Chapter 4 Experiment Two: Lexical ambiguity resolution with syntactic category
4.2.3 Disambiguating words
Experiment 2: Statistic results of all effects for the probability measures on post-target words
4.2.3 Disambiguating words 4.2.3.1 Duration measures
Means and standard errors of both the first-pass and second-pass duration measures for each condition on disambiguating words are shown in Table 25. The
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
statistic results of all analyzed effects for the duration measures on disambiguating words are listed in Table 26. For the first-pass duration measures, both the main effect of syntactic category and that of ambiguity were non-significant in any measure.
However, SFD and GD revealed significant syntactic category effect of homographs’
dominant meaning when the target word was a verb (ts > 1.9), indicating that the disambiguating words obtained more fixation durations in VV condition than in NV condition. In addition, GD revealed a reverse syntactic category effect of target word when the homographs’ dominant meaning was a noun (b = - .07, SE = .04, t = -1.86), indicating that the disambiguating words obtained more gaze durations in NN condition than in NV condition.
For the second-pass duration measures, GPT revealed significant subordinate-by-dominant syntactic category interaction (b = .08, SE = .03, t = 2.59).
When the target word was a verb, the disambiguating words obtained more go-past times in VV condition than in NV condition (b = .12, SE = .04, t = 2.72). When the homographs’ dominant meaning was a noun, the disambiguating words obtained more go-past times and total viewing times in NN condition than in NV condition (ts <
-1.8). When the homographs’ dominant meaning was a verb, the disambiguating words obtained more go-past times in VV condition than in VN condition (b = .08, SE
= .04, t = 1.83). On the other hand, GPT also revealed marginally significant main
‧
effect of ambiguity (b = .04, SE = .02, t = 1.86), significant ambiguity effect when the homographs’ dominant meaning was a noun (b = .06, SE = .03, t = 2.17), and marginally significant ambiguity effect for NN (b = .07, SE = .04, t = 1.93). RRT
showed no effect.
Table 25
Experiment 2: Means and standard errors of the duration measures for all conditions on disambiguating words
Duration measures (ms) Type
Ambiguity Effect size
A UA A－UA Note. A = ambiguous words; UA = unambiguous words
‧
Experiment 2: Statistic results of all effects for the duration measures on disambiguating words
Means and standard errors of both the first-pass and second-pass probability measures for each condition on disambiguating words are shown in Table 27. The statistic results of all analyzed effects for the probability measures on disambiguating
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
words are listed in Table 28. For the first-pass probability measures, SKIP revealed marginally significant syntactic category effect of homographs’ dominant meaning (b
= .24, SE = .13, p = .07), indicating that the disambiguating words were less skipped in NN condition than in VN condition. RFR showed no effect.
For the second-pass probability measures, both the main effect of syntactic category and that of ambiguity were non-significant in any measure. However, RRR and ROR revealed significant subordinate-by-dominant syntactic category interaction (ps < .05). When the target word was a verb, the disambiguating words obtained more probability of rereading and regression-out in VV condition than in NV condition (ps
< .05). When the homographs’ dominant meaning was a noun, the disambiguating words obtained more probability of rereading and regression-in in NN condition than in NV condition (RRR: b = - .53, SE = .24, p = .03; RIR: b = - .52, SE = .28, p = .06).
When the homographs’ dominant meaning was a verb, the disambiguating words obtained more probability of regression-out in VV condition than in VN condition (b
= .58, SE = .25, p = .02). On the other hand, RIR and ROR revealed ambiguity-by-dominant syntactic category interaction (ps < .05). ROR also revealed the significant interaction when the target word was a noun (b = - .45, SE = .21, p
= .03), indicating larger ambiguity effect for the disambiguating words in NN condition than in VN condition. In addition, RIR and ROR revealed significant
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
ambiguity effect when the homographs’ dominant meaning was a noun (ps < .05).
ROR also revealed significant ambiguity effect for NN (b = .60, SE = .29, p = .04).
Table 27
Experiment 2: Means and standard errors of the probability measures for all conditions on disambiguating words
Probability measures (%) Type
Ambiguity Effect size
A UA A－UA
First-pass SKIP NN 25.38 (2.21) 24.16 (2.18) 1.22 VV 27.25 (2.26) 26.15 (2.23) 1.10 VN 30.18 (2.32) 27.95 (2.28) 2.23 NV 29.59 (2.31) 28.90 (2.30) 0.69 RFR NN 15.12 (2.10) 11.99 (1.90) 3.13 VV 12.77 (1.99) 12.54 (1.96) 0.23 VN 13.60 (2.08) 10.71 (1.85) 2.89 NV 12.04 (1.97) 8.39 (1.68) 3.65 Second-pass RRR NN 23.91 (2.17) 27.18 (2.29) -3.27 VV 30.45 (2.36) 27.46 (2.27) 2.99 VN 25.13 (2.23) 24.94 (2.21) 0.19 NV 18.65 (1.99) 18.90 (2.01) -0.25 RIR NN 12.60 (1.68) 16.36 (1.90) -3.76 VV 14.44 (1.80) 13.73 (1.75) 0.71 VN 15.08 (1.84) 15.06 (1.83) 0.02 NV 10.10 (1.54) 11.02 (1.61) -0.92 ROR NN 14.93 (2.10) 10.03 (1.77) 4.90 VV 14.95 (2.13) 13.94 (2.05) 1.01 VN 8.24 (1.69) 10.83 (1.87) -2.59 NV 11.44 (1.94) 8.09 (1.66) 3.35 Note. A = ambiguous words; UA = unambiguous words
‧
Experiment 2: Statistic results of all effects for the probability measures on disambiguating words
Effects (p value)
1st-pass 2nd-pass
SKIP RFR RRR RIR ROR
(Intercept) <2e-16 <2e-16 <2e-16 <2e-16 <2e-16
amb<A-UA> 0.99 0.11 0.64 0.23 0.12
The ambiguity effects for the four types of homographs were observed in different eye movement measures and regions. Specifically, for NN homographs, the SBE did not occur until the second-pass reading in the post-target and disambiguating region,
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
indicating that readers encountered a conflict when integrating the dominant meaning into the subordinate-biased succeeding context. Similarly, the SBE for VV homographs emerged in the second-pass reading in the target and post-target region.
For VN homographs, the SBE was pervasively shown in the first-pass and second-pass measures in the target and post-target region, suggesting that the dominant meaning was instantly activated even though the preceding context was semantically-neutral but syntactically-biased toward the subordinate meaning. Finally, for NV homographs, the SBE was not apparent but still occurred in the second-pass reading in the post-target region, revealing that the interference from the activation of the dominant meaning.
Experiment 2 demonstrated that the SBE for NV occurred much later and less obviously than that for VN, the pattern of which was opposite to the findings of Experiment 1. A possible explanation may relate to the activation speed of the alternative meanings. Previous findings have demonstrated that nouns can be activated faster than verbs due to its less processing difficulty (Rösler et al., 2001; J. A.
Sereno, 1999; J. A. Sereno & Jongman, 1997). In the case of NV, the dominant meaning is a noun, which can be activated very fast, while the subordinate meaning is a verb, whose activation needs time. Presumably, the noun meaning of NV can be activated much faster than the verb meaning. As a result, in the semantically-neutral
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
context as in Experiment 2, readers may easily select the noun meaning for NV, resulting in less obvious SBE in the first-pass reading. However, readers may encounter a conflict when integrating the noun meaning into the succeeding context, which was biased toward the subordinate verb meaning. Thus, the SBE occurred in the second-pass reading. In the case of VN, the meaning dominance may ‘reorder’ the activation speed of the noun and verb meaning. The activation of the verb meaning can be boosted with the help of its high frequency, while the activation of the noun meaning would slow down due to its low frequency. As a result, the activation speeds of the alternative meanings might be somewhat equivalent. When the preceding context was semantically-neutral, VN would be like a balanced homograph. The SBE for VN occurred instantly in the first-pass reading because of the competition between the two meanings, and the SBE lasted to the second-pass reading since readers made regressions to integrate the subordinate meaning. These findings suggest that the SBE for SCA words can be modulated by meaning frequency and unequal processing difficulties of the noun and the verb meaning. In addition, the syntactic category constraint seemed not very influential during the meaning resolution of Chinese SCA words. Otherwise, the results should have been similar in both experiments. If the syntactic category constraint had been very constraining, VN should have had less obvious ambiguity effect because the subordinate meaning can be easily selected.
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
Identical to Experiment 1, the results of Experiment 2 also agree with the assumptions of the constraint-satisfaction models. Without semantic bias in the preceding context, the syntactic category constraint and the meaning dominance jointly determine lexical ambiguity resolution. For NN and VV, the meaning dominance can guide the ambiguity resolution at first because the syntactic category constraint is useless to distinguish between the alternative meanings. Our results showed no SBE in the first-pass reading in the target region, verifying that only the dominant meaning was activated. However, the dominant meaning is inconsistent with the subordinate-biased semantic information in the succeeding context, contributing to the SBE in the second-pass reading in all analyzed regions. These results are congruent with the findings of Duffy et al. (1988) and Rayner and Duffy (1986), in which equal time was spent in the target region but more time was spent in the subordinate-biased disambiguating region when preceded by the biased ambiguous words than by the unambiguous words, given a neutral preceding context was used. For SCA words, the meaning dominance also influences the semantic ambiguity resolution, though the syntactic category constraint can distinguish between the alternative meanings. Inconsistent with the findings of Folk and Morris (2003) and Jones et al. (2012), our demonstration of instant and long-lasting SBE for VN and relatively-late SBE for NV provides strong evidence against the syntax-first models.
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
The syntactic category constraint in the preceding context cannot mediate the semantic resolution of Chinese SCA words. In addition, the SBE for VN, compared to that in Experiment 1, occurred earlier in the target region and lasted longer in Experiment 2, suggesting that a lack of semantic information would lead readers to experience greater difficulty in resolving the semantic ambiguity of SCA words. This finding supports the constraint-satisfaction models and corresponds to that of Lee and Federmeier (2009), in which a sustained frontal negativity elicited by SCA words was largely reduced when additional semantic information was available in contexts.
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
116
General Discussion
The present study aims to investigate whether the syntactic category constraint determines the semantic resolution of Chinese SCA words and whether syntactic category of alternative meanings of Chinese homographs influences the SBE during lexical ambiguity resolution. In Experiment 1, using a semantically- and syntactically- biased preceding context, we demonstrated a delay of SBE for lexical-semantic (NN/VV homograph) and syntactic category (VN/NV homograph) ambiguous words.
The SBE for the four types of homographs was delayed until the second-pass reading in the post-target region. In addition, the SBE for SCA words was modulated by syntactic category of the alternative meanings. NV had the largest SBE among the 4 types in the target region, while VN had the smallest SBE among the 4 types in the post-target region. In Experiment 2, we used a semantically-neutral but syntactically-biased preceding context and provided disambiguating semantic information in the succeeding context. The results showed that the SBE for VN appeared from the first-pass reading in the target region and lasted until the second-pass reading in the target and post-target region, while the SBE for the other types of homographs did not occur until the second-pass reading in all analyzed
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
regions. The SBE effects of Chinese SCA words found in different measures and regions in two experiments apparently favor the constraint-satisfaction models.
Meaning frequency and syntactic category of the meanings conveyed in a homograph mainly determine the speed of lexical ambiguity resolution. The contextual constraint from semantic rather than syntactic category can influence the lexical resolution during the reading of sentences.
5.1 Syntax-first or Constraint-satisfaction
Our findings support constraint-satisfaction models. In Experiment 1, consistent with the findings in the study of Folk and Morris (2003), the SBE for SCA words was not found in the target region but was instead delayed until the second-pass reading in the post-target region. Folk and Morris suggested the absence of SBE for SCA words in the target region indicate the prior syntactic category constraint can mediate the semantic resolution of SCA words. Nevertheless, this suggestion cannot explain why the SBE for SCA words was still present in the post-target region, which provides evidence against syntax-first models. Under the assumption of constraint-satisfaction models, the absence of SBE for SCA words in the target region might result from strongly-biased semantic and syntactic category constraints. That is, the preceding constraints are so strong that the activation of dominant meaning is delayed. As a
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
result, in Experiment 2, when the preceding context provided only syntactic category constraint, the SBE for SCA words was present instantly in the target region, indicating the preceding semantic constraint indeed plays a role during the resolution of SCA words.
The previous eye-movement studies reported inconsistent findings when SCA words are preceded by syntactic category constraint alone (Folk & Morris, 2003;
Jones et al., 2012; Stites, Federmeier, & Stine-Morrow, 2013). One possible reason is that the case in which noun is biased and the case in which verb is biased are lumped together in the study of Jones et al. (2012). Given that the access of verb meaning is not as fast as that of noun meaning, the meaning resolution of SCA words would be different in the two cases. Thus, the findings in Experiment 1 of Jones et al. (2012) that the SBE did not emerge on GD in the target region might be simply because the noun and the verb meaning of SCA words were used half and half in each condition.
Another reason is that different kinds of “semantically-neutral” context are employed which may still differ at the degree of message-level information. For instance, in the studies of Folk and Morris (2003) and Jones et al. (2012), plausible context but semantically-neutral toward either meaning of SCA words were used, which can still provide semantics for sentence comprehension. In contrast, in the study of Stites et al.
(2013), syntactic prose context was used, which consisted of meaningless
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
combination of content words, not providing any semantics for sentence comprehension. It remains possibility that readers employ different reading strategies when they encounter the SCA words in sentences using different kinds of
“semantically-neutral” context.
To sum up, our findings in both experiments support constraint-satisfaction models and disagree with syntax-first models. The dominant meaning can be activated even if it is not favored by the preceding contexts. The syntactic category constraint alone cannot determine the semantic resolution of SCA words. Semantic constraint, meaning frequency, and inherent processing difficulty are primary sources determining ambiguity resolution.
5.2 Syntactic category constraint
The functional primacy of syntactic category assignment has been supported by research on English and Indo-European languages (Frazier, 1979, 1987, 1989; Frazier
& Rayner, 1982; Friederici, 1995, 2002; Friederici et al., 1996; Friederici et al., 1999;
Rayner et al., 1983). For example, when a target word violates both syntactic category and semantic constraints from the preceding context, N400 was not observed; instead, ELAN was found around 100-300 ms after target word onset (Friederici et al., 1999).
These studies suggested that syntactic analysis precedes semantic analysis during
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
sentence comprehension. Does syntactic category constraint play a role in Chinese sentence comprehension? Under the assumption of constraint-satisfaction models, the answer is definitely yes. However, syntactic category constraint seems less influential than probabilistic constraints during Chinese sentence comprehension. On the one hand, both Experiment 1 and Experiment 2 demonstrated that the SBE for SCA words cannot be successfully eliminated by syntactic category constraint, suggesting that the dominant meaning can still be activated though being inconsistent with the preceding syntactic category bias. Thus, this finding rejects the functional primacy of syntactic category assignment and supports the influence of probabilistic constraints. On the other hand, both Experiment 1 and Experiment 2 revealed unequal effect sizes of the ambiguity effect for VN and NV homographs. If the syntactic category constraint indeed played a very influential role in the semantic resolution of SCA words, the subordinate meaning would be selectively activated either in the case of VN or NV homographs. The difference between VN and NV homographs indicates that the syntactic category of alternative meanings is more influential than the preceding syntactic category constraint. Our Experiment 2 demonstrated the SBE for strongly-biased VN, indicating that the probabilistic constraints are more influential than the syntactic category constraint during Chinese sentence comprehension.
In addition, several eye-movement and ERP studies have also revealed that the
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
functional primacy of syntactic category assignment was not so obvious in Chinese as compared to in Indo-European languages (Li, 1998; Wang et al., 2013; Yu & Zhang, 2008; Zhang et al., 2013; Zhang et al., 2010). Instead of using SCA words as probes, these studies employed the violation paradigm in which the target word / character either conforms to or violates the preceding context with regard to semantic and syntactic category constraints. Eye movement research showed that the violation effect in both the pure semantic violation condition and the semantic + syntactic violation one did not differ from each other in terms of the appearance time (Yang et al., 2009), indicating that the syntactic processing does not precede the semantic processing in Chinese reading. Similarly, ERP research demonstrated a larger N400 for the semantic + syntactic violation condition in comparison with the pure syntactic violation condition (Zhang et al., 2010), indicating that the semantic integration is not blocked by the syntactic category violation. The authors suggest the lack of syntactic primacy in Chinese might result from language-specific characteristics. Unlike some Indo-European languages (e.g., French, German), Chinese has few morpho-syntactic markers and vague boundary of syntactic category, so it is possible that readers rely on the syntactic category constraints to a less degree during Chinese reading.
In sum, although constraint-satisfaction models allow syntactic category constraints to be considered in Chinese sentence comprehension, some
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
language-specific properties reduce its availability.
5.3 Semantic constraint
The present study verifies that semantic constraint plays an influential role during ambiguity resolution. For lexical-semantic ambiguous words (NN/VV), both Experiment 1 and 2 showed the SBE was absent on the first-pass measures for the target and present on the second-pass measures. Our findings seem inconsistent with those in the previous studies (Duffy et al., 2001; Duffy et al., 1988; 盧怡璇, 2012), in
which the SBE for NN and VV was obviously observed on the measures of initial processing in the target region. One possible reason is that the preceding semantic constraint we used in Experiment 1 might be so strong that the subordinate meaning could be selectively activated. A comparison of the first-pass duration measures between Experiment 1 and 2 may shed light on the strength of semantic constraint (see Table 29). First, regardless of ambiguity, the durations were increased in Experiment 2, indicating readers needed more time to access a word when the preceding context provided no semantic information. Second, Experiment 1 showed a reverse pattern for the ambiguity effect (A < UA), but this pattern was eliminated in Experiment 2. This finding also indicated that the semantic constraint in Experiment 1
which the SBE for NN and VV was obviously observed on the measures of initial processing in the target region. One possible reason is that the preceding semantic constraint we used in Experiment 1 might be so strong that the subordinate meaning could be selectively activated. A comparison of the first-pass duration measures between Experiment 1 and 2 may shed light on the strength of semantic constraint (see Table 29). First, regardless of ambiguity, the durations were increased in Experiment 2, indicating readers needed more time to access a word when the preceding context provided no semantic information. Second, Experiment 1 showed a reverse pattern for the ambiguity effect (A < UA), but this pattern was eliminated in Experiment 2. This finding also indicated that the semantic constraint in Experiment 1