Chapter 3 Experiment One: Syntactic category and semantic constraint effects
3.4 Discussion
Experiment 1: Statistic results of all effects for the probability measures on post-target words
Experiment 1 demonstrated the syntactic category effects (V > N) of the target words and the dominant meaning of homographs on different ROIs. In the target region, the syntactic category effect of the target words reflected in the first-pass
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
probability measure RFR, indicating that verbs are inherently more difficult to process than nouns. This finding is consistent with the study of J. A. Sereno (1999), in which subjects responded to nouns significantly faster than to verbs in both the categorization task and the lexical decision task. In addition, no significant syntactic category effect of the dominant meaning was found in the target region, indicating that the dominant meaning is not instantly activated when readers first encounter homographs in sentences. On the other hand, the syntactic category effect of the dominant meaning shown in the post-target region reveals that the activation of nouns is easier and faster than that of verbs. During the first-pass reading on the post-target, the lower skipping rate and higher refixation rate shown when the dominant meaning of target words was a noun indicated that the dominant meaning, if a noun, had already been activated. The inconsistency of syntactic category between the dominant meaning and the contextual constraint resulted in the spillover. However, it was not until the second-pass reading (RRR & RIR) on the post-target that the processing cost (V > N) of the dominant verb meaning occurred. These results suggested that it takes more efforts for readers to process a verb than a noun. Our findings is in agreement with the Chinese study of 張亞旭 et al. (2003), in which Chinese NV caused larger
processing difficulty than VN in the disambiguating regions. In their study, NV and VN were preceded by a neutral context and followed by a verb-biased context. The
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
results showed that both NV and VN caused longer reading times relative to unambiguous controls in the disambiguating regions, indicating that readers selected the noun meaning for both SCA words. However, the increased reading times caused by NV lasted longer than those caused by VN. Thus, they suggested that it was easier for readers to assign the syntactic category in the case of NV. In summary, syntactic category would influence the processing of homographs. The syntactic category of the target words has an immediate effect in the target region, while that of the dominant meanings exerts influences in the post-target region.
For the ambiguity effects, Experiment 1 demonstrated a delay of the subordinate bias effect (A > UA). Generally, the SBE for the four types of homographs did not occur until the second-pass reading in the post-target region. The absence of the SBE in the target region might result from different reasons for different types of homographs. For NN homographs, inconsistent with the predictions of both models, the SBE was delayed until the second-pass reading in the post-target region. This might result from the strongly biased semantic context, since the syntactic bias from the preceding context is unhelpful for the selection of appropriate meaning. The elimination of the SBE has been reported by some researcher using strong contextual manipulations or priming words (Seidenberg et al., 1982; Vu & Kellas, 1999; Vu, Kellas, Metcalf, & Herman, 2000; Vu, Kellas, & Paul, 1998). For VV homographs,
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
both the strongly biased semantic context and the processing difficulty of verbs may cause the absence of the SBE. Like NN homographs, the syntactic bias from the preceding context cannot select an appropriate meaning. In addition, previous research has demonstrated that the resolution of VV homographs is delayed (Pickering & Frisson, 2001). For SCA words, VN and NV homographs, though the SBE was absent on the target, it was still reflected in RRT and ROR in the post-target region, indicating that the syntactic category constraint from the preceding context does not prevent the activation of the dominant meaning. As for the absence of the SBE in the target region, it is possible that the processing difficulty of verb meaning delays the competition between alternative meanings.
A further examination on the effect size of ambiguity reveals a difference between VN and NV homographs. Numerically, NV obtained a larger size of the ambiguity effect than VN. In the target region, both the first-pass (FFD, SFD, GD, RFR) and second-pass (GPT, RRR, RIR) measures revealed a pattern of SBE for NV, but this was not the case for VN. In the post-target region, the ambiguity effect also showed a larger size for NV compared to VN in GD, GPT, and ROR. This finding suggested that some probabilistic constraints, other than the contextual constraints, may also influence the semantic resolution of SCA words. The inherent processing difficulty and meaning frequency are related to the activation speed of the alternative meanings
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
of SCA words. When the dominant meaning is a noun and the subordinate meaning is a verb (i.e., NV), the dominant meaning can be activated much faster than the subordinate meaning. Thus, with the strongly subordinate-biased preceding context, the dominant meaning would compete with the context-intended subordinate meaning, resulting in a larger ambiguity effect. Our finding is consistent with the results obtained in 張亞旭 et al. (2003) that the increased reading times caused by NV
lasted longer than those caused by VN.
With respect to the sentence processing models, our results agree with the assumptions of the constraint-satisfaction models. Non-syntactic information is immediately considered during lexical ambiguity resolution. Our findings showed the absence of the SBE in the target region, indicating the influence of semantic constraint from the preceding context. The subordinate-biased semantic information guided the readers to activate the subordinate meaning at the beginning. As a result, for NN and VV, readers did not encounter meaning competition in the target region. In addition, our findings also showed the SBE was delayed until the second-pass reading on the post-target, as opposed to the prediction of syntax-first models. The syntactic category constraint cannot determine the semantic resolution of SCA words. Thus, for VN and NV, meaning competition still happened and the SBE arose in the post-target region. Taken together, the findings in Experiment 1 imply that semantic constraint
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
exerts an early influence on lexical ambiguity resolution and that the syntactic category constraint cannot determine the semantic resolution of SCA words.
In fact, Folk and Morris (2003) also observed the delayed SBE for VN in semantically- and syntactically-biased context, and they suggested that the SBE was delayed because syntactic category assignment can restrict semantic resolution to the syntactically appropriate meaning of SCA words. Nevertheless, this explanation cannot explicate the unequal effect sizes of the SBE between VN and NV in our study.
Although the SBE was absent in the first-pass reading on target words, it remains unclear whether the delay of the SBE for SCA words in the semantically- and syntactically-biased context was due to the semantic constraint, the syntactic category constraint, or both. Some previous research has suggested that a strongly biased context may result in selective access or eliminate the SBE during lexical ambiguity resolution. For example, Seidenberg et al. (1982) found lexical ambiguity resolution was selective when the preceding context involved a lexical-priming word (i.e., a word semantically-related to either meaning). Vu et al. (1998) manipulated the semantic specificity of verb and subject noun to investigate whether lexical ambiguity resolution can be influenced by the strength of sentence constraints. The results showed a multiple access in the ambiguous condition (e.g., He located the bat.) but a selective access in the strongly-biased condition (e.g., dominant-biased: He splintered
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
the bat.; subordinate-biased: He wounded the bat.) whenever the dominant or
subordinate meaning was context-intended. Both studies supported the idea that strong semantic constraints can restrict lexical ambiguity resolution to the context-appropriate meaning, which may be responsible for the absence of the SBE in the first-pass reading on target words in Experiment 1.
In addition, compared to English and Indo-European languages, Chinese morphology allows less transparency on syntactic category. The language-specific properties may decrease the functional primacy of syntactic category assignment during Chinese sentence comprehension. Some previous empirical findings have suggested that syntactic processing does not precede semantic processing in Chinese (Wang et al., 2013; Yang et al., 2009; Yu & Zhang, 2008; Zhang et al., 2013; Zhang et al., 2010). Thus, in Experiment 2, we used sentences in which only the syntactic category constraint preceded the targets to examine whether the strong semantic constraint can be responsible for the absence of the SBE in Experiment 1 and whether the syntactic category constraint is sufficient to determine the semantic resolution of SCA words.
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
80
Experiment Two:
Lexical ambiguity resolution with syntactic category constraint
Experiment 2 examined whether the syntactic category constraint alone can determine the meaning resolution of syntactic category ambiguity. The biased homographs were embedded in sentences in which the preceding context provided no semantic information but syntactic category information consistent with the subordinate meaning. Disambiguating semantic information was only available in the succeeding context. Frequency-matched unambiguous words were used as control words.
According to both syntax-first and constraint-satisfaction models, the SBE was not expected for NN and VV until the subordinate-biased disambiguating information was encountered. In the target and post-target regions, readers would initially select the dominant meaning, since the preceding syntactic category constraint cannot distinguish between the alternative meanings of homographs. In the disambiguating region, the subordinate meaning would be activated and compete with the dominant meaning, resulting in a first-pass processing difficulty. Afterwards, readers would experience difficulty in integrating the subordinate meaning into the sentences, so a
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
second-pass processing difficulty was expected in all analyzed regions.
On the other hand, for VN and NV, syntax-first models predicted neither first-pass nor second-pass processing difficulties in any region, since the syntactic category constraint from the preceding context determined the subordinate meaning from the very beginning. In contrast, constraint-satisfaction models predicted both first-pass and second-pass processing difficulties in all analyzed regions, since readers consider the syntactic category constraint and meaning dominance simultaneously. In the target region, the subordinate meaning was supported by syntactic category constraint and competed with the dominant meaning, resulting in a first-pass processing difficulty which might spill over to the post-target region. When the disambiguating information was available, readers may spend time reading back and forth to disambiguate, resulting in second-pass processing difficulties in all analyzed regions.
4.1 Method
4.1.1 Participants
Forty undergraduate and graduate students (7 males and 33 females) aged between 18-25 years old (mean age = 21) were paid to participate in Experiment 2.
All participants were native speakers of Mandarin Chinese and had normal or
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
corrected-to-normal vision. None had participated in any previous norming study and Experiment 1.
4.1.2 Materials and design
Eighty biased homographs, including 4 types (NN, VV, VN, and NV), and 80 corresponding unambiguous words (NNN, VVV, NVN, and VNV) were used as target words in Experiment 2. Most of the words were the same as Experiment 1. For better control over the word properties, 10 homographs (2 NN, 2 VV, 1 VN and 5 NV) and 15 unambiguous words (5 NNN, 4 VVV, 1 NVN and 5 VNV) used in Experiment 1 were replaced with new words in Experiment 2. Table 14 presents the means and standard deviations of word-form frequency, word stroke, and frequency of the first character (C1F) for homographs and unambiguous words used in Experiment 2. The two-way analysis of variance (ANOVA) with word type and ambiguity as independent variables were conducted for the word properties. For word-form frequency, word stroke and C1F, the main effects and the interaction were non-significant (Fs < 3, ps
> .07), indicating there were no significant differences among word types and between homographs and unambiguous words.
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
Table 14
Means (and standard deviations) of word properties for each condition in Experiment 2
Word property
Type Ambiguity Frequency Stroke C1F
NN A 7.82 (8.08) 18.25 (5.64) 973.11 (965.64)
UA 7.96 (8.66) 18.50 (4.36) 913.17 (792.01)
VV A 9.31 (8.98) 20.60 (5.11) 739.54 (781.72)
UA 8.55 (7.92) 20.55 (4.48) 697.45 (933.79) VN A 12.83 (10.74) 18.90 (5.04) 910.94 (841.67) UA 13.11 (12.50) 19.40 (5.62) 765.33 (780.79) NV A 10.74 (7.70) 20.15 (4.83) 1040.50 (928.43)
UA 7.84 (7.85) 20.75 (4.34) 773.59 (769.04) Note. A = ambiguous words; UA = unambiguous words; C1F = frequency of the first character
Sentences were designed in such a way that the preceding context provided no semantic but syntactic category information consistent with the subordinate meaning of the homographs. For syntactic category bias, as in Experiment 1, we used a pre-target word to bias the syntactic category of the targets toward the syntactic category of homographs’ subordinate meaning. A disambiguating word, which was semantically related to the subordinate meaning of the homographs, was located in the succeeding context. An intervening region (at least four characters) was inserted between the ambiguous word and the disambiguating word. Pre-target, target and disambiguating regions were all disyllabic Chinese words. A total of 80 sentence frames were constructed where the homograph and the corresponding unambiguous word can fit the context into the same position. The beginning of target words and
‧
disambiguating words was respectively located in the range of 12th to 14th and of 18th to 21st character in the sentences, which consisted of 25 to 27 characters. The construction of experimental lists was identical to Experiment 1. Table 15 shows the example sentences of each condition.
Table 15
Examples of targets and sentences for each condition in Experiment 2
Sentence
Type Conditions Preceding context Target
intervening
Note. A = ambiguous words; UA = unambiguous words
4.1.3 Norming studies of ambiguous words and sentential contexts
Prior to the eye-tracking experiment, seven norming studies were carried out to ensure that word semantics and sentential contexts are successfully manipulated.
Table 16 shows the summarized results of all norming studies in Experiment 2. A total of 160 Chinese sentences were chosen for Experiment 2, including 80 homograph-embedded (20 for each type) and 80 corresponding unambiguous word-embedded sentences (see Appendix E).
‧
Means (and standard deviations) of words and sentences in all norming studies for Experiment 2
UA Identical to the corresponding homographs 4. Semantic bias (preceding)
(1 = not related; 7 = much related)
5. Syntactic category of target word
(1 = not plausible; 7 = very plausible)
A 5.79 (0.38) 5.61 (0.53) 5.75 (0.41) 5.79 (0.34) UA 5.80 (0.35) 6.03 (0.36) 5.88 (0.44) 5.79 (0.45)
Note. A = ambiguous words; UA = unambiguous words; A_dom = dominant meaning of ambiguous words; A_sub = subordinate meaning of ambiguous words
4.1.3.1 Norming study 1: Meaning dominance
Sixteen new ambiguous words and 22 unambiguous words were rated by 40 Mandarin-speaking participants. The procedure and the criteria for selecting word materials were identical to Norming study 1 in Experiment 1. Overall, eleven biased ambiguous words and 20 unambiguous words met the criteria. Eventually, ten
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
homographs (2 NN, 2 VV, 1 VN and 5 NV) and 15 unambiguous words (5 NNN, 4 VVV, 1 NVN and 5 VNV) used in Experiment 1 were replaced with new words. Means of meaning dominance were 0.86, 0.88, 0.85, and 0.92 for NN, VV, VN, and NV, indicating that they were highly-biased homographs since one of the meanings of homographs accounted for over 85 % of the subjects’ interpretations.
4.1.3.2 Norming study 2: Meaning relatedness
Meaning relatedness of 23 new ambiguous words was rated by 40 Mandarin-speaking participants, who also participated in Norming study 1. Two lists were constructed to avoid the participants rating both the meaning dominance and the meaning relatedness of same words. The procedure and the criteria for selecting word materials were identical to Norming study 2 in Experiment 1. Overall, twenty biased ambiguous words were qualified homographs and 10 were substituted for the homographs used in Experiment 1. The one-way ANOVA on meaning relatedness showed no significant difference across four types of homographs, F(3, 76) = .82, p
= .49.
4.1.3.3 Norming study 3: Word predictability and syntactic category bias
Forty Mandarin-speaking participants participated in Norming study 3－7; none
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
of them had participated in any previous norming study, nor the eye-tracking experiment. Forty new sentence frames were constructed in a way that both the homograph and the corresponding unambiguous word can fit into the same one. A new questionnaire was created, in which the preceding context, the definitions of the dominant and the subordinate meaning of the homograph, the target word (either ambiguous or unambiguous), the succeeding context, a parallel display of the disambiguating word and the target word (either the homograph with its subordinate meaning or the unambiguous word), and the entire sentence were orderly presented.
The procedure and the criteria were identical to Norming study 3 in Experiment 1.
The two-way analysis of variance (ANOVA) with word type and ambiguity as independent variables was conducted for word predictability and for syntactic category bias. For word predictability and syntactic category bias, the main effects and the interaction were non-significant (Fs < 3, ps > .10), indicating there were no significant differences among word types and between homographs and unambiguous words.
4.1.3.4 Norming study 4: Semantic bias
The task of rating semantic bias was altered in Experiment 2. For the homograph-embedded sentences, the participants were asked to rate the semantic
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
relatedness between the preceding context and both meanings of the biased homographs, respectively. For the unambiguous word-embedded sentences, they were asked to rate the semantic relatedness between the preceding context and the unambiguous word. The reason we changed the task was to avoid the influence of meaning dominance and syntactic category constraint as much as possible. If, as in the semantic bias rating in Experiment 1, the participants had been asked to choose one definition which they considered more consistent with the preceding context, they would have probably chosen the dominant meaning for NN and VV homographs or the subordinate meaning for VN and NV homographs, which was consistent with the syntactic category constraint. As a result, directly evaluating the semantic relatedness on a 7-point scale (where 1 = not related; 7 = much related) might be a better way to probe whether the preceding context was indeed semantically neutral. In addition, the participants were asked to write down any word in the preceding context which they considered semantically related to the target words (either the dominant or the subordinate meaning for the homographs). On the other hand, we followed the procedure of the semantic bias rating in Experiment 1 to probe the semantic bias in the succeeding context.
The two-way analysis of variance (ANOVA) with word type and ambiguity as independent variables was conducted for semantic bias. There was a significant
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
difference among word types, F(3, 228) = 3.90, p < .01. Post hoc comparisons using the Bonferroni test revealed that VV obtained significantly lower semantic bias than VN (p < .01). The main effect of ambiguity (A_dom, A_sub, UA) was also significant, F(2, 228) = 6.24, p < .01. Post hoc comparisons using the Bonferroni test revealed
that the dominant meaning of homographs obtained significantly lower semantic bias than did the subordinate meaning of homographs (p < .01) and unambiguous words (p
< .05). However, the interaction was non-significant, F(6, 228) = .73, p = .62.
4.1.3.5 Norming study 5: Syntactic category judgment of target words
The procedure and the criteria were identical to Norming study 5 in Experiment 1. The two-way analysis of variance (ANOVA) with word type and ambiguity as independent variables was conducted for syntactic category judgment. There was a significant difference among word types, F(3, 152) = 9.45, p < .001. Post hoc
The procedure and the criteria were identical to Norming study 5 in Experiment 1. The two-way analysis of variance (ANOVA) with word type and ambiguity as independent variables was conducted for syntactic category judgment. There was a significant difference among word types, F(3, 152) = 9.45, p < .001. Post hoc