note: Difference score to UR p < .05 for participants*
Difference score to UR p < .01 for participants **
Difference score to UR p < .001 for participants ***
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
The average utterance duration of target homographs was 800 ms, so the linguistic context began to affect looking to SR prior to homograph offset. It is likely that context influences reflect initial lexical access of spoken word before the information of whole word was available. In our experiment, the meaning relatedness of SR was not different from that of DR according to the rating results. Therefore, it is likely to rule out the possibility of any preference of looks towards either semantic associate. In the subordinate biasing context, subordinate meaning was activated earlier than dominant meaning before the ambiguous word was available. In subsequent, from 901 ms to 1300 ms, both meanings are activated in the same time window. At later time, the meaning was revised and selected in order to arrive at a coherent contextual interpretation. As for the status of the unselected meaning under the constraining context of subordinate meaning, we may also provide the temporal evidence of dominant meaning activation the time course from 901 ms to 1300 ms. It appears that the subordinate-biased context did not completely eliminate the dominant meaning, instead, it was activated independently and took its advantage gradually on the basis of lexical dominance, namely, the stronger strength of form and meaning mapping. In short, the dominant meaning was activated but was delayed because of the contextual biasing to the subordinated meaning. This evidence seems to rule out the selective access model, which posits the dominant meaning should not be
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
activated at all. On the contrary, our data tend to support the predictions of the fate of unselected meaning proposed by reordered access model. That is, even the strong subordinate biased context did not override the automatic activation of the dominant meaning. However, the fixation proportions to the dominant meaning are rather fewer than those to the subordinate meaning. It seems that we cannot completely rule out the possibility that the initial activation of the dominant meaning was modulated by the contextual information. It is likely that a strong subordinate biasing context may decrease activation of the dominant meaning.
Falk Huettig and McQueen (2007) have shown that semantic information was not retrieved when using printed words as visual stimuli. They suggested that because reading a word provided much more direct access to phonological knowledge. It was true for alphabetical languages and may be not the case for Chinese. We found fixation proportions towards the printed word fully matched with the spoken target diverged from those towards the unrelated distractors at 400 ms. The results were also similar to Mcqueen and Viebahn (2007) which found significant fixation proportions towards offset mismatched bisyllabic word at about 400 ms (corresponding to the onset of the final phoneme). The semantic information was retrieved at 500 ms in the present study, finding that phonological and semantic information are accessed at different time and play comparable roles during word recognition in Chinese.
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
79
6.1 Dynamic processing of context influence and meaning dominance
The results from the present two eye-tracking experiments demonstrate the interaction of contextual influence and lexical activation during lexical ambiguity resolution. In Experiment 1, the fixation times of homographs were compared to those of low frequency unambiguous words. Subordinate bias effect emerged consistently as the control used form-matched unambiguous word (Sereno et al., 1992). Sereno et al. (2006) found SBE was not attenuated even in a strongly biasing context. They suggested the reason may be due to the special situation that the word form of ambiguous word was a HF word but its functional link to context was subordinate in terms of meaning. However, the results from Experiment 1 indicated that SBE occurred as the homograph was an LF word both in terms of its word form and meaning.
Experiment 2 provides a comprehensive time course of lexical ambiguity resolution on spoken word recognition which reveals the temporal information of the contextual influence on lexical activations. When listeners are given sufficient contextual information, they produce a greater fixation proportions towards the contextually-selected semantic associative words. The consistent results were found in
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
both homograph and monograph conditions that context influences occur from about 500 ms, shortly after the acoustic onset of the target word. This may indicate that at an earlier stage, the context is acting on the access of the subordinate meaning of the homograph. We found the dominant meaning was activated from 901 ms to 1300 ms, approximately after the acoustic offset of the spoken target and before the completion of next word. The converging evidence from both visual and auditory presenting experiments shows the robust SBE and the activation of the dominant meaning.
We then compare our results with those in cross-modal priming studies to gauge the theoretical implications of temporal dynamics of lexical activation and contextual influence. Generally, two levels of semantic access are distinguished based on the results of lexical decision studies. Pre-lexical stage involves word recognition and meaning activation, while post-lexical stage deals with semantic selection and integration. Onifer and Swinney (1981) presented sentences that biased for the dominant or for the subordinate meaning of an ambiguous word. The results demonstrated the activation of multiple meanings irrespective of the context when presenting visual target word at the auditory offset of ambiguous words. According to the exhaustive or multiple access model, context can only penetrate lexical activation at post-lexical stage, but not at the earlier stage. It is likely that the frequency effect emerged after the multiple meanings associated with a word were accessed.
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
Seidenberg et al. (1982) delayed the presentation of visual target until 200 ms or more after the spoken homophone, by that time, demonstrating that a single meaning had been selected after initially activating multiple candidates. It is assumed to reflect post-lexical stage of using context to select an appropriate meaning.
From the results of our Experiment 2, sentential information aids the processing of Chinese homographs from early on within the acoustic boundary of the homograph in natural speech. The influence of sentential context is thus pushed to a much earlier stage than what has been proposed by the multiple access model. At a subsequent time, after the homograph is being heard, the dominant meaning is activated, thus semantic competition occurs. However, it was hard to separate the stage of this activation possibly occurred at the level of lexical or post-lexical processing. So it may be more likely to view the lexical ambiguity resolution as the continuous graded constraint of context and frequency effects rather than an order-based of two-stage processing for different meanings (Mirman, 2008). In terms of continuous graded constraints, it seems that both contextual bias and meaning dominance are used in parallel by the comprehenders. Two implications thus can be drawn from these findings, first, contextual information affects the ambiguity resolution occurring early before the acoustic offset of homograph. Second, the context reorders the processing of different meanings, the unselected but frequent meaning becomes activated later.
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
6.2 The competition account of subordinate bias effect
Both reordered-access and selective-access models predicted the rapid and early use of context. Consistent with both models, experiment one have shown context affected the fixation proportions on the subordinate meaning (SR associate), starting from 500 ms after the homograph onset. However, the discrepancy of the two models lies in whether the dominant meaning was activated. Our results are more consistent with the reordered access model according to the two findings reported here. First, in Experiment one, longer processing time is demonstrated in both target and post-target region when readers process ambiguous words. Under the selective view, no initial processing time cost should be observed when the context is sufficiently constraining.
Second, in Experiment two, the dominant meaning (DR associate) attracted more looks of fixations than those to the unrelated distractors and above chance level even in the subordinate-biased context. Two meanings were activated at the same time, therefore, competition or processing difficulty occurred which were evident in longer fixation durations in experiment one, supporting the reordered access model. In sum, the dominant meaning was still available though it was delayed. Therefore, the selective access view was ruled out and the homographs were not merely treated as low frequency words.
Another issue was how the context affected the status of contextually-
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
inappropriate meaning. According to reordered access model, two meanings are activated independently. Contexts speed access to the appropriate meaning, while no effect was on the activation of the inappropriate meaning. On the other hand, based on selective access model, given sufficiently constraining contextual information, only the contextually-appropriate meaning should be activated; therefore, the inappropriate
inappropriate meaning. According to reordered access model, two meanings are activated independently. Contexts speed access to the appropriate meaning, while no effect was on the activation of the inappropriate meaning. On the other hand, based on selective access model, given sufficiently constraining contextual information, only the contextually-appropriate meaning should be activated; therefore, the inappropriate