Chapter 1 Introduction
1.2 Research questions
國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
models or the constraint-satisfaction models can explain lexical ambiguity resolution in sentences.
1.2 Research questions
In the present study, we conducted two eye movement experiments using the normal reading task to investigate the influence of syntactic category and semantic constraints on Chinese lexical ambiguity resolution. In Experiment 1, four types of disyllabic homographs (NN, VV, VN and NV)1 were embedded into sentences involving syntactic category and semantic biases toward the subordinate meaning, while in Experiment 2 only the syntactic category constraints preceded the homographs. Each homograph was assigned a frequency-matched unambiguous word as control, which can fit into the same sentence frame. Specific research questions are listed as follows:
(1) Would syntactic category constraints determine the semantic resolution of Chinese SCA words? The syntax-first (garden-path) models predict the absence of the SBE, while the constraint-satisfaction models predict the presence of the SBE.
(2) Would syntactic category of alternative meanings of Chinese homographs influence the SBE during lexical ambiguity resolution?
1 NN, VV, VN and NV are used as marks for the four types of homographs. The first and the second letter represent the syntactic category of the dominant and the subordinate meaning, respectively. For example, VN represents the homograph whose dominant meaning is a verb and subordinate meaning is a noun.
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
8
Literature Review
2.1 Mechanisms of sentence processing
The primary task of comprehending a sentence is to integrate plenty of information from different sources (e.g., syntactic structure, word semantics, real-world experience, etc.). Some researchers have been debating how syntactic and semantic information interact during sentence comprehension. With different assumptions, two primary classes of models were proposed: syntax-first models and constraint-satisfaction models. The following sections introduce the two models and review some empirical evidence.
2.1.1 Syntax-first models
The most influential model of syntax-first accounts is the garden-path model (Frazier, 1979, 1987, 1989; Frazier & Rayner, 1982; Rayner et al., 1983), which was originally proposed to explain the resolution of syntactic ambiguity. In this model, sentence comprehension is assumed to engage two serial processing stages. At the first stage, sentence comprehension relies on construction of the simplest sentence structure, which is based on the syntactic category information. More importantly, the
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
mechanism of syntactic processing is modular, impervious to non-syntactic contextual information (e.g., lexical-semantic, pragmatic information). Therefore, the language processor considers only available structural information to generate a single candidate structure. At the second stage, non-syntactic information is taken into consideration. If the initial candidate structure turns out to be inconsistent with non-syntactic information that is processed during the second stage, reanalysis occurs and leads to processing difficulty.
Rayner et al. (1983) conducted an eye-tracking experiment and supported the idea that semantic and pragmatic information cannot influence the initial syntactic analysis during sentence comprehension. In their experiment, they manipulated the relative likelihood of possible real world events (i.e., the plausibility of sentences) and the appearance of relative pronoun, constructing four types of sentences: (1) reduced implausible: The dealer sold the car wasn’t sure that it was safe.; (2) reduced plausible: The teenager sold the car wasn’t sure that it was safe.; (3) unreduced plausible: The teenager who was sold the car wasn’t sure that it was safe.; and (4) active implausible: The teenager sold the car but wasn’t sure that it was safe.. In terms of the total reading time (per character) on whole sentence, they found no differences between the reduced implausible and the reduced plausible sentences, while both types of reduced sentences were read longer than were the unreduced
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
plausible and active implausible sentences. In addition, the reading times on the unreduced plausible sentence were longer than on the active implausible sentences.
They further discovered that the longer reading times on both types of reduced sentences resulted from their longer reading times in the disambiguating region (e.g., wasn’t sure that it was safe), which indicated that readers initially constructed the
simplest structure (i.e., the active sentence structure) and encountered processing difficulty when realizing the sentence was a reduced relative-clause sentence. And the lack of increased reading times on the implausible sentences suggested that the plausibility did not influence the initial syntactic analysis. In sum, the study of Rayner et al. (1983) supported the idea that during sentence comprehension the initial syntactic analysis was independent and impervious to semantic analysis.
Recently, Friederici et al. (1996) and Friederici (1995, 2002) also proposed a neurocognitive model of sentence processing on the basis of electrophysiological data.
In this model, three phases are proposed to specify the time course of syntactic and semantic processes (see Figure 1). During phase 1, syntactic structure is constructed based on syntactic category information. Violation of syntactic category information usually elicits an early left-anterior negativity (ELAN) around 100-300 ms after target word onset (Friederici, Pfeifer, & Hahne, 1993; Hahne & Friederici, 1999; Hahne &
Jescheniak, 2001). During phase 2, morphosyntactic and semantic information is
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
processed. Words that are semantically inconsistent with the prior context usually elicit a larger negative brain potential (N400) (300-500 ms), peaking around 400 ms (Kutas & Federmeier, 2000; Kutas & Hillyard, 1984; Neville, Nicol, Barss, Forster, &
Garrett, 1991). Finally, phase 3 represents integration and reanalysis of various types of information. Both syntactically anomalous and garden-path sentences usually elicit a positive brain potential (P600), peaking around 600 ms (Hagoort, Brown, &
Groothusen, 1993; Osterhout & Holcomb, 1992; Osterhout, Holcomb, & Swinney, 1994). Interestingly, when a word violates both syntactic category and semantic constraints in a sentence, only an ELAN but no N400 appears (Friederici et al., 1999).
This finding supports syntax-first models, in which syntactic analysis is assumed to precede semantic analysis.
Figure 1. A simplified diagram of Friederici’s auditory sentence processing model (adapted from Friederici (2002), Figure 1)
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
2.1.2 Constraint-satisfaction models
An alternative class of models, constraint-satisfaction models, is proposed to challenge syntax-first models in the 1990s (Macdonald, 1993; Macdonald et al., 1994;
Macdonald & Seidenberg, 2006; Trueswell & Tanenhaus, 1994; Trueswell et al., 1994;
Trueswell et al., 1993). According to constraint-satisfaction models, two stages of sentence comprehension are unnecessary. Sentence processing engages one mechanism, in which all sources of information are available to interact among one another. Non-syntactic constraints need not wait until a second stage to influence ambiguity resolution. One of the important assumptions in constraint-satisfaction models is that multiple candidates are activated initially and weighted by probabilistic constraints (e.g., frequency, plausibility, etc.). When the probabilistic constraints are strongly consistent with one analysis, processing is easy because no selection is needed. In contrast, when multiple analyses are equally supported by constraints, processing difficulty takes place due to the competition between multiple candidates.
Trueswell et al. (1994) conducted an eye movement experiment and suggested that the lexical-semantic information (e.g., thematic-role) can be used by readers immediately during comprehending relative clauses. They manipulated the animacy of subject noun (animate: defendant vs. inanimate: evidence), the relative clause type (reduced vs. unreduced), and the type of verb in the relative clause (ambiguous
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
regular verb: examined vs. unambiguous irregular verb: drawn). Comparing reduced- with unreduced-relative clauses with animate subject nouns (e.g., reduced: The defendant examined by the lawyer turned out to be unreliable.; unreduced: The
defendant that was examined by the lawyer turned out to be unreliable.), they found
longer first-pass reading times for reduced-relative clauses in the disambiguating region and longer second-pass reading times both in the verb and the disambiguating regions. These results indicated that readers began to encounter processing difficulty because they mistook the verb as a main verb initially. For inanimate subject nouns, neither first-pass nor second-pass reading times showed significant differences between reduced- and unreduced-relative clauses in any region. Thus, there seemed no processing difficulty with the inanimate nouns. Nevertheless, Trueswell et al.
discovered that their inanimate stimuli varied in patient-typicality (i.e., whether the noun is a good patient or theme with regard to the verb following). For example, their norming data of patient-typicality indicated the textbook－loved pair obtained a low score, resulting in weak semantic fit, while the evidence－examined pair received a
high score, resulting in strong semantic fit. Further analyses of first-pass reading times revealed that the lack of processing difficulty remained only in the case of strong semantic fit. In the case of weak semantic fit, reduced-relative clauses required longer first-pass reading times than did the unreduced both in the verb and the
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
disambiguating regions. Therefore, the study of Trueswell et al. (1994) suggested the influence of semantic information in an early stage of sentence comprehension.
In sum, the two primary models of sentence processing have been proposed to elucidate how syntactic and semantic information interact during sentence comprehension. Nevertheless, most evidence comes from the findings of syntactic ambiguity resolution. It remains unclear whether these sentence processing models can explain the resolution of lexical ambiguity. In the next section, some crucial issues and findings of lexical ambiguity resolution are reviewed first, and then some implications of distinct sentence processing models for lexical ambiguity resolution are put forward.
2.2 Issues of lexical ambiguity resolution
Over the past three decades, psycholinguists and neurolinguists have been interested in lexical ambiguity resolution by using various methodologies, such as cross-modal priming, eye-tracking and event-related potentials (ERPs) methods. A primary issue concerns whether multiple meanings of an ambiguous word are activated at the same time. Previous research has demonstrated that meaning dominance, syntactic category and contextual constraints have an impact on semantic resolution of ambiguous words.
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
2.2.1 Meaning dominance
Meaning dominance was defined as the relative frequency of usage of alternative meanings involved in an ambiguous word. It can be assessed by using the free-association technique, in which subjects write down the first meaning that came to mind when given the ambiguous word in isolation (Gawlickgrendell & Woltz, 1994;
Hogaboam & Perfetti, 1975). Based on meaning dominance, two types of homographs can be differentiated: balanced homographs and biased homographs. Balanced homographs have multiple meanings whose frequencies of usage are equal, while
biased homographs have one dominant (high-frequency) meaning and other
subordinate (low-frequency) meanings. For example, “同志” is a balanced Chinese homograph with two equally-frequent meanings (comrade/homosexual), and “儀表”is
a biased Chinese homograph with a dominant meaning (appearance) and a subordinate meaning (instrument panel).
Previous research has demonstrated that meaning dominance would influence the activations of alternative meanings of homographs (Carpenter & Daneman, 1981;
Duffy et al., 1988; Rayner & Duffy, 1986). In a neutral context (i.e., the context where no disambiguating information is provided), the alternative meanings of a balanced homograph were activated simultaneously (Rayner & Duffy, 1986; Swinney, 1979), while the dominant meaning of a biased homograph was retrieved prior to the
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
subordinate meaning (Hogaboam & Perfetti, 1975; Simpson, 1981). Swinney (1979) investigated the semantic resolution of balanced homographs by using a cross-modal priming paradigm with a lexical decision task. In the cross-modal priming paradigm, subjects listened to sentences containing lexical ambiguities (e.g., bugs) followed by a visually-presented target (e.g., contextually-appropriate: ant;
contextually-inappropriate: spy; unrelated: sew). Their task was to determine whether the visually-presented target formed a word or not. The results demonstrated facilitated lexical decisions for the visual targets immediately following the end of the homographs, no matter when the target was related to the contextually-intended meaning or the contextually-unintended meaning of the ambiguity. Also using the lexical decision task, Simpson (1981) examined meaning activation of biased homographs in Experiment 1 by presenting the homographs in isolation (e.g., bank), which were immediately followed by two types of target words (e.g., dominant-related: money; subordinate-related: river). In addition to the prime-target related pairs, prime-target unrelated pairs (e.g., prime: calf; dominant-unrelated targets: money; subordinate-unrelated target: river) were used as controls. Compared to the unrelated controls, only when the target word was related to the dominant meaning was the lexical decision latency facilitated. This study supported the ordered access model of lexical ambiguity resolution (Hogaboam & Perfetti, 1975),
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
suggesting that biased homographs were resolved on the basis of the relative frequency of the meanings when there was no disambiguating context. The retrieval of the dominant meaning was faster than that of the subordinate meaning.
2.2.2 Syntactic category
Words from different syntactic categories are linguistically, psychologically and neurologically distinct. Some electrophysiological and imaging studies have reported distinct brain responses (ERPs / activation areas) to nouns and verbs (Federmeier, Segal, Lombrozo, & Kutas, 2000; Liu, Hua, & Weekes, 2007; Rösler, Streb, & Haan, 2001; Yokoyama et al., 2006). For example, Federmeier et al. (2000) found ERPs elicited by unambiguous nouns were more negative than those elicited by unambiguous verbs between 250 and 450 ms. In addition, unambiguous verbs preceded by appropriate syntactic context elicited a left-lateralized frontal positivity which was not found for unambiguous nouns. Rösler et al. (2001) also discovered that N400 elicited by nouns had a larger amplitude than that elicited by verbs, and that reaction times were shorter to noun targets than to verb targets in their primed lexical decision task. These findings seem to suggest that nouns and verbs are processed differently in the brain.
As reviewed by Vigliocco, Vinson, Druks, Barber, and Cappa (2011), it may take
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
more efforts to process verbs than to process nouns in sentences. Semantically, typical nouns often refer to objects, while verbs are usually used to express actions where participants need to integrate several events. Syntactically, verbs need to assign thematic roles (e.g., agent, patient, theme, etc.) to other words (often nouns) in sentences. Morphologically, verbs are more complex and have more inflections than nouns, especially in the Indo-European languages (e.g., German, French, etc.).
In terms of behavioral differences, a lot of empirical research has reported a processing advantage of nouns over verbs, i.e., nouns were processed faster than verbs.
This processing advantage of nouns were repeatedly replicated in different experimental tasks, such as lexical decision (Rösler et al., 2001; J. A. Sereno, 1999; J.
A. Sereno & Jongman, 1997), noun/verb categorization (J. A. Sereno, 1999), and semantic categorization (Tyler, Russell, Fadili, & Moss, 2001). For example, J. A.
Sereno and Jongman (1997) controlled overall frequency of occurrence and number of letters for unambiguous pure nouns and unambiguous pure verbs in a lexical decision task and found the latencies to nouns were significantly shorter than those to verbs whenever the stimuli were monosyllabic or disyllabic. Their results indicated that processing verbs was more demanding than processing nouns, which they attributed to differences in inflection structure between nouns and verbs. According to them, in English nouns were used with its bare form more frequently than verbs, so
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
the responses to nouns were faster than to verbs when the bare form was presented.
Moreover, J. A. Sereno (1999) also replicated the processing advantage of nouns both in a noun/verb categorization task and a lexical decision task. In their experiments, nouns and verbs with either high- or low-frequency were presented to either the left or right visual fields of subjects. Both tasks revealed very similar results: the responses to nouns were significantly faster than those to verbs; however, in the noun/verb categorization task, the advantage of nouns vanished when the stimuli had low-frequency. In addition, the hemispheric difference was shown only for verbs: the responses to verbs were significantly faster in the case of right visual field compared to the case of left visual field.
Other research investigated reasons for the processing difference between nouns and verbs. For instance, Kauschke and Stenneken (2008) used two types of German nouns (man-made vs. biological objects) and verbs (transitive vs. intransitive verbs) with their bare form in a visual lexical decision task. In addition to the processing advantage of nouns over verbs, their results also showed that the responses to intransitive verbs were significantly faster than those to transitive verbs, indicating the influence of syntactic factor since transitive and intransitive verbs differ in their argument structure. In a second experiment, they used inflected word forms by adding plural-suffixes and personal-suffixes to nouns and verbs respectively, constructing
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
three subsets (e.g., NELK-EN—RUPFEN: pinks—(to/we/they) pick;
KLUB-S—SIEG-T: clubs—(he/she) wins; HENGST-E—HÜPF-E: stallions—(I) bounce). Their reaction time data again demonstrated the noun advantage in all subsets, suggesting that the noun-verb discrepancies in processing cannot be attributed to word form and morphological complexity.
In addition, from the perspective of language acquisition, the evidence that verbs are acquired later than nouns may reflect the psychological complexity of verbs (Gentner, 1982, 2006). Evidence from eye movement showed that the main verb in simple active sentences received longer fixation durations than other grammatical elements (e.g., function word, the subject, the object, etc.), indicating readers paid more visual attention to the main verb in sentences (Rayner, 1977). Therefore, it is reasonable to consider more processing loads for verbs than for nouns during sentence comprehension.
Actually, the larger difficulties of processing verbs compared to nouns also influence the resolution of ambiguous words. Pickering and Frisson (2001) conducted eye-tracking experiments to investigate the processing of verbs, including verbs with multiple meanings (homographs), verbs with multiple senses (polysemy), and unambiguous verbs. The two types of ambiguous verbs were biased in terms of the frequencies of alternative interpretations. They manipulated whether preceding
‧
context or succeeding context containing disambiguating information and whether dominant or subordinate interpretation was supported by the disambiguating information. Thus, in the case of ambiguous verbs, four conditions were formed: (a) supportive-dominant, (b) supportive-subordinate, (c) neutral-dominant, and (d) neutral-subordinate. For unambiguous verbs, the presence of disambiguating information in preceding context was the only variable (supportive vs. neutral). For the homographic verbs, they found no context and meaning frequency effects in the measures of initial processing (i.e., first-pass time and first-pass regressions) in the verb region; instead, in the region immediately following the verbs, the context effects were observed on first-pass time and the meaning frequency effects emerged on both total time and second-pass time. Table 1 shows their findings for homographic verbs.
Table 1
The summarized findings for homographic verbs from Pickering and Frisson (2001)
Measures
(Gaze duration) Context Context &
Frequencyb Context First-pass
regression Context
Second-pass time
(Rereading time) Context Frequencya
Total time Context &
Frequencya
Note. Context: neutral > supportive; Frequencya: subordinate > dominant; Frequencyb: subordinate < dominant
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
Pickering and Frisson (2001) suggested that the resolution of homographic verbs was delayed in comparison with that of homographic nouns, since the SBE was not observed in the verb region but in the following region. The delayed resolution allowed alternative meanings of homographic verbs to reach a high level of activation.
As a result, the meaning frequency effects did not occur immediately in the verb region. The context effects observed in Region 3 indicated a selection process between multiple meanings. The late occurrence of the meaning frequency effects suggested that the dominant meanings were easier to be integrated into the sentence context than the subordinate meaning.
In addition to ambiguous verbs, another type of homograph involving a verb meaning is syntactic category ambiguous words (SCA words), whose multiple meanings cross different syntactic categories (e.g., watch). People may encounter syntactic category ambiguities when there are SCA words in the sentences. Previous literature on the resolution of syntactic category ambiguities yielded different findings.
Frazier and Rayner (1987) proposed a delay model: When no disambiguating
Frazier and Rayner (1987) proposed a delay model: When no disambiguating