Chapter 1 Introduction
1.2 The issues in Chinese relative clauses processing
國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
processing (e.g., Korean: Kwon, Polinsky, & Kluender, 2004; Japanese: Ueno &
Garnsey, 2008).
1.2 The issues in Chinese relative clauses processing
As a head-final language, Chinese relative clause processing has undergone heated debate. (Examples of Chinese relative clauses are presented in item b of (1) and (2).) Previous experimental research with the self-paced reading task had mixed findings. Lin (2006, 2008) found a preference for SRCs, based on the effects found at the relativizer and the head noun. However, Hsiao & Gibson (2003) indicated that
ORCs are easier, according to the effects observed at the region before the relativizer
DE.
The discrepancy implies that, in addition to head position, more factors should be taken into consideration in Chinese. First, Chinese is unique in embedding the head-final relative clauses in a typical SVO word order, which is the one-and-the-only combination (known so far) in the world (Dryer, 1992, 2003). Besides, word order is the sole structural cue available for readers since Chinese does not have case marking or agreement. The other head-final languages, such as Japanese and Korean, have a SOV typical word order and a case and/or agreement system. Therefore, the
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
distinction between the RC types in Chinese relies only on the gap position. In consequence, SRCs comprise a “V-N” pattern (identified as “V-O”) and ORCs comprise an “N-V” pattern (identified as “S-V”), representing a non-canonical and canonical word order respectively. According to Bever’s (1970) canonical word order account, the non-canonical word order of SRCs may lead to an earlier disambiguation than ORCs. ORCs, with a canonical word order “N-V”, may be interpreted as an “S-V”
main clause. ORCs may be disambiguated when the relativizer occurs, or at least when the head occurs and may also face perspective shifting at the main verb. The head of object relative clauses may further suffer from the memory load resulted from the perspective shift from object to subject when integrating the main clause (MacWhinney, 1977).
Second, although word order in Chinese is the sole structural cue in sentence processing, it is not fixed and tidy; rather, several alternatives may be driven by pragmatics. Accumulating evidence shows that semantic cues, such as plausibility of argument-verb configuration, animacy configuration, thematic role structure, and frequency-bias of verbs, also modulate sentence processing in English (e.g., Garnsey, Pearlmutter, Myers, & Lotocky, 1997b; Traxler & Pickering, 1996; Traxler, Williams, Blozis, & Morris, 2005; J. C. Trueswell, Tanenhaus, & Garnsey, 1994). For example,
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
Traxler, Morris, & Seely’s (2002) eye movement experiments showed that when only one of the two animate nouns in the relative clause can be a plausible agent of the verb, readers speed up the reanalysis caused by the ORC difficulty in English. When the two critical nouns contrast in animacy, an animate head deteriorates the ORC penalty but an inanimate head mitigates the ORC difficulty. In Chinese, animacy was also found to play a relational role as its effects occurred at the second noun and it highly correlated with thematic role assignment (Philipp et al. (2008)). If semantic cues of plausibility from the argument-verb relation play a crucial role in Chinese, then there is a bias of thematic role assignment in sentence processing that is closely related to the real world experience. Certain verbs are inclined to have one argument be the agent and the other the patient. If argument animacy is highly correlated with thematic role assignment, animate arguments bias towards an agent role while inanimate arguments bias towards a patient role (c.f., Valin & LaPolla, 1997). In Chinese, little experimental research has been carried out to explore how and when semantic cues function during the processing of RCs.
Third, temporal ambiguities may come from different sources at different positions as the sentence unfolds incrementally. On the one hand, SRCs and ORCs in Chinese have their respective ambiguities resulting from diverse structural cues to
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
solve at different positions. The “preference” for SRCs or ORCs in processing may not fairly describe Chinese relative clause processing. On the other hand, semantic cues provided by the argument-verb relation or the argument animacy may alleviate or aggravate the structural ambiguities at different positions. The biases relating to thematic role assignment leads readers to develop an interpretation based on real life experience or cognitive conceptions. As the sentence unfolds word by word, such an interpretation “evolves” accordingly and checks with the structural cues to ascertain or to revise. Both structural cues and semantic cues may initiate their influences on processing at different positions in a sentence and lead readers up the garden-path.
In sum, the head-final syntactic property is not the only prominent factor affecting RC processing in Chinese. Chinese is unique in combining this property with SVO basic word order. As an isolating language, it is also distinct in not having a case or agreement system. The fact that it relies solely on word order structurally and that its word order is alterable by pragmatics implies that semantic cues may play a crucial role in sentence processing. Moreover, as readers comprehend sentences incrementally, various structural cues and semantic cues may launch at different positions in the sentence to lead, and sometimes mislead, readers during the course of processing. This thesis aims to address: (a) the syntactic ambiguities of Chinese RCs
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
at different positions in the sentence based on its unique syntactic properties, (b) the influence of the cues of semantic plausibility based on the argument-verb relation in Chinese RC processing, (c) the influence of animacy cues on Chinese RC processing, (d) the ambiguities contributed by both syntactic and semantic information at different positions as readers read incrementally and evolve their interpretations over time.
The following parts of the thesis are organized into four chapters. Chapter 2 reviews previous research on relative clauses from the perspectives of processing and grammar. Chapter 3 and Chapter 4 present two eye-tracking experiments on Chinese RCs with the manipulation of semantic plausibility of argument-verb relation and animacy respectively. The last chapter, Chapter 5, offers a conclusion and general discussion based on the findings of the experiments.
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
Chapter 2
Literature Review
2.1 Processing of Relative Clauses
Relative clauses are a major focus in the sentence processing research because they often induce garden-path effect. The difficulty of processing relative clauses may be attributed to the complexity of the embedded structure. In languages that adopt the gap strategy such as English and Chinese, the relativized NP leaves a gap in the relative clause. This gap should be filled by the head noun in order to understand the sentence. In the following subsections, subject-object asymmetry in relative clause processing is presented in Section 2.1.1. Then, four sentence processing models are introduced in Section 2.1.2. The research that focused on semantic cues in sentence processing is reviewed in Section 2.1.3.
2.1.1 Subject-object asymmetry
In most languages, the noun phrase in the subject and object position can be relativized (cf. Comrie & Keenan, 1979; Keenan & Comrie, 1979). Relative clauses with subject extraction, as in (3), are called subject relative clauses (SRCs). They are
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
RCs with a gap in the subject position. The relative clauses with object extraction in (4) are object relative clauses, with a gap in the object position. In example (3) and (4), English example sentences (a) are accompanied by its corresponding Chinese examples (b).
(3) Subject relative clause (SRC) (subject-extraction)
a. The pastor [RC who hugged the orphan] missed his grandmother very much.
b. [RC 抱住了 孤兒 的] 牧師 很 懷念 祖母。
Baozhule guer de mushi hen huainian zumu.
hug orphan DE pastor very miss grandmother
(4) Object relative clause (ORC) (object-extraction)
a. The orphan [RC who the pastor hugged ] missed his grandmother very much.
b. [RC牧師 抱住 的] 孤兒 很 懷念 祖母。
Mushi baozhu de guer hen huainian zumu.
pastor hug DE orphan Very miss grandmother
Previous research has found a subject-object asymmetry in processing relative clauses. A preference (or priority) for sentential subject is found in typological studies (Comrie & Keenan, 1979; Keenan & Comrie, 1979). A universal tendency called Noun Phrase Accessibility Hierarchy was proposed with a survey on more than 100 languages. The hierarchy ranks from Subject to Object of comparison as follows:
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
Subject, Direct Object, Indirect Object, Oblique Object, Possessor, and Object of
comparison. According to this hypothesized hierarchy, a language which can
relativize a given position in the hierarchy can also relativize all higher positions.
Almost all of the languages under survey are capable of relativizing subjects while noun phrases in other positions are less likely to be relativized. From the perspective of typology, such a universal tendency implies that subject relative clause is easier for sentence processing.
Tavakolian (1981) also observed a preference of children to treat the sentential subject as the subject for both the interpretations of the main clause and the relative clause. Hence, they interpret both subject relatives and object relatives in a common
conjoined clause. For example, “The pastor hugged the orphan and the pastor missed
Grandmother very much.” is interpreted for both SRC “The pastor who hugged the
orphan missed Grandmother very much.” and ORC “The pastor who the orphan
hugged missed Grandmother very much.” Such a subject preference account predicts
that the subject relative is easier than the object relative in head-final languages, as well as in head-initial languages.
In addition, the subject-object asymmetry is also found in psycholinguistic studies. This subject preference for relative clause processing is consistently found not
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
only in head-initial languages such as European languages but also in head-final languages such as Japanese and Korean (e.g., Englsih: Gibson, Pearlmutter, Canseco-Gonzalez, & Hickok, 1996; Korean: Lee, Lee, & Gordon, 2007; Dutch: Mak, et al., 2002; Japanese: Ozeki & Shirai, 2007; German: Schriefers, et al., 1995).
However, studies of Chinese relative clause showed inconsistent results and different theoretical explanations were proposed. This debate on the asymmetry issue in Chinese relative clause processing is discussed in Section 2.3.2.
2.1.2 Sentence Processing Models
In the following subsections, we summarize the theoretical backgrounds for four prominent branches of theories, including syntax-based theories, memory-based resources theories, perspective-shifting theory, and interactionist approaches.
2.1.2.1 Syntax-based theories
The term “garden-path” derives from the best known model of language comprehension, the Garden-Path Theory, developed by Frazier and colleagues (Lyn Frazier & Charles Clifton, 1996; Lyn Frazier & Fodor, 1978; Lyn Frazier & Villiers, 1990). In the model, two stages of sentence parsing mechanism were proposed to
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
explain the ambiguity resolution. In the first stage, the sentence parser establishes an initial interpretation based on the available information for the moment. Then, as more information approaching, the initial interpretation is revised if it turns out to be inconsistent with the information that arrives later. That is, if the initial syntactic analysis fails, a reanalysis is required. Since the parser in the two-stage models (Fernanda Ferreira & Henderson, 1993; Lyn Frazier & Charles Clifton, 1996; Lyn Frazier & Fodor, 1978; Lyn Frazier & Villiers, 1990) uses only the word class information (for example, noun-verb-noun) to assign syntactic position (noun-verb-noun as subject-verb-direct object) in the first stage, a certain types of sentences with specific misleading word orders are subject to, or even doomed to, occasion a garden-path effect.
For example, in a classic example in English The horse raced past the barn fell., Readers are very likely to parse this sentence as an intransitive sentence, but then stumble at the word “fell” and regress to look for other possible structures. This sentence is a reduced relative clause sentence with a passive participle, in which relativizer and the copula were missing. The correct reading of this sentence should be
“The horse (that was) raced past the barn fell.” Such a reduced relative clause without a relativizer often leads to comprehension difficulty.
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
An important hypothesis within the Garden-Path Theory is the Active Filler Strategy (Lyn Frazier & Charles Clifton, 1996; Lyn Frazier & Clifton, 1989). The head noun (as a filler) initiates a search for the gap. To explain the relative asymmetry in English, a subject preference accompanies the filler’s search. Under this circumstance, readers would tend to insert the sentential subject as the subject of the relative clause, since a “doubtful gap” would appear immediately after the relativizer.
When reading the object relative clauses, this subject preference leads to a syntactic misanalysis and the second-stage reanalysis is then required, and so are more reading times.
Based on the syntax of relative clauses, Bever (1970) proposed a canonical word order account that may explain the difference between subject and object relative clauses. In English, the word order of the relative pronoun “who” (denoting the subject, S) plus the verb (V) and object (O) accords with the basic word order SVO in subject relatives. As in the example (5a), who
hugged the orphan is in an order of
S-V-O, just like the canonical word order in English.‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
(5) a. subject relative clause (SRC)
The pastor [RC who hugged the orphan] missed his grandmother very much.
who (S) V O
b. object relative clause (ORC)
The orphan [RC who the pastor hugged ] missed his grandmother very much.
who (O) S V
In contrast, the object relative has a non-canonical word order, with “who”, subject,
and verb in a line (OSV), which leads to greater difficulty. That is, as in (5b), who the
pastor hugged
is in an O-S-V order, which is not the canonical word order. Becauseobject relative clauses in English is different from its canonical word order, more processing efforts is needed and it is more likely to be led down the garden path, compared with subject relative clause. As the examples show, Bever’s (1970) canonical word order account can also explain the subject preference in English.
2.1.2.2 Perspective-shifting theory
The perspective-shifting theory (MacWhinney, 1977) regards the mapping of the syntactic subject onto the reader’s perspective. It is assumed that speakers and listeners prefer to take the viewpoint of active agents, rather than passive recipients.
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
Perspective shifts are time-consuming and costly for working memory. Readers have to shift perspectives in object relative clauses, while the perspective of the sentential subject in subject relative clauses is maintained. For example, readers have to shift from subject perspective to object perspective for the sentential subject ‘the lion’ in the ORC (6b). However, readers do not have to shift the perspective in the SRC (6a).
(6) Relative clause in terms of perspective of the sentential subject
a. SRC: The dog [RC that jumps over the pig] bumps into the lion. (SÆS) b. ORC: The lion [RC that the horse bumps into] jumps over the giraffe. (SÆO)
Therefore, object relative clauses are more difficult than subject relative clauses. The maintenance of an inconsistent viewpoint, memory demands, and syntactic complexity contributes to the difficulty of processing relative clauses.
2.1.2.3 Memory-based resources theories
Daneman and Carpenter (1980) proposed a working memory test especially for reading. Since poor readers may process in an inefficient way, they can maintain less additional information in working memory. Thus, this working memory test for reading, called reading span test, poses heavy demands for processing and storage in order to measure the working memory capacity of reading. In the test, readers were
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
asked to recall the last word of each sentence after reading a set of sentences and doing comprehension questions right after reading each sentence. A set of sentences ranges from two sentences to six. Each reader started with the two-sentence set. The number of sentences in a set increases until the reader was unable to recall the final words. A correlation between the scores of reading span and reading comprehension performance was found in this study. Readers with smaller reading span had a worse performance in reading comprehension.
The correlation between reading span tests and comprehension of relative clauses in English was also established (King & Just, 1991). Individual differences in syntactic processing were found in relative clause reading. Reading the verbs in object relative clauses takes more time for readers with low reading span. Though they spend more reading time, their comprehension is less accurate. This finding supports the Capacity Constrained Model for sentence parsing (also called CC READER) (Just &
Carpenter, 1991). CC READER successfully models the individual difference in the processing of object relative clauses. If the activation limit in the model is set high enough, the model will always make the correct thematic role assignments. If the activation limit is lowered, the model starts to make errors because of the capacity limit.
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
Gibson’s (1998; See also Gibson & Pearlmutter, 1998; Hsiao & Gibson, 2003) Dependency Locality Theory (DLT) predicts ORC difficulty as well, but it is based on a different cause. Increased working memory demands are imposed on processing object relative constructions because the memory cost needed for the intervening discourse referents is greater. These intervening discourse referents are located between the filler and the gap, causing temporarily incomplete dependencies. For instance, the unattached constituents involved in the English ORC, as shown in (7b),
are who, the pastor, hugged. In contrast, the SRC (7a) involves only the relativizer
who between the filler and the gap. Syntactically, the unattached constituents, or the
intervening discourse referents, are on the terminal nodes.
(7) a. subject relative clause (SRC)
The pastor [RC who hugged the orphan] missed his grandmother very much.
filler gap
b. object relative clause (ORC)
The orphan [RC who the pastor hugged ] missed his grandmother very much.
filler gap
According to DLT, the unbounded dependencies result in the cost of storage and integration. On the one hand, storage resources are required to keep track of the syntactic head, so the fact that object relatives carry the filler unattached longer than
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
the subject relatives explains the asymmetry. On the other hand, the effort of connecting an incoming word to its head in the current structure calls for integration resources. Indexical pronouns, such as I, you, and me suffer less integration load.
Gordon et al. (2006) views the effect of indexical pronouns in a slightly different way. It is considered an effect led by the similarity-based interference. That is, indexical pronouns are easier to process because they are distinct from other NPs, such as common nouns. They are from different categories in terms of similarity-based interference. Gordon et al. (2006) demonstrated three categories of the sort: indexical pronoun, short name (e.g., Tom), common noun (e.g., pastor).
Difficulty in object relatives is greatly decreased if the head noun and the noun phrase in the relative clause came from different categories.
2.1.2.4 Interactionist approaches
In addition to syntactic structure of the sentence, there are many other types of information available to reach a correct interpretation of the sentence when readers process a sentence. These sources of information include semantic content of the words in the sentence and pragmatic influences from the discourse. Recent studies have suggested that readers utilize multiple sources of information interactively to
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
constraint interpretations (e.g., Boland, Tanenhaus, Carlson, & Garnsey, 1989; Boland, Tanenhaus, Garnsey, & Carlson, 1995; Gennari & MacDonald, 2008; Pearlmutter &
Gibson, 2001; J. C. Trueswell, et al., 1994; John C. Trueswell, Tanenhaus, & Kello, 1993).
Interactionist approaches propose that syntactic and non-syntactic cues are simultaneously available as soon as each word is recognized. In contrast, the serial parsing accounts, which are assumed by the syntax-based theories, divide the syntactic parsing into two stages. In the first stage, only the syntactic cues are used. It is not until reanalysis is required that the semantic cues such as thematic roles become available in the second stage. An important difference between the serial parsing accounts and the interactionist approaches is that the semantic cues participate in syntactic processing from the beginning.
2.1.3 Semantic cues in Sentence Processing
Previous studies conducting eye-tracking experiments or self-paced reading experiments have shown that semantic cues, such as thematic roles, verb types, verb bias towards a certain structure, plausibility, and animacy of subject and object, do play a role in syntactic processing and may interact with each other or with syntactic
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
cues. For example, Trueswell, Tanenhaus, and Kello (1993) showed that verb bias, namely the frequency with which a particular verb is used in a particular structure, has rapid effects on temporary ambiguity resolution. Garnsey, et al. (1997a) replicated their results and further showed that plausibility of particular verb-object combination interact with verb bias. Wilson & Garnsey (2009) investigated verb bias effects in simple direct object sentences and found evidence that attested the influence of combinational properties of individual words on the earliest stages of sentence comprehension.
In the present study, the influence of plausibility and animacy on relative clause processing is the focus of our interest. Traxler, Morris, & Seely (2002) found that plausibility speed up the recovery from misanalysis and that animacy configuration modulates sentence processing. In their study, Traxler, et al. (2002) manipulated plausibility of argument-verb relation (specifically, the plausibility for both critical NPs or only one of them being the agent of the relative-clause verb) and animacy in
In the present study, the influence of plausibility and animacy on relative clause processing is the focus of our interest. Traxler, Morris, & Seely (2002) found that plausibility speed up the recovery from misanalysis and that animacy configuration modulates sentence processing. In their study, Traxler, et al. (2002) manipulated plausibility of argument-verb relation (specifically, the plausibility for both critical NPs or only one of them being the agent of the relative-clause verb) and animacy in