Antecedent Priming at Trace Positions in Japanese LongDistance Scrambling
Yoko Nakano Department of Language and Linguistics, University of Essex, UK email@example.com
Claudia Felser Department of Language and Linguistics, University of Essex, UK firstname.lastname@example.org
Harald Clahsen Department of Language and Linguistics, University of Essex, UK email@example.com
Published in Journal of Psycholinguistic Research 31 (2002), 531-571
? Plenum Press, 2002.
Address for correspondence: Harald Clahsen Department of Language and Linguistics University of Essex Wivenhoe Park Colchester CO4 3SQ UK Phone: +44 1206 872228 Fax: +44 1206 873598 E-mail: firstname.lastname@example.org
We report the results from three cross-modal lexical decision experiments investigating antecedent priming effects in Japanese. In the first two experiments we examined antecedent reactivation at the preverbal trace position in long-distance scrambling sentences. We found an interaction between the participants' working memory (WM) span and antecedent priming. For the High Span Group, the magnitude of antecedent priming at the trace position was significantly larger than at the earlier control position; for the Low Span Group, on the other hand, there was no such difference. In a third experiment, we examined whether similar reactivation effects could be observed for argument expressions that are not base-generated adjacent to the verb. Contrary to scrambled objects, subject NPs in canonically ordered sentences were not reactivated at the preverbal test position in either of the two participant groups. We argue that the priming effect observed in the High Span Group supports a trace-based account of long-distance scrambling. The degree of complexity of the experimental sentences was such, however, that they exceeded the memory span of the Low Span Group. We conclude that argument traces access their antecedents irrespective of the position of their subcategorizers.
In generative-transformational theories of grammar (Chomsky 1981, 1995), syntactically displaced constituents are assumed to be linked to their original structural position through a movement chain, with the original position hosting a phonetically unrealized 'trace' of the moved element. For displaced arguments, the trace at the foot of the chain marks the position at which they are assigned a thematic role by their subcategorizer. Under this view, the mental representation of an interrogative sentence such as (1a) below includes a trace (marked ti ) of the moved constituent which book in direct object position, as indicated in (1b).
Which book did Mary say John had read? [which book]i did Mary say [ John had read ti ]
Within Chomsky's (1995, and later) minimalist framework, movement traces are conceived of as silent copies of their antecedents. 1 Several researchers have argued that empty categories (or syntactic 'gaps') form part of the processing of sentence structure by showing that at trace sites, the moved constituent is reactivated (Bever & McElree 1988, Swinney et al. 1987, Nicol 1993, Clahsen & Featherston 1999, among others). The Trace Reactivation Hypothesis (TRH) has been contested by those who argue that antecedent reactivation is triggered by lexical rather than structural information. According to the Direct Association Hypothesis (DAH), antecedent reactivation effects are best explained in terms of direct associations between a subcategorizing head and its arguments (Pickering & Barry 1991, Pickering 1993, Sag & Fodor 1994, Traxler & Pickering 1996). Under this view, as soon as the subcategorizer is processed, a displaced argument will be linked to it directly, without the mediation of traces. The DAH is consistent with lexicalist syntactic frameworks
such as Head-Driven Phrase Structure Grammar (Pollard & Sag 1994) insofar as these theories do not postulate movement traces as part of the structural representation of sentences. Although lexicalist hypotheses such as the DAH, and the structure-based TRH make rather different assumptions about the way syntactic dependencies are processed, the two types of hypotheses turn out to be difficult to distinguish empirically. According to the TRH, encountering a potential 'filler' (such as which book in  above) will trigger the prediction of a syntactic gap. The filler will then be kept in short-term memory until a suitable gap is found, where the moved item is reconstituted. 2 For head-initial languages such as English, however, it is virtually impossible to determine whether such antecedent reactivation effects are due to the existence of traces, or result from lexically-driven direct association, since a putative object trace will normally be located immediately after the verb, which is also the lexical subcategorizer (compare, for instance, Gibson & Hickok 1993, Sag & Fodor 1994). Hence, in order to determine whether or not filler-gap dependencies are established during online sentence processing, cross-linguistic investigation is necessary. Specifically, examining the processing of languages in which subcategorized arguments originate before, rather than after, their subcategorizer should help resolve the above psycholinguistic controversy (see e.g. Clahsen & Featherston 1999, Muckel & Pechmann 2001, for German). In head-final languages like Japanese, for instance, objects precede the subcategorizing verb, so that according to a trace-based account, we would expect reactivation of a dislocated (or 'scrambled') direct object before the verb is encountered. If, on the other hand, a displaced element is linked directly to its subcategorizer, we expect to find no reactivation effects until after the verb has been processed. The purpose of the present study is to examine whether displaced objects in Japanese longdistance scrambled sentences are reactivated at the preverbal trace position. We will also discuss the implications of our findings for the linguistic analysis of Japanese.
Object scrambling in Japanese
Japanese is a head-final language with a relatively free word order. In clauses headed by a ditransitive verb, the canonical order is subject (S) - indirect object (IO) - direct object (DO) verb (V). The three argument noun phrases (S, IO, and DO), however, can, in principle, appear in any order. The operation that changes the canonical ordering of the verb's arguments to a non-canonical ordering is called scrambling, and sentences that display this type of non-canonical word order are called scrambled sentences. 3 In the generative-transformational literature, it is usually assumed that scrambling configurations are derived by movement. Like all syntactically displaced constituents, scrambled arguments will leave behind a trace at the extraction site (compare e.g. Saito 1985, 1992; Mahajan 1990, 1994; Ueyama 1994, Nemoto 1995). 4 In Japanese, we can distinguish two main types of scrambling operation, local and non-local. Local, or clause-bound, scrambling is also known as short-distance (SD) scrambling, whereas scrambling that crosses one or more clause boundaries is referred to as long-distance (LD) scrambling (Harada 1977, Saito 1985, 1992). The two types of scrambling are illustrated in (2b-f) and (3b), respectively. The particles attached to the noun phrases are case markers: The suffix -ga is a nominative case marker, -ni is a dative case marker, and -o marks accusative case.
Short-distance Scrambling a. John-ga Tom-ni banana-o ageta (koto).
John-NOM Tom-DAT banana-ACC gave (fact) “(the fact that) John gave a banana to Tom.” b. c. John-ga banana-oi Tom-ni ti ageta (koto). Banana-oi John-ga Tom-ni ti ageta (koto).
d. e. f.
Tom-nik John-ga tk banana-o ageta (koto). Tom-nik banana-oi John-ga tk ti ageta (koto). Banana-oi Tom-nik John-ga tk ti ageta (koto).
Long-distance Scrambling a. Tom-ga Mary-ni [ John-ga sono saru-ni banana-o
John-NOM that monkey-DAT banana-ACC
gave-COMP tell-PRES-PROG (fact)
“(the fact that) Tom is telling Mary that John gave the banana (to) that monkey.”
banana-ACC Tom-NOM Mary-DAT
John-NOM that monkey-DAT
gave-COMP tell-PRES-PROG (fact)
Lit. “(the fact that) the banana, Tom is telling Mary that John gave (to) that monkey.”
While opinions have recently become divided as to whether short scrambling configurations result from syntactic movement or whether they are base-generated (see Tonoike 1997, 6
among others), long-distance scrambling out of finite clauses exhibits several characteristics typical of A'-movement (Mahajan 1990, Saito 1992, Tada 1993). 5 Observe, for instance, that in (4) below the scrambled object karera-o 'them' fails to qualify as a binder of the anaphor otagai 'each other' despite the fact that it c-commands the anaphor (example from Saito 1992: 76).
*Karerai-o [ Masao-ga
[ Hanako-ga ti hihansita to ] itta ]
they-ACC Masao-NOM each other's teacher-DAT Hanako-NOM criticized that told "Them, Masao told each other's teacher that Hanako criticized."
Example (5) illustrates Saito's (1992) and Tada's (1993) observation that a long-distance scrambled quantified object is unable to take scope over the matrix subject. That is, the existential quantifier dareka 'someone' unambiguously takes scope over the universal quantifier daremo 'everyone', contrary to what one might expect given the relative hierarchical prominence of the latter.
[ John-ga ti aisiteiru to ] itta ] that said
everyone-ACC someone-NOM John-NOM love "Everyone, someone said that John loves."
(someone > everyone)
Along with other so-called reconstruction effects, the above examples suggest that longdistance scrambling is effectively undone at Logical Form, the level of representation relevant to sentence interpretation. The observation that long-distance scrambling is semantically vacuous thus constitutes linguistic evidence that long-distance scrambled constituents are
indeed eventually reconstituted at their original position. This observation by itself does not, however, tell us anything about the position of the trace relative to the verb, or about the role of verb information in the formation of long-distance filler-gap dependencies. Investigating the processing of scrambled sentences in head-final languages such as Japanese makes it possible to distinguish the TRH and lexicalist accounts such as the DAH empirically. As in Japanese the direct object gap precedes the verb, the Trace Reactivation Hypothesis predicts that the antecedent should be reactivated immediately after the embedded indirect object (i.e., the second dative NP) has been processed. The Direct Association Hypothesis, by contrast, does not predict any priming effects at the preverbal position. Thus our main research question is whether in Japanese long-distance scrambling structures, a displaced direct object is reactivated at the hypothesized trace position. 6
Previous psycholinguistic studies on Japanese scrambling
Previous research on the processing of scrambled sentences has focused mainly on shortdistance scrambling, and has produced mixed results. Yamashita (1997), for example, using the self-paced reading paradigm, found no significant differences in reading times for scrambled versus canonically ordered sentences. Mazuka et al. (2001) and Miyamoto & Takahashi (2002), on the other hand, found evidence that scrambled sentences are more difficult to process than their canonically ordered counterparts. Mazuka et al. report results from self-paced reading and eye-tracking experiments which show that for OSV scrambled sentences, both the reading time and the number of regressive eye movements increased on the nominative NP (i.e., the subject). This suggests that in OSV short scrambling structures, the second argument NP may be a locus of processing difficulty. Mazuka et al. interpret their results as supporting an incremental processing model, and as evidence for a configurational structure of Japanese. They suggest that the observed processing difficulty could be due to the
generation of an additional node for the scrambled constituent, to the presence of an object trace, or both. Also using self-paced reading, Miyamoto & Takahashi (2002) observed increased reading times at the position of the NP preceding the subcategorizing verb - the segment that included the hypothesized object gap - in short scrambling sentences. The authors compared reading times for (embedded) scrambled relative clauses containing a ditransitive verb and their canonically ordered counterparts, which corresponded to the schematic forms shown in (6) and (7), respectively.
[CP Subject [RC ecj NPi-ACC NP-DAT gapi ditransitive-verb ] NPj-ACC verb-Comp ] NP-NOM report-verb.
[CP Subject [RC ecj NP-DAT NP-ACC ditransitive-verb ] NPj-ACC verb-Comp ] NP-NOM report-verb.
The slow-down in reading times at the position of the dative-marked object in the scrambled condition (6) demonstrates, according to Miyamoto & Takahashi, that scrambling does indeed lead to a relatively higher processing cost. The authors attribute the observed increase in reading times at the preverbal segment to the increase in working memory load associated with the processing of filler-gap dependencies (compare Gibson 1998, and section 8.2 below). This interpretation gains further support from the results of their second experiment, which showed that reading times on the critical segment became even longer if the distance between the filler and the gap was increased (see Gibson & Warren 1998 for a similar finding on whdependencies in English). Given that the slow-down in reading times occurred prior to the
appearance of the verb, Miyamoto & Takahashi's results cannot be explained in terms of direct lexical association. Studies that aimed specifically to investigate the psychological reality of traces of scrambling have also yielded contradictory results. Nakayama (1995), using probe recognition tasks, failed to find any antecedent priming effect in Japanese short scrambling sentences such as (8b) (the italicized words were used as probes).
Today question-ACC gave
teacher-NOM homework-ACC forgot
Today homework-ACC forgot
student-ACC question-ACC gave
“Today, the teacher who gave the questions (to the student) scolded the student who forgot the homework.“ (Nakayama 1995: 265)
The results from Sakamoto's (1995) end-of-sentence naming study, on the other hand, suggest that traces of scrambled objects can serve as antecedents for null subject PRO in control
structures, which can be interpreted as indirect evidence for the existence of movement traces. Observe, however, that like Nakayama, Sakamoto only obtained the participants' responses at the end of the sentence, so that no conclusions can be drawn from his results as to the exact location of the trace, or the role of verb information during online processing. Moreover, using the dichotic listening technique, Oda et al. (1997) and Ninose et al. (1998) found a different pattern of results from Sakamoto (1995), namely a preference for interpreting the main clause subject (rather than the object, as in Sakamoto's study) as the controller of PRO in both canonically ordered and scrambled sentences. Further evidence for antecedent priming in short scrambling structures has been provided by Clahsen & Featherston (1999) for German. Like in Japanese, VPs in German are standardly assumed to be head-final. In a cross-modal lexical priming task, Clahsen & Featherston found antecedent reactivation effects at the hypothesized preverbal trace position in sentences that contained a scrambled direct object.
In sum, whereas there is some (albeit inconclusive) evidence for trace reactivation in short-distance scrambling constructions, little is known to date about the processing of long-distance scrambling structures such as (3b) above. Note that contrary to short scrambling, long-distance scrambling of this type is not available in German or most other Indo-European languages. From a syntactic perspective, studying long-distance scrambling is potentially more revealing as there is a broad consensus among generative-transformational linguists that longdistance scrambled word orders are indeed derived by syntactic movement. 7 In order to shed more light on
how scrambled sentences are processed, the present study investigates antecedent priming effects in Japanese long-distance scrambling sentences using the cross-modal lexical priming technique.
In the cross-modal lexical priming (CMLP) procedure we used, test sentences are presented to the participants auditorily. While listening to the sentences, participants perform a lexical
decision task on visually presented target words (compare Swinney 1979, Swinney et al. 1987). The CMLP technique has frequently been used to investigate antecedent-reactivation effects in on-line sentence processing (Nicol & Swinney 1989, Hickok et al. 1992, Nicol 1993, Zurif et al. 1994, Love & Swinney 1996, Clahsen & Featherston 1999, among others). Swinney et al., for example, investigated the processing of sentences such as (9) below, which contains an object-relative clause.
The policeman saw #1 [the boy]i #2 that the crowd #3 at the party #4 accused #5 (ti) of the crime.
At various test points during the sentence - including the hypothesized trace position (= #5) a target word would appear on a screen, which was either semantically related (e.g. girl) or unrelated to the head noun of the filler the boy. Participants were asked to judge whether the target words were real words or non-words, and then to press either the 'yes' or the 'no' button as quickly as possible. Lexical decision time is thought to reflect the time it takes for the target word to be recognized. The recognition of semantically related words should be facilitated at the point where the antecedent is retrieved from short-term memory and reactivated, an effect that is commonly referred to as (semantic) priming. One of the advantages of using the CMLP technique to tap into online processing is that participants can process the auditory stimuli at a natural speed. Note, however, that the CMLP procedure has been subject to methodological criticism in the past. Specifically, McKoon et al. (1994) and McKoon & Ratcliff (1994) have argued that apparent antecedent reactivation effects may just be artifacts of the particular method used for selecting control words. That is, the reason why semantically related (as opposed to unrelated) targets often trigger shorter reaction times may simply be that they fit better into the current sentential context, and hence can be integrated more easily than poorly-fitting control words. 12
Nicol et al. (1994), however, suggest that variables other than 'goodness of fit' might have been responsible for McKoon et al.'s failure to replicate earlier results from Nicol & Swinney (1989), such as the possibility that the test points were not optimally chosen, and the fact that McKoon et al.'s stimuli sentences were relatively more difficult to comprehend than those used by Nicol & Swinney. They also point out that the word-by-word visual presentation technique used by McKoon & Ratcliff to back up their claim is not in fact comparable to the CMLP procedure. Love & Swinney (1996) and Nicol et al. (1997) have demonstrated that 'goodness of fit' does not significantly influence lexical decision times in the CMLP task, which led them to conclude that the cross-modal priming technique is indeed sensitive to online syntactic parsing rather than to integration processes. Observe further that in Japanese object scrambling sentences, the hypothesized trace position precedes the verb, so that 'goodness of fit' would be considerably more difficult to determine at this position than it would be for postverbal object traces in head-initial languages like English.
Design and materials
In Experiments I & II, we investigated antecedent priming effects in Japanese long-distance scrambling sentences such as (10) below, in which a direct object is moved from inside an embedded clause across the main clause subject.
[CP/IP futari-me-no hito-ga
And then lemon-ACC
the second person-NOM M.C.-DAT
[CP sono kodomo-ga onna-no hito-ni that child-NOM female person-DAT ^
ti nedatte-iru to ] asking ^^
COMP ] answered]
Lit. "And then, a lemon, the second person answered to the Master of Ceremonies that that child was asking the woman for."
Here, the direct object remon 'lemon' has scrambled out of the embedded clause and has been placed in front of the matrix subject futari-me-no hito-ga 'the second person'. In Experiment III, we examined whether or not in canonically ordered sentences the subject NP would be reactivated at the preverbal test position. The design was adopted from Clahsen & Featherston’s (1999) study. Design and materials used in Experiments I & II were identical. 40 experimental sentences were constructed, all of which were structurally identical to (10) above, with both the embedded clause and the matrix clause being headed by a ditransitive verb (see Appendix 2 for a complete list of the test sentences used in Experiments I & II). The dative-marked NP in the embedded clause (onna-no-hito-ni 'woman' in ) that immediately preceded the trace position was more than 7 morae in length. This was to increase the distance between filler and gap, so as to reduce as much as possible any priming effect of the residue of the first activation. The materials used in Experiment III were different, and will be described in detail in section 7. Pretests including grammaticality judgement and listening comprehension tests were conducted for stimulus selection, and our experimental sentences were selected only from
those materials where error rates had been less than 10 percent. The test sentences were embedded in short pieces of discourse in order to make them sound more natural. The texts were read by a female native speaker of Japanese at normal speed, and recorded on a SONY digital tape recorder. Each text was between 38 and 63 bunsetsu in length, and was semantically and pragmatially coherent. 8 In addition to the experimental sentences, 40 filler and 4 practice sentences were constructed, yielding a total number of 84 short texts (for an example text, see Appendix 1). The test and filler items were mixed and pseudo-randomized. For Experiments I & II, 40 nouns which are commonly written in katakana were chosen as target test words from Wydell (1991). The 40 words comprised 10 groups of 4 words each that were matched with respect to frequency, and number of characters and morae. The target words were counter-balanced across participant groups, that is, if one of the four words was used as an experimental target word in one of the participant groups, then the other three were used as controls in the same group. Each word was used as an experimental target word only once in each participant group. Additionally, a separate set of non-experimental 'filler' targets was constructed comprising 66 real words and 81 non-words written in kana, and 83 real words and 81 non-words written in kanji. The experimental target words were either identical (e.g. remon 'lemon') or semantically unrelated (e.g. sonata 'sonata') to the head noun of the displaced object, and were presented at the hypothesized trace position (marked ^^ in  above) or at a control position (marked ^) 500 ms earlier. Our main reason for choosing identical rather than semantically related target words was in order to eliminate the additional processing step of establishing a semantic association between the trace of the filler and the target word (compare Clahsen & Featherston 1999). Given that according to the copy theory of movement (Chomsky 1995), the 'gap' actually contains a silent copy of the displaced constituent, using identical targets is the most direct way of testing whether or not such a copy forms part of the mental representation of the sentence during online processing. Note that
any amount of priming due to the formal or semantic identity of filler and target will be factored out by comparing priming effects at the trace position with those on the same target word at the control position. Moreover, in order to prevent participants from developing a conscious strategy based on the realization that some target words would appear in the input twice, we used a higher than usual number of filler targets (approximately 85 percent). In the following we will refer to the trace position as the Late Point, and to the control position as the Early Point. If in long-distance scrambling sentences such as (10), the antecedent is reactivated at the putative trace site, lexical decision times to identical target words should be shorter at the Late Point than at the Early Point. For unrelated targets, on the other hand, no such difference between the Early and Late test positions is expected.
134 participants were recruited from the undergraduate and graduate communities of the University of Essex and from the Institute of International Education in London. All of them were right-handed native speakers of Japanese with normal or corrected-to-normal vision who had been educated in Japan up to high school level. They were paid a small fee for their participation. 80 participants took part in Experiment I (59 women, 21 men, mean age 25.3), 54 in Experiment II (40 women, 14 men, mean age 26.7), and 54 in Experiment III (40 women, 14 men, mean age 26.7).
The experiments were carried out in a dedicated room without external disturbance. The participants were given a written instruction in Japanese, in which they were asked to make lexical decisions on visually presented target words. The participants were seated in front of a 17'' monitor, and instructed to listen carefully to the pre-recorded texts over headphones.
Every six to eight seconds, a visual target word appeared on the screen (green colour 36 pt. Mincho, black background), and remained there for 400 ms. The experimental test points were at the offset of the second ni-marked NP, and the control test points were 500 ms prior to them. Participants had to make a lexical decision on the visual target by pressing a green button for the real words or a red button for the nonwords on a push-button box. The presentation of the stimuli and the measurement of reaction times was controlled by the NESU software package (Baumann et al. 1993). To prevent the participants from attending only to the lexical decision task, the tape recorder stopped every 5 to 9 texts, and the participants were asked to write a short summary of the last text they had heard before the tape recorder stopped.
For all experiments, the following types of data were discarded prior to statistical analysis: (i) data from participants who had confused the red and green buttons, who had failed to complete summary writings, or whose overall error rate exceeded 20 percent; (ii) responses involving erroneous lexical decisions; (iii) errors resulting from computer malfunction; and (iv) responses to items for which the overall error rate was larger than 20 percent. Statistical analyses were conducted on the basis of raw reaction times converted into logarithms (note that all reaction times given in milliseconds were calculated by anti-log conversion). In all three experiments, we used the same criterion for outliers. For each condition, reaction times outside the range of two standard deviations from a participant's mean for this condition were excluded from any further analysis. For the remaining data we performed two separate analyses of variance (ANOVAs), one for subjects and one for items. Unless otherwise stated, post-hoc pairwise comparisons were performed using Fishers‘ Least Significant Difference test.
80 native speakers of Japanese took part in Experiment I. As mentioned above, the experimental sentences all corresponded to the schematic form of (11) below, with the embedded direct object NP preceding all other argument NPs.
NP-oi [ NP-ga NP-ni [ NP-ga NP-ni ti V COMP ] V ]
The total number of conditions was four, resulting from two test points and two target types. Due to the absence of any potential source of priming, we expect reaction times for unrelated targets to be longer than for identical ones. Secondly, the Trace Reactivation Hypothesis predicts that the Late-Identical condition should be faster than any other condition. That is, if antecedents are reactivated at trace positions, we expect to find a priming effect for identical target words at the Late Point, but not earlier on in the sentence. The data from seven participants were discarded either because they had confused the green and red buttons (3 cases) or had been unable to complete summary writings (2 cases), or because of computer error (2 cases). The data from ten participants whose overall error rate in the lexical decision task was higher than 20 percent were also removed prior to analysis. The data from the remaining 63 participants were included in the statistical analyses. Erroneous lexical decisions and extreme reaction times (together making up 8.21% of the data) were excluded from the database.
Results Figure 1 presents the mean lexical decision times for all four conditions.
Figure 1: Mean lexical decision times (in ms.) to visual targets for all four conditions
850 800 750 700 650 600 550 500 Early Test Point Late Target Type Unrelated Identical
Two separate two-way ANOVAs were conducted on these data (F1 for subject analysis, and F2 for item analysis), which revealed significant main effects of both Target Type (F1 (1,62)=30.20, p<0.001; F2 (1,37)=13.81, p<0.001) and Test Point (F1 (1,62)=8.30, p=0.004; F2 (1,37)=4.88, p=0.027). That is, lexical decision times for identical targets were shorter than for unrelated ones, and reaction times at the Early position were shorter than at the Late position. The interaction between Test Point and Target Type (F1 (1,62)=0.61, p=0.4358; F2 (1,37)=0.70, p=0.40) did not prove significant though. That is, contrary to the prediction made by the Trace Reactivation Hypothesis, the mean reaction time for the Late-Identical condition was not faster than for the Early-Identical condition.
Post-hoc analyses Further analysis of the data was prompted by the fact that when asked to comment on the experiment afterwards, several participants reported that they had found the sentences very difficult to understand. Interestingly though, not all participants had experienced this difficulty. In order to test how individual differences between participants may have affected 19
our results, we examined the reaction times to the non-experimental real-word targets used in the experiment. Notice that this provided us with a data set that was independent of the experimental constructions used to test antecedent reactivation. A cluster analysis on the mean reaction times for 106 non-experimental target words (excluding those that were identical or semantically related to nouns occurring in the auditory materials presented simultaneously) indicated the existence of two participant groups.
Figure 2: Dendrogram for mean lexical decision times to non-experimental real-word targets
C A S E ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ?
Rescaled Distance Cluster Combine 0 5 10 15 20 25 +---------+---------+---------+---------+---------+ -+ -+ -+ -+ -+-+ -+ I -+ I -+ I -+ I -+ I -+ I -+ +---------------------------------------------+ -+ I I -+ I I -+ I I -+ I I -+ I I -+ I I -+-+ I -+ I -+ I -+ I -+ I -+ I -+-+ I -+ I I -+ I I -+ +-----------+ I -+ I I I -+ I I I -+-+ I I -+ I I -+ I I -+ I I -+ I I -+ I I -+ I I -+ +---------------------------------+ -+ I -+ I -+ I -+-+ I -+ I I -+ I I -+ I I -+ I I -+ I I -+ +-----------+ -+ I -+ I -+ I -+ I -+ I -+-+ -+ -+ -+ -+ -+ -+ -+ -+ -+
In Figure 2, each dot under case denotes each participant, and connected horizontal lines denote combined cases. The actual distances are rescaled to numbers between 0 and 25, preserving the ratio of the distances between steps. The connected two long horizontal lines of the uppermost clusters indicate a large distance between the two clusters, which is a clear sign of the existence of two groups.
Figures 3 and 4 present mean lexical decision times to test target words for the two groups as defined by the above cluster analysis. They show that Group 1 (N=24) produced faster lexical decision times to test targets than Group 2 (N=39). Given the existence of two groups of participants, let us now re-examine the four experimental conditions separately for each group.
Figure 3: Mean lexical decision times (in ms.) to experimental target words for Group 1
850 800 750 700 650 600 550 500 Early Test Point Late Target Type Unrelated Identical
Figure 4: Mean lexical decision times (in ms.) to experimental target words for Group 2
850 800 750 700 650 600 550 500 Early Test Point Late Target Type Unrelated Identical
For both Group 1 and Group 2 we found a significant main effect of Target Type (Group 1: F1 (1,23)=41.99, p<0.001; F2 (1,37)=25.99, p<0.001; Group 2: F1 (1,38)=5.75, p=0.017; F2 (1,37)=5.85, p=0.016), a significant effect of Test Point for Group 1 but not for Group 2 (Group 1: F1 (1,23)=0.87, p=0.35; F2 (1,37)=0.57, p=0.45; Group 2: F1 (1,38)=16.02, p<0.001; F2 (1,37)=18.11, p<0.001). Only for Group 1, however, did we find a significant Test Point x Target Type interaction (F1 (1,23)=4.54, p=0.033; F2 (1,37)=3.99, p=0.046). Pairwise comparisons revealed that for identical targets, Group 1 produced significantly shorter reaction times at the Late test point than at the earlier control position (579 vs. 549 ms, MSe=0.0051, p<0.05). This result indicates that in Group 1, antecedents were reactivated at the preverbal trace position. In Group 2, on the other hand, there was no reliable interaction of Test Point and Target Type (F1 (1,38)=0.18, p=0.67; F2 (1,37)=0.22, p=0.64). Here the LateIdentical condition actually produced significantly longer reaction times than the EarlyIdentical condition (781 vs. 818 ms, MSe=0.0078, p<0.05), indicating that in this group, the antecedent was not reactivated at the hypothesized trace position.
Discussion The results from Experiment I indicate that only one group of participants (Group 1) showed reactivation effects during the processing of long-distance scrambled sentences. In Group 2, on the other hand, there was no evidence for antecedent reactivation. Notice also that in Group 2, overall reaction times were a magnitude longer than those of Group 1, with a mean of 810 ms against 588 ms for Group 1. Given the fairly high degree of syntactic complexity of the construction under investigation, it is conceivable that Group 2 participants found the experimental sentences more difficult to process than Group 1 participants. Observe that our experimental sentences contain several potential sources of processing difficulty. First, consider the distance between the filler and the associated gap. When
processing object-scrambled sentences, the presence of an accusative-marked NP in sentenceinitial position signals to the parser the possibility that this constituent has been displaced. Both the gap-filling account and lexicalist accounts such as the DAH assume that during online sentence processing, a displaced element, or potential filler, will be stored in short-term memory until a suitable gap (or, according to the DAH, the subcategorizing head) is found, at which stage the filler is retrieved from memory and reactivated. It is reasonable to assume that keeping a filler in working memory incurs a processing cost, and that the size of this cost is partly a function of the distance between the filler and the corresponding gap or subcategorizer - possibly, up to a point at which the dependency can no longer be successfully established (compare e.g. Berwick & Weinberg 1984, Gibson 1998). 9 Secondly, note that long-distance scrambling sentences typically exhibit a high degree of syntactic ambiguity. Due to the head-final nature of the language, all our experimental sentences are multiply ambiguous at the outset, and likely to induce garden path effects requiring subsequent reanalysis. On the assumption that at any point during processing, the parser will opt for the simplest grammatically possible analysis (Frazier 1987a, among others), the incoming string of argument noun phrases NP-o - NP-ga - NP-ni will initially be interpreted as belonging to the same clause. However, given the high degree of temporary ambiguity exhibited by free word order languages like Japanese, alternative analyses are conceivable as well. Only the second incoming nominative NP signals to the parser that it is dealing with an embedded rather than with a monoclausal structure, thus possibly triggering reanalysis. Evidence for the existence of such a garden-path effect comes from Yamashita's (1997) finding that in processing embedded structures, reading times on the second nominative NP increase. Third, sentences like (10) above involve centre-embedding, with the complement clause separating the matrix verb from its arguments. The well-known processing difficulty
associated with nested dependencies, or centre-embedding structures, is usually attributed to processor and/or WM overload (Gibson 1998, among others). Working memory resources are required both for the temporary storage of the intermediate products of processing, and as a working space for online computation. Thus, variation in individual working memory capacity might affect the efficiency of processing. King & Just (1991) report that participants with low working memory capacity were slower than participants with high working memory capacity in processing syntactically complex constructions such as object relative clauses in English. They also found that if a structure exceeded a certain complexity limit, processing collapsed altogether. Participants with a high working memory capacity, by contrast, were processing the same constructions faster and successfully. As we saw above, long-distance scrambling sentences tend to make fairly heavy demands on working memory, and hence would seem to constitute something of a challenge for the human parser. This would explain why the occurrence of such sentences in natural corpora is relatively rare (compare Yamashita 2000), and why many speakers find them somewhat unnatural. The results from Experiment I show that in Group 1, the 'fast' group, antecedents were reactivated at the trace position, whereas in Group 2, the 'slow' group, they were not, suggesting a possible breakdown of normal automatic processing in the latter. Specifically, it may have been the case that Group 2 participants could not sustain the prediction of a syntactic gap for long enough, or were unable to retrieve the antecedent from working memory when finally encountering the direct object gap. The purpose of Experiment II was to test the hypothesized interaction between working memory and antecedent reactivation.
The specific hypothesis to be tested in Experiment II was that at least those participants with a high WM capacity should exhibit antecedent reactivation effects in processing long-distance
scrambled sentences. The design, materials and procedure were identical to Experiment I. Only this time, participants were divided into different groups according to their working memory capacity. Working memory spans were measured using the Japanese version of Daneman & Carpenter's (1980) reading span test (Osaka & Osaka 1992, Osaka 1998). 109 Japanese native speakers participated in the test. Test materials consisted of sets of two, three, four or five sentences. The test sentences were displayed on a computer screen one by one. Each of the sentences contained one underlined word. Participants were asked to read the sentences aloud, and to memorize the underlined word in each sentence at the same time. At the end of each set of sentences the participants were required to recall the underlined words in each set in the order in which they had appeared. Reading span scores reflect the maximum number of sentences for which the participants could successfully recall the target words (compare Osaka & Osaka 1992). On the basis of these scores, the participants were divided into three groups: a Low Span Group (mean score ≤ 2.5 words), a Medium Span Group (mean score = 3.0 or 3.5 words), and a High Span Group (mean score ≥ 4 words). The 37 Low Span participants and 17 High Span participants went on to participate in the second cross-modal priming experiment. The data from two participants were eliminated prior to analysis. One had been unable to complete summary writings, and one had confused the green and red buttons throughout the experiment. The responses to one item were removed because of computer error. The data from 52 participants were included in the statistical analysis. Erroneous lexical decisions and outlier responses (together comprising 6.9% of the data) were excluded from further analysis.
Results Figures 5 and 6 present the mean reaction times for High and Low Span Groups for all four conditions.
Figure 5: Mean lexical decision times (in ms) for High Span Group
850 800 750 700 650 600 550 500 Early Test Point Late Target Type Unrelated Identical
Figure 6: Mean lexical decision times (in ms) for Low Span Group
850 800 750 700 650 600 550 500 Early Test Point Late Target Type Unrelated Identical
Two separate three-way ANOVAs were performed on these data, with Working Memory, Target Type and Test Point as main factors. We found significant main effects of Working Memory (F1 (1,50)=570.11, p<0.001, F2 (1,38)=221.33, p<0.001) and of Target Type (F1
(1,50)=41.45, p<0.001, F2 (1,38)=16.39, p<0.001), as well as a significant Working Memory x Target Type x Test Point interaction (F1 (1,50)=9.75, p=0.0018, F2 (1,38)=4.02, p=0.045). Since the interaction among the three main factors had proved significant, further analyses were conducted. Two-way ANOVAs on the results from the High Span Group revealed a significant interaction between Target Type and Test Point (F1 (1,15)=6.105, p=0.026, F2 (1,38)=4.51, p=0.040). There was also a significant main effect of Target Type (F1 (1,15)=23.88, p<0.001, F2 (1,38)=19.57, p<0.001), but no significant main effect of Test Point (F1 (1,15)=0.82, p=0.38, F2 (1,38)=1.417, p=0.24). Pairwise statistical comparisons showed that the Late-Identical condition produced significantly faster reaction times than the LateUnrelated condition (585 vs. 644 ms; MSe=0.0037, p<0.05), and within the Identical condition there were faster responses at the Late test point than at the Early test point (585 vs. 611 ms; MSe=0.0037, p<0.05). For the Low Span participant group, we also found a significant main effect of Target Type (F1 (1,35)=9.205, p=0.005, F2 (1,38)=7.71, p=0.008), and no reliable effect of Test Point (F1 (1,35)=4.179, p=0.049, F2 (1,38)=0.725, p=0.40). In contrast to the results from the High Span Group, however, the overall interaction of the two main factors did not reach significance for the Low Span Group (F1 (1,35)=3.171, p=0.084, F2 (1,38)=4.101, p=0.050). Pairwise comparisons showed that the Late-Identical condition produced significantly slower reaction times than the Early-Identical condition (746 vs. 719 ms; MSe=0.0053, p<0.05). Thus whereas for the High Span Group, reaction times for identical targets were faster at the Late position than at the Early position, the reaction times produced by the Low Span participants were actually slower at the Late position than at the earlier control position. These results indicate that antecedents were reactivated at the purported trace position in the High Span Group but not in the Low Span Group. 10
To further elucidate the differences between the two participant groups, we compared overall response times in the lexical decision task for the High and the Low Span Groups. Pairwise statistical comparisons using independent samples t-tests revealed that for High Span participants, reaction times to test items were significantly faster than for Low Span participants across all four conditions (with mean RTs of 619 vs. 747 ms; t(50)=2.895, p=0.003). To summarize, we found that in High WM Span participants, the reaction times for identical target words were faster at the preverbal test position than at the control position, indicating that there was a priming effect at the trace position. In Low WM Span participants, by contrast, the situation was actually the reverse: Reaction times for identical target words were slower at the trace position than at the earlier control position. For the Low Span Group, then, no antecedent priming effect could be observed. These results are compatible with those from Experiment I. In addition, we found that in our participants, a high WM capacity correlated with overall faster lexical decision times. The interaction between WM capacity and antecedent priming observed in Experiment II confirms the hypothesis that trace reactivation in the long-distance scrambling construction depends upon the participants' WM capacity.
While the priming effects observed in Experiments I & II are consistent with the Trace Reactivation Hypothesis, we cannot completely rule out the possibility that they reflect verbdriven integration processes rather than structural processing. One could argue, for example, that after processing the second dative NP, the parser anticipates the (type of) verb to come and begins reconstructing the anticipated verb's argument structure. In this case, we would expect that other arguments of the verb will be reactivated at the preverbal test position as well. The specific prediction made by the TRH, on the other hand, is that only the displaced
object should be reactivated at the preverbal position, but not any of the verb's other arguments. To test whether the observed priming effect can be attributed to a reconstruction of the verb's argument structure at this point, or to the presence of a preverbal object trace, we examined whether or not in canonically ordered sentences such as (12), the subject NP would be reactivated at the preverbal test position (marked #).
Homon-saki-no byoin-de ooji-ga
kodomotachi-ni omocha-o # age-ta toy-ACC give-PAST
visit-place-GEN hospital-in prince-NOM children-DAT
“In the hospital he visited, the prince gave the children some toys.”
The rationale here was as follows. Since the verb is the lexical subcategorizer for all argument NPs, including the subject, then if the priming effect observed in our experiments was triggered by verb information rather than by the presence of an object trace, we would expect to find a similar priming effect for a 'distant' subject NP. The Trace Reactivation Hypothesis, by contrast, predicts that the subject NP should not be reactivated at the position immediately preceding the verb, since even under the VP-internal subject hypothesis (Koopman & Sportiche 1991, among others) the subject is not assumed to originate - and hence, to leave a trace - between the direct object and the verb. 11 The test sentences and target words were prepared in the same manner as in Experiments I & II. 28 test sentences (7x4 groups) were constructed and were embedded in short texts (a complete list of the critical sentences used in Experiment III is provided in Appendix 3). For ease of comparison with the test sentences used in Experiments I & II, only three-place predicates were used as matrix verbs. Visual targets were written either in katakana only or in kanji, and were either semantically unrelated or identical to the head noun of the subject NPs. 28 words which were semantically unrelated to the head noun of the subject NPs were chosen
from Wydell's (1991) corpus and matched to the identical target words (8 pairs in katakana and 20 pairs in kanji). The semantically unrelated target words written in katakana were matched for frequency and number of morae (three or four, with each mora corresponding to one character). The target words written in kanji were matched for frequency, number of characters (two), and (as closely as possible) for the number of strokes used in writing the characters. 12 The critical test position was at the offset of the accusative-marked object NP preceding the verb, and the control position was located 500 ms earlier. The experimental procedure was identical to that of Experiments I and II. The 54 participants who took part in Experiment III had been classified as either High or Low Span participants on the basis of the memory test described in section 6 above. The data from 4 participants were discarded either because they had confused the green and red buttons (1 case), had been unable to complete summary writings (1 case), or because their overall error rate in the lexical decision task was higher than 20 percent (2 cases). The data obtained from the remaining 50 participants were included in the statistical analysis. Erroneous lexical decisions and extreme reaction times were removed from the database, as well as one item for which the error rate exceeded 20 percent (together these amounted to 2.4% of the data).
Results Figures 7 and 8 present the mean reaction times for the two participant groups in the four experimental conditions.
Figure 7: Mean lexical decision times for High Span Group (in ms.)
850 800 750 700 650 600 550 500 Early Test Point Late Target Type Unrelated Identical
Figure 8: Mean lexical decision times for Low Span Group (in ms.)
850 800 750 700 650 600 550 500 Early Test Point Late Target Type Unrelated Identical
There was no significant interaction of Test Point x Target Type x Memory (F1 (1,48)=0.13, p=0.71, F2 (1,54)=0.56 p=0.45). Significant main effects were found for Working Memory (F1 (1,48)=170.98, p<0.001, F2 (1,54)=93.68 p<0.001) and Target Type (F1 (1,48)=89.01, p<0.001, F2(1,54)=51.39 p<0.001). Test Point x Target Type interactions did not prove significant in either of the groups (F1 (1,13)=1.51, p=0.22, F2 (1,54)=0.67, p=0.41 for the 31
High Span Group, and F1 (1,35)=0.34, p=0.56, F2 (1,54)=0.03, p=0.85 for the Low Span Group), nor was there a significant main effect of Test Point. For both groups, we found a significant main effect of Target Type though, which indicates that in the Identical condition, there was an effect of repetition, as one would expect. In short, there was no priming effect at the preverbal test position, even for the High Span Group that had shown antecedent reactivation for dislocated objects in Experiment II. These results are incompatible with the view that the priming effect observed in Experiments I & II was due to the reconstruction of the verb's argument structure at the preverbal test position. Thus the results from Experiment III support our conclusion that the antecedent reactivation effects observed in Experiments I & II resulted from trace reactivation rather than from lexically-driven integration processes.
General discussion Trace reactivation versus direct lexical association
Our most interesting finding was the priming effect found for displaced direct objects at the hypothesized preverbal trace position in Japanese long-distance scrambled sentences. The view that empty categories form part of the mental representations that are constructed during the online processing of filler-gap dependencies provides a straightforward explanation for our findings. When object-scrambled sentences are processed, the presence of an accusativemarked NP in sentence-initial position signals to the parser that this constituent has been displaced, and triggers the prediction of a preverbal gap. In our experimental sentences, the distance between the filler and its associated gap is rather long, with the scrambled object having crossed two clause boundaries. As further incoming material is processed, additional processing cost is incurred by the high degree of structural ambiguity of the incoming string of argument NPs, and by the centre-embedding structure. Thus it is conceivable that the
degree of complexity of our test sentences was such that it led to processor overload in some individuals - those with a relatively lower WM capacity. Lexicalist accounts of antecedent reactivation such as the DAH, by contrast, are unable to explain our findings, since the observed reactivation effect was not triggered by the verb. This is confirmed by the results from Experiment III, which show that contrary to scrambled objects, a distant but non-scrambled subject NP is not reactivated at the preverbal test position. Another possible explanation for the antecedent reactivation effect observed in Experiments I & II merits consideration though. Given that our target words were identical to the antecedents, it is conceivable that the observed priming effect was not due to antecedent reactivation, but to some residual level of activation of the displaced element that was maintained throughout the sentence (compare Nicol 1993). Although reaction times to identical targets were faster in all conditions, possibly reflecting a certain amount of maintained activation, this does not explain the results on object traces from the High Span Group. Contrary to what we would expect, High Span participants' reaction times were actually shorter at the Late Position than at the Early Position. Hence the view that antecedent reactivation effects are due to maintained activation only cannot explain our findings, either. The results for the Low Span Group, on the other hand, are consistent with the idea of there being a certain (but gradually decreasing) level of maintained activation: Whereas the Target Type effect found at the earlier control position could well be due to residual activation of the scrambled object, by the time the gap position was reached, the antecedent had faded from these participants' working memory.
The results from our experiments also indicate an interaction between individual working memory capacity and antedecent reactivation effects. We argue that whereas High Span participants successfully established filler-gap dependencies during the online processing of long-distance scrambled sentences, normal processing broke down at some point in the Low Span Group, most likely due to processing overload. Processing overload effects occur if the structure to be processed exceeds a certain complexity limit, that is, if the processing effort required to parse a sentence successfully exceeds the available amount of computational or memory resources (King & Just 1991, Just & Carpenter 1992, Gibson 1998). A number of scholars have tried to define a psycholinguistically relevant syntactic complexity metric (Miller & Chomsky 1963, Frazier 1985, Hawkins 1994, Gibson 1991, 1998). According to Gibson's (1991) case-based account of processing overload, for example, more than four local theta-violations are likely to cause the processor to break down (a local thetaviolation occurs if a case-marked noun phrase is processed before its theta-assigner). Note that in processing the critical sentences used in Experiments I & II, the parser encounters five case-marked argument NPs before receiving the first theta-assigning verb. 13 Gibson's (1998) Syntactic Prediction Locality Theory (SPLT) allows for the relative processing cost at each point in a sentence to be calculated. Computational cost is made up from two components: (i) integration cost (a function of the number of new discourse referents introduced between a dependent element and its associated head or host), and (ii) memory cost (a function of the number of obligatory syntactic predictions held in working memory). For illustration, consider the memory cost incurred during the processing of longdistance scrambled sentences such as (10) (repeated below for convenience).
[CP/IP futari-me-no hito-ga
And then lemon-ACC
the second person-NOM M.C.-DAT
[CP sono kodomo-ga onna-no hito-ni that child-NOM female person-DAT ^
ti nedatte-iru to ] asking ^^
COMP ] answered]
Lit. "And then, a lemon, the second person answered to the Master of Ceremonies that that child was asking the woman for."
Memory cost is at its maximum at the beginning of the embedded clause. The minimum number of obligatory syntactic predictions that must be kept in working memory during the processing of the second NP-ga - NP-ni sequence is four: (i) the prediction of the matrix verb, (ii) the prediction of the embedded verb, (iii) the prediction of the complementizer, and (iv) the prediction of the object gap. Recall further that to the extent that sentences of this type induce a garden path effect, some amount of reanalysis may also be required as the second nominative NP is encountered, which is likely to use up additional processing resources. It is conceivable that, at parsing stages where computational resources are pushed to their limit, old predictions will be dropped (i.e., eliminated from memory space) first. In our experimental sentences, the oldest prediction (apart from the prediction of the matrix verb, which Gibson 1998 suggests may not in fact 'count') is the prediction of the object gap. The idea that long-distance scrambled sentences of the type we used in our experiments may cause processor overload is consistent with the fact that many speakers of Japanese consider them rather unnatural when presented in isolation, and with the comments made by some of our participants to the effect that they had experienced comprehension difficulties while
listening to the experimental texts. Observe, however, that processing does not appear to have collapsed completely in our Low Span participants - if it had, they would presumably not have been able to understand the test sentences, as witnessed by their ability to produce accurate summaries of the critical texts. Hence it is possible that all that went wrong in the Low Span participants was that they 'forgot' the antecedent halfway through parsing the sentence, while nevertheless being able to construct an (otherwise grammatical) mental representation of the sentence. They may, for instance, have postulated a genuine null object (i.e., a null pronominal) in place of an object trace, which was then not associated with any particular discourse referent until the sentence-final wrap-up stage. 14 Our results are thus compatible with studies showing that associating a dislocated constituent with the corresponding gap becomes more difficult if the distance between the two is increased (Gibson & Warren 1998, Kaan et al. 2000, Miyamoto & Takahashi 2002, among others). Other studies lend support to our hypothesis that the distance between the scrambled object and its trace might have exceeded the memory span of some of our participants, effectively rendering the antecedent inaccessible to automatic processing. Waters & Caplan (1996), for instance, examined the relationship between Daneman-Carpenter reading span scores and the efficiency of processing different types of sentence. They argue that although the reading span test was designed to measure efficiency of processing, it actually measures only storage. This suggests that although reading span scores may predict for how long participants are likely to keep antecedents in their short-term memory, they do not necessarily serve as predictors for the efficiency of processing as such. 15 A recent study by McElree (2000) using the response-signal speed-accuracy tradeoff procedure, which measures accuracy as a function of processing speed, provides further evidence that under certain conditions, antecedents held in working memory may be lost or difficult to retrieve during the processing of filler-gap dependencies.
Implications for linguistic theory
The results from Experiments I-III have some interesting implications for linguistic theory in general, and for the analysis of Japanese in particular. First of all, our findings suggest that the human parser does indeed make use of empty categories during online sentence processing, and are thus consistent with grammatical theories that include movement traces in their theoretical inventory. Secondly, our results support the view that Japanese has a configurational rather than a 'flat' structure (compare [13a] vs. [13b]).
S/VP 2 NP-ga VP 2 NP-ni V 2 NP-o V
S wp NP-ga NP-ni NP-o V
According to the non-configurational hypothesis proposed (among others) by Hale (1980) and Farmer (1984), Japanese lacks a VP node, and scrambled word orders are base-generated. Since all arguments stand in exactly the same hierarchical relation to the verb, and the linearization of arguments is essentially arbitrary, the non-configurational hypothesis predicts no difference with respect to antecedent reactivation between distant subject and object NPs neither should be reactivated at the preverbal test position. The trace reactivation effects we found, however, support a configurational structure of Japanese along the lines of (15a), with the direct object being the most deeply embedded argument. Our results are furthermore difficult to reconcile with Kayne's (1994) Universal Base Hypothesis, according to which the underlying word order is universally SVO. If direct objects originate after rather than before the verb in Japanese, as is the case in head-initial languages such as English, we would not expect to find any reactivation effects for scrambled 37
objects at the preverbal test position. The fact that we did find evidence for trace reactivation at this position, however, supports the assumption that the base order of Japanese VP is headfinal.
We have argued that the observed priming effect in the High Span Group supports a tracebased account of Japanese long-distance scrambling, and are inconsistent with lexicalist hypotheses such as the DAH. Given the head-final nature of Japanese, our results suggest that argument traces access their antecedents irrespective of the position of their subcategorizers. Our results also support Love & Swinney's (1996) and Nicol et al.'s (1997) claim that the CMLP method is sensitive to structural processing. To the extent that results from psycholinguistics can inform linguistic theory, our findings are consistent with theoretical frameworks which assume that displaced constituents are related to their subcategorizers via empty categories rather than through direct lexical association. They also support the view that the structure of Japanese is configurational, and argue against Kayne's (1994) claim that direct objects are universally generated postverbally. Regarding the absence of antecedent reactivation effects in the Low Span participants, we suggested that our experimental sentences exceeded the memory span of this group.
Acknowledgements We are grateful to Martin Atkinson, Bob Borsley, Sam Featherston, Roger Hawkins, Ed Miyamoto, Martin Pickering and Andrew Radford for helpful comments and discussion, to Axel Huth for providing technical support, and to Leah Roberts for helping with the English translations. Further thanks are due to Taeko Wydell for permission to use her frequency corpus, to the Institute of International Education in London for their cooperation in connection with the data collection, and to Yoshiro Nakano for his support.
Appendix 1: An Example Text The following example text is taken from Exerperiments I and II. The experimental sentence is shown in italics, and visual target words in boxes, at the positions at which they appeared on the screen. Ichiroo-wa kafunshoo-de, kafun-no tobikau haru-niwa, iroirona kusuri ya iryooyoogu-ga nai to, ポンネル (nonce word)
netsu-o dashiteshimau. Kotoshi-mo haru-ni natta node, kusuriya-ni iku to, kafunshoo-no 和紙 (washi “Japanese paper”)
yobooyoogu-ga miataranai. Sorede, kusuriya-no shujin-ni dooshitanoka kiite-mita. Suruto, 粟矛 ( nonce word) shujin-ga Ichiro-ni, kyuunyuuki-wa zenbu ureteshimatta to kotaeta. Sorekara, subete-no masuku-o, shujin-ga Ichiro-ni baito-no tenin-ga kireina joseikyaku-ni ageteshimatta to itta. Shujin-wa sono マスク?ドラム “mask/drum”) baito-no tenin-o sokkoku kubi-ni shita-sooda. ランプ (ranpu “lamp”) (masuku/doramu
(Ichiro suffers from hay fever, and during the spring when pollen is in the air, he gets a fever if he cannot take various medications. When spring came that year, he went to a pharmacy but couldn’t find any preventive medicine. He asked what had happened. The shopkeeper told Ichiro that all the inhalers had sold out. And, the shopkeeper told Ichiro, that a part-time shop assistant had given all the masks to a pretty woman. He said that he had immediately fired that part-time assistant.) Appendix 2: Experimental Sentences for Experiments I and II Visual target words are shown in a box (Identical / Unrelated) for each experimental clause. There were two test points, at which one of the target words was shown to participants: (i) the gap position (Late) shown by #, and (ii) a control position (Early) 500 ms prior to (i). Target types and test points were counterbalanced across presentation lists. Each participant heard the same sentences and saw the same target words only once. 1. Kyodaina tento-o otetsudaisan-ga keiji-ni Izumi-yoogisha-ga korosareta-yuujin-ni # kashiteita to kateita. (A gigantic tent, the maid answered to the detective that the suspect Izumi had lent to his murdered friend.) tento/jamu "tent/jam" 2. Puuru-o Hashimoto-san-ga choochoo-ni machi-de ichiban-no okanemochi-ga yoogo-gakkoo-no kodomotachi-ni # kashita to hanashita. (A (swimming) pool, Mr. Hashimoto told the mayor that the richest man in the town had lent to the children of the special school.) puuru/pedaru "(swimming) pool/pedal"
3. Hebi-no ekisu-o tsuuyaku-ga Buraun-san-ni kitooshi-ga kazehiki-no kodomo-ni # ataeta to oshieta. ((In an unknown jungle) the extract from a snake, the interpreter told Mr. Brown that the witch doctor had given to the child (who) caught a cold.) ekisu/taoru "extract/towel" 4. Subete-no masuku-o shujin-ga Ichiro-ni baito-no tenin-ga kireina josei-ni # ageteshimatta to itta. (All the masks, the owner told Ichiro that a part-time shop assistant had given to a pretty woman.) masuku/doramu "mask/drum" 5. Gojuu-man-en no doresu-o kooan-iinkai-ga seken-ni roodoo-daijin-ga Ginza-no hosutesu-ni # okutta to happyooshita. (A five-hundred-thousand-yen dress, the public safety commission revealed to the public that the Labor Minister had given to a hostess in Ginza.) doresu/karee "dress/curry" 6. Nekutai-o futari-me-no tenin-ga tantoosha-ni sanjuu-dai-no josei-ga booifurendo-ni # ageta to hanashita. (A necktie, the second shop assistant told the person in charge that a woman in her 30s had given to her boyfriend.) nekutai/robotto "necktie/robot" 7. Omocha-no oobun-o futari-me-no kaitoosha-ga shikaisha-ni Santakuroosu-ga neteiru kodomo-ni # todoketeiru to kotaita. (A toy oven, the second participant answered to the game show host that Father Christmas had delivered to a sleeping child.) oobun/suraido "oven/slide" 8. Posuto-o betsuno kisha-ga dokusha-ni daijin-ga kootaishi-fusai-ni # omiseshita to hookokushita. ((In the exhibition of new products of the Ministry of Posts and Telecommunication), a mailbox, another journalist reported to the readers that the Minister had shown to the Crown Prince and Princess.) posuto/booru "mailbox/ball" 9. Ebi-no pirafu-o otoo-san-ga okaasan-ni jijo-no Hiroko-ga otooto-no Hiroshi-ni # tsukutta to itta. (Shrimp pilaf, Father told Mother that their second daughter Hiroko had made for Hiroshi.) pirafu/baketsu "pilaf/bucket" 10. Kabocha-no suupu-o okaasan-ga Kenichi-kun-ni obaachan-ga inu-no Arekusandaa-ni # yatteshimatta to itteiru. (Pumpkin soup, Mother told Kenichi that Grandmother had given to their dog Alexander.) suupu/gitaa "soup/guitar" 11. Remon-o futarime-no hito-ga shikai-sha-ni sono-kodomo-ga onnanohito-ni # nedatte-iru to kotaeta. (A lemon, the second person answered to the Master of Ceremonies that the child was asking the woman for.) remon/sonata "lemon/sonata" 12. Rimokon-no robotto-o otoosan-ga okaasan-ni ani-no John-ga otooto-no Bobu-ni # yuzutta to itta. (The remote-controlled robot, Father told Mother that the elder brother John had given to his younger brother Bob.) robotto/nekutai "robot/necktie" 13. Keeki-o sannin-me-no hito-ga shikaisha-ni yasashisoona okaasan-ga kyuu-sai-no kodomo-ni # yaiteiru to kotaeta (A cake, the third person answered to the Master of Ceremonies that a kind-looking mother baked for a nine-year-old child.) keeki/suriru "cake/thrill" 14. Shatsu-o Suuzan-ga ojiisan-ni Tomu-ga Santakuroosu-ni # tanonda to itta. (A shirt, Susan told her grandfather that Tom had asked Father Christmas for.) shatsu/naifu "shirt/knife" 15. Pedaru-o Kitaro-san-ga Haruzoo-san-ni shachoo-ga shokudoo-no kokku-ni # ageta to hanashita.
((In a bicycle factory,) a pedal, Kitaro told Haruzo that the company director had given to a cook in the canteen.) pedaru/puuru "pedal/(swimming) pool" 16. Karee-o shoogakusei-no onnanoko-ga kurasumeeto-ni otoosan-ga tomobataraki-no okaasan-ni # tsukutta to itta. (Curry, an elementary school girl told her classmate that her father made for her mother (who) also had a job.) karee/doresu "curry/dress" 17. Medaru-o anaunsaa-ga shichoo-sha-ni tennooheika-ga rokujussai-no josei-ni # sazuketa to tsutaeta. (A medal, the announcer told the audience that the Emperor had handed to the sixty-year-old woman.) medaru/koppu "medal/cup" 18. Suteeki-o hisho-ga Takayama-sensei-ni Kageyama-san-ga futari-no joshigakusei-ni # ogotta to hanashita (Beefsteak, the secretary told Professor Takayama that Dr. Kageyama had treated the two female students to.) suteeki/purinto "beefsteak/handout" 19. Burezaa-o Kinoshita-san-ga Tanaka-san-ni ojiisan-ga jinan-no Kenji-ni # nokoshita to itta. (A jacket, Ms Kinoshita was told Mr. Tanaka that her grandfather had left to his second son Kenji.) burezaa/kaaten "jacket/curtain" 20. Kookoku-no purinto-o Takako-san-ga okaasan-ni ginkooin-ga san-nin tsuukonin-ni # kubatteita to hanashita. ((On a street) handouts, Takako told her mother that three bank clerks were giving out to passersby.) purinto/suteeki "handout/beefsteak" 21. Suriru-o sono komedian-ga Yatabe-san-ni jibun-no tomodachi-ga kodomotachi-ni # teikyooshiteiru to itta. ((In the conversation about a friend who controls a roller coaster in an amusement park,…) excitement, the comedian told Mr. Yatabe that his own friend was giving to children.) suriru/keeki "thrill/cake" 22. Shinpin-no naifu-o okaasan-ga obaasan-ni obasan-ga mei-no Makiko-san-ni # katta to itta. (A brand new knife, the mother told the grandmother that the aunt had bought for her niece Makiko.) naifu/shatsu "knife/shirt" 23. Yakyuu-no booru-o choojo-ga Joonzu-san-ni Betii-ga booifurendo-ni # nedatta to hanashita. (A baseball, the eldest daughter told Mr. Jones that Betty had asked her boyfriend for.) booru/posuto "ball/mailbox" 24. Nooto-o Satoru-kun-ga okaasan-ni kurasu-no zenin-ga tannin-no sensei-ni # dashita to kotaeta. (A notebook, Satoru answered to his mother that everybody in the class had submitted to their class teacher.) nooto/biiru "notebook/beer" 25. Tomato-o Bobu-ga Takashi-ni chuugakusei-ga tsuukoo-nin-ni # nageta to itteita. (A tomato, Bob was telling Takashi that a junior-high school kid had thrown at a passer-by.) tomato/taitsu "tomato/tights" 26. Akai koppu-o Yuki-san-ga Kaori-san-ni okaasan-ga oneesan-ni # ageta to itta. (A red cup, Yuki told Kaori that her mother had given to her elder sister.) koppu/medaru "cup/medal" 27. Jamu-o gesuto-ga anaunsaa-ni hoomuherupaa-no hito-ga hitorigurashi-no roojin-ni # okutta to itta. (Jam, the guest speaker told the announcer that a home help had presented to an elderly person living alone.) jamu/tento "jam/tent" 28. Gitaa-o Takai-san-ga Rikarudo-ni Karurosu-ga daigaku-no sensei-ni # azuketeiru to kotaeta.
(A guitar, Mr. Takai answered to Ricardo that Carlos had his university teacher keep for him.) gitaa/suupu "guitar/soup" 29. Tsumetai biiru-o futarime-no hito-ga shikai-sha-ni sono onnano-hito-ga danseikyaku-ni # dashiteiru to kotaeta (Cold beer, the second person answered to the Master of Ceremonies that the woman was serving to the male customer.) biiru/nooto "beer/notebook" 30. Himitsu-no suraido-o sono mokugekisha-ga Hoomuzu-ni san-nin-me-no supai-ga kurofuku-no josei-ni # watashiteita to itta. (A confidential slide, the witness told Holmes that the third spy had been handing to the woman in black.) suraido/oobun "slide/oven" 31. Afurika-no doramu-o Matsuda-sensei-ga sono kyookan-ni san-nin-me-no jisshusei-ga sannensei-ni # oshieteiru to itta. (The African drum, Mr. Matsuda told the teacher that the third practice teacher was teaching thirdyear students (how to play).) doramu/masuku "drum/mask" 32. Masshirona taoru-o anaunsaa-ga shichoosha-ni dansei-no fan-ga yuushooshita rikishi-ni # nageta to tsutaeta. (A towel in shining white, the announcer told the (TV) viewers that a male fan had thrown to the sumo champion.) taoru/ekisu "towel/extract" 33. Renzu-o sono otokono hito-ga Saito-san-ni dairide kita musuko-ga mise-no shujin-ni # tanondeoita to itta. (A lens, the man told Miss Saito that his son, whom he had sent, had ordered from the shop.) renzu/soosu "lens/sauce" 34. Ookina baketsu-o Maikeru-san-ga dooryoo-ni okaasan-ga nenchoo-no kodomo-ni # tewatashiteita to hanashita. (A big bucket, Michel told his colleague that the mother had been handing to the older kid.) baketsu/pirafu "bucket/pilaf" 35. Mootsuaruto-no sonata-o sono okaasan-ga Akiko-san-ni onaji sennsei-ga sono onnanoko-ni # oshieta to kotaeta. (Mozart’s sonatas, the mother answered to Akiko that the female teacher had taught to the girl.) sonata/remon "sonata/lemon" 36. Murasaki to kiiro-no taitsu-o minna-ga bucho-ni Minoru-ga Piitaa Pan-ni # eranda to ittekita. (In the setting that in a drama club the members had some complaints against the bad taste of their dress coordinator, (a pair of) purple and yellow tights, everybody came to tell to the head of the club that Minoru had chosen for Peter Pan.) taitsu/tomato "tights/tomato" 37. Hanagara-no kaaten-o Takahashi-san-ga reji-no hito-ni imooto-ga wakai tenin-ni # tanonda to hanashita. (A floral curtain, Takashi was telling a person at a cash desk his younger sister had ordered from a young shop assistant.) kaaten/ burezaa "curtain/jacket" 38. Sono kanja-no karute-o Kimura-san-ga keiji-ni jimuin-ga kazoku-no hito-ni # miseta to hanashita. (The patient’s medical report, Mr. Kimura told the detective that an administrator had shown to the (patient’s) family.) karute/jazu "medical report/jazz"
39. Wasabi-no soosu-o sono resutoran-no maneejaa-ga Kimura-san-ni onaji hito-ga wakai joseikyaku-ni # susumeta to kotaeta. (Horseradish sauce, the restaurant manager answered to Mr. Kimura that the same person had recommended to the young woman.)
40. Jazu-o gaido-san-ga Taniguchi-san-tachi-ni aru mumei-no wakai kokujin-ga utsukushii koibito-ni # sasageta to hanashita. ((A piece of) jazz, a guide told Ms. Taniguchi and the others that an unknown young black man had dedicated to his beautiful sweetheart.) jazu/karute "jazz/medical record"
Appendix 3: Experimental Sentences for Experiment III Visual target words are shown in a box (Identical / Unrelated) for each experimental clause. There were two test points, at which one of the target words was shown to participants: (i) the gap position (Late) shown by #, and (ii) a control position (Early) 500 ms prior to (i). Target types and test points were counterbalanced across presentation lists. 1. 2. 3. 4. 5. 6. 7. 8. 9. Ano kami-no nagai moderu-ga boku-ni nage kisu-o # okuttanda. (That model with long hair threw me a kiss.)
Kono aida-no joyuu-ga mata otokonoko-ni purezento-o # agete-iru. (The actress that I met the other day is giving a boy a present again.)
Ano ishiku-no Deibiddo-ga Suuzan-ni bara-no hanataba-o # ageta-n-desu-tte. (I heard (that) that engraver gave Susan a rose bouquet.)
ishiku/taiko "engraver/ancient times"
Ano kaeru-ga uchi-no niwa-ni ookina mesu-o # tsuretekita-wayo. (That frog brought a female (frog) to our garden.) kaeru/rifuto "frog/lift" Kesa, ano tsuuyaku-ga nyuusu-no anaunsaa-ni genkoo-o # watashiteita. (This morning, the interpreter was handing the newsreader a manuscript.)
Ano ryoshi-ga gomibako-ni biiru-no akikan-o # nageireteita. (That fisherman was throwing an empty beer bottle into the rubbish bin.)
Ano kitsune-ga hoka-no nakama-ni esa-o # agenai. (That fox won’t give others any food.) kitsune/sotee "fox/sauté" Junsa-ga shoonen-tachi-ni juusho to shozoku-o # shitsumonshita. (The policeman asked the boys for their addresses and affiliation.) junsa/kootaku "policeman/gloss" Kono aida-no fujin-ga uchi-no mise-ni takkyuu-no booru-o # chuumonshita-zo. (The lady (who visited us) the other day ordered from our shop table-tennis balls.)
fujin/kooshuu "lady/the public"
10. Nezumi-ga tenjoo-ura-ni yasai-no kuzu-o # mottekiteita. (The mouse brought the remains of vegetables into the attic.) nezumi/haabu "mouse/herb"
11. Ano niwashi-ga wakai hito-ni kaabingu-no shikata-o # oshieterundawayo. (That gardener is teaching young people the method of curving.) niwashi/monogatari "gardener/story"
12. Ano haiyuyu-ga charitii konsaato-ni shoogaiji-o # shootaishita-zo. (That actor invited (some) disabled children to the charity concert.) haiyuu/kioku "actor/memory" 13. Suzume-ga sono nikai-no tobukuro-ni sanbiki aomushi-o # hakondekita. (The sparrow brought three caterpillars to the window container.) suzume/penki "sparrow/paint 14. Nohara-de, hitsuji-ga kodomotachi-ni michi-o # oshieru. ((In children’s literature…) In the field, a sheep is telling children one way.) hitsuji/nozuru "sheep/nozzle" 15. Gorira-ga okyaku-ni banana-no kawa-o # nagetsuketa. (The gorilla threw a banana skin at a visitor.) ゴリラ gorira/korona "gorilla/corona" 16. Ano naasu-ga onnanoko-ni oiwai-o # ageta-zo. (That nurse gave the girl a present.) naasu/terasu "nurse/balcony" 17. Hyooshooshiki-de, yakuin-ga nii-no hito-ni medaru-o # watashita. (At the awards ceremony, a committee member gave the second winner a medal.) yakuin/zuga "committee member/drawing" 18. Kaizoku-ga shoonin-ni nusunda koshoo-o # uritsukeyootoshita. (The pirates tried to sell the merchants some stolen pepper.) kaizoku/gakusei "pirate/maestro" 19. Gakusha-ga Itaria-seifu-ni choosa-no hookokusho-o # teishutsushita. (The scholar submitted the report of (his) investigation to the Italian government.) gakusha/kuuki "scholar/air" 20. Dokusha-ga sono komatte-iru hitotachi-ni enjokin-o # okuttekita. (The readers sent the people in trouble some money as aids.) dokusha/riyuu "reader/reason" 21. Ooji-ga kodomotachi-ni omocha-o # ageta. (The prince gave the children (some) toys.) ooji/kaseki "prince/fossil" 22. Hitori-no deshi-ga sono okyakusama-ni ocha-o # dashimashita. (One apprentice served the visitor with (some) tea.) deshi/zukoo "apprentice/Art(class)((at elementary school))" 23. Noomin-ga sakumotsu-ni mizu-o # yatteitarisuru. (Farmers are watering the crops.) 農民 noomin/suuji "farmer/number" 24. Kakkoku-no shisetsu-ga kootei-no shitsuji-ni oosama-kara-no tegami-o # watashimasu. (The ambassadors of (various) countries hand the emperor’s butler letters from their kings.) shisetsu/teiri "ambassador/theorem" 25. Hitori-no shinshi-ga madoguchi-no oneesan-ni nanika-o # watashimashita. (A gentleman handed the lady at the reception something.) shinshi/shushi "gentleman/seed" 26. Hitori-no kooin-ga Katoo-san-ni mikan-o # nagemashita. (A factory worker threw Mr. Kato an orange.) 工員 kooin/hitogaki "factory worker/crowd" 27. Shufu-ga Burajiru-kara-no ijuusha-ni nihongo-o # oshieteimasu. (A house wife is teaching Brazilian immigrants Japanese.) 主婦 shufu/iryoo "house wife/clothing" 28. Jooou-ga endoo-no hitobito-ni egao-o # miseteita. (The Queen was smiling at the people by the roadside.) joo-oo/akubi "queen/yawn"
References Babyonyshev, M., & E. Gibson (1995). Processing overload in Japanese. In C. Schuetze, J. Ganger & K. Broihier (Eds.), Papers on language processing and acquisition. MIT Working Papers in Linguistics 21, Cambridge, MA: MITWPL. Baumann, H., Nagengast, J., and G. Klaas (1993). New Experimental Setup (NESU). Ms., Max-Planck Institute for Psycholinguistics, Nijmegen. Berwick, R. C., & Weinberg, A. S. (1984). The Grammatical Basis of Linguistic Performance. Cambridge: MA: MIT Press. Bever, T. G., and B. McElree (1988). Empty categories access their antecedents during comprehension. Linguistic Inquiry 19, 35-43 Butler, H. Y. (1994). Processing of Japanese and Korean. Unpublished doctoral dissertation, Ohio State University, Columbus, Ohio. Caplan, D., and Waters, G. S. (1999). Verbal working memory and sentence comprehension. Behavioral and Brain Sciences 22, 77-126. Chomsky, N. (1981). Lectures on Government and Binding. Dordrecht: Foris. Chomsky, N. (1995). The Minimalist Program. Cambridge, MA: MIT Press. Chomsky, N. (1998). Minimalist Inquiries: The Framework. MIT Occasional Papers 15, Cambridge, MA: MITWPL. Chomsky, N., and H. Lasnik (1993). Principles and parameters theory. In J. Jacobs, A. von Stechow, W. Sternefeld and T. Vennemann (Eds.) Syntax: An International Handbook of Contemporary Research, 505-569. Berlin: de Gruyter. Clahsen, H., and S. Featherston (1999). Antecedent priming at trace positions: evidence from German scrambling. Journal of Psycholinguistic Research, 28, 415-437. Daneman, M., and P. Carpenter (1980). Individual differences in working memory and reading. Journal of Verbal Learning and Verbal Behavior 19, 450-466. De Vincenzi, M. (1991). Filler-gap dependencies in a null-subject language: referential and non-referential WHs. Journal of Psycholinguistic Research 20, 197-213. Farmer, A. K. (1984). Modularity in Syntax: A Study of Japanese and Syntax. Cambridge, MA: MIT Press. Frazier, L. (1985). Syntactic complexity. In D. Dowty, L. Karttunen and A. Zwicky (Eds.) Natural Language Parsing. Cambridge: Cambridge University Press. Frazier, L. (1987a). Sentence processing: A tutorial overview. In M. Coltheart (ed.) Attention and Performance XII. Hillsdale, NJ: Lawrence Erlbaum Associates. Frazier, L. (1987b). Syntactic processing: Evidence from Dutch. Natural Language & Linguistic Theory 5, 519-559. Gibson, E. (1991) A Computational Theory of Human Linguistic Processing: Memory Limitations and Breakdown. PhD Dissertation, Carnegie Mellon University. Gibson, E. (1998). Linguistic complexity: Locality and syntactic dependencies. Cognition 68, 1-76. Gibson, E., and G. Hickok (1993). Sentence processing with empty categories. Language and Cognitive Processes 8, 147-161. 45
Gibson, E., and Warren, T. (1998). The psychological reality of intermediate linguistic structure in long-distance extractions. Paper presented at the 11th Annual CUNY Conference on Human Sentence Processing, New Brunswick, NJ. Hale, K. (1980). Remarks on Japanese phrase structure: Comments on the papers on Japanese syntax. MIT Working Papers in Linguistics 2, 185-203. Harada, S.-I. (1977). Nihongo ni henkei wa hituyoo da. Gengo 6(10), 88-95; 6(11), 96-103. Hawkins, J. A. (1994). A Performance Theory of Order and Constituency. Cambridge: Cambridge University Press. Hickok, G., E. Canseco-Gonzales, E. Zurif, and J. Grimshaw (1992) Modularity in locating wh-gaps. Journal of Psycholinguistic Research 21, 545-561. Just, M. A., and P. A. Carpenter (1992). A capacity theory of comprehension: individual differences in working memory. Psychological Review, 99, 122-149. Kaan, E., Harris, A., Gibson, E., and Holcomb, P. (2000). The P600 as an index of syntactic integration difficulty. Language and Cognitive Processes 15, 159-201. Kayne, R. (1994). The Antisymmetry of Syntax. Cambridge, MA: MIT Press. King, J., and M. A. Just (1991). Individual differences in syntactic processing: The role of working memory. Journal of Memory and Language, 30, 580-602. Koopman, H., and D. Sportiche (1991). The position of subjects. Lingua 85, 211-258. Love, T., and D. Swinney (1996). Co-reference processing and levels of analysis on objectrelative constructions: Demonstration of antecedent reactivation with the cross modal priming paradigm. Journal of Psycholinguistic Research 25, 5-24. Mahajan, A. (1990). The A/A' Distinction and Movement Theory. MIT Dissertation, Cambridge, MA. Mahajan, A. (1994). ‘Toward a unified theory of scrambling’, in N. Corver and H. van Riemsdijk (Eds.) Studies on Scrambling. Berlin: Mouton de Gruyter. Mazuka, R., and N. Nagai (Eds.) (1995). Japanese Sentence Processing. Hillsdale, NJ: Lawrence Erlbaum Associates. Mazuka, R., Itoh, K. and T. Kondo (2001). Costs of scrambling in Japanese sentence processing. In M. Nakayama (Ed.), Sentence processing in East Asian languages. Stanford: CSLI. McElree, B. (2000). Sentence comprehension is mediated by content-addressable memory structures. Journal of Psycholinguistic Research, 29, 111-123. McKoon, G., and R. Ratcliff (1994). Sentential context and on-line lexical decision. Journal of Experimental Psychology: Learning, Memory and Cognition, 20, 1239-1243. McKoon, G., Ratcliff, R., and D. Allbritton (1996). Sentential context effects on lexical decisions with a cross-modal instead of an all-visual procedure. Journal of Experimental Psychology: Leaning, Memory, and Cognition 22, 1494-1497. McKoon, G., Ratcliff, R., and G. Ward (1994). Testing theories of language processing: An empirical investigation of the on-line lexical decision task. Journal of Experimental Psychology: Learning, Memory, and Cognition 20, 1219-1228.
Miller, G. A., and N. Chomsky (1963). Finitary models of language users. In R. D. Luce, R. R. Bush, and E. Galanter (Eds.) Handbook of Mathematical Psychology, Vol. 2. New York: Wiley. Miyamato, E. T. and S. Takahashi (2002). Sources of difficulty in the processing of scrambling in Japanese. In M. Nakayama (Ed.), Sentence processing in East Asian languages. Stanford: CSLI. Muckel, S. and T. Pechmann (2001). Trace identification in German: Prosodic hints help before the verb is being processed. Poster presented at the 14th Annual CUNY Conference on Human Sentence Processing, Philadelphia, PA.. Nakayama, M. (1990). Accessibility to the antecedent in Japanese sentence processing. Unpublished ms., Ohio State University, Columbus, Ohio. Nakayama, M. (1995). Scrambling and probe recognition. In R. Mazuka & N. Nagai (Eds.), Japanese sentence processing. Hillsdale, NJ: Lawrence Erlbaum Associates. Nemoto, N. (1995). Scrambling in Japanese, AGRoP, and economy of derivation. Lingua 97, 257-273. Nemoto, N. (1999). Scrambling. In N. Tsujimura (ed). The Handbook of Japanese Linguistics. Oxford: Blackwell, 121-153. Nicol, J. (1993). Reconsidering reactivation. In G. A. R. Shillcock (Ed.), Cognitive Models of Speech Processing: The Second Sperlonga Meeting (pp. 321-347). Hove: Erlbaum. Nicol, J., Fodor, J. D., and D. Swinney (1994). Using cross-modal lexical decision tasks to investigate sentence processing. Journal of Experimental Psychology: Learning, Memory, and Cognition, 20, 1229-1238. Nicol, J., and Swinney, D. (1989). The role of structure in coreference assignment during sentence comprehension. Journal of Psycholinguistic Research, 18, 5-19. Nicol, J., D. Swinney, T. Love, and L. Hald (1997). Examination of sentence processing with continuous vs. interrupted presentation paradigms. Center for Human Information Processing, Technical Report #97-3, La Jolla, U.C.S.D. Ninose, Y., Oda, J., Sakaki, Y., Sakamoto, T., and Gyoba, J. (1998). On the real-time processing of empty subjects in Japanese using a dichotic-listening method (2). Cognitive Studies 5, 82-88. Oda, J., Ninose, Y., Sakaki, Y., Gyoba, J. and Sakamoto, T. (1997). On the real-time processing of empty subjects in Japanese using a dichotic listening method. Cognitive Studies 4, 58-63. Osaka, M. (1998). Yomi to wakingu memori [Reading and Working Memory]. In N. Osaka (Ed.), Yomi - No to kokoro-no johoshori [Reading - information processing of brain and mind] (pp. 239-262). Tokyo: Asakura. Osaka, M., and Osaka, N. (1992). Language-independent working memory as measured by Japanese and English reading span tests. Bulletin of the Psychonomic Society 30, 287289. Pickering, M. (1993). Direct association and sentence processing: A reply to Gorrell and to Gibson and Hickok. Language and cognitive Processes, 8, 163-196. Pickering, M., and G. Barry (1991). Sentence processing without empty categories. Language and Cognitive Processes, 6, 229-259.
Pollard, C., and I. A. Sag (1994). Head-Driven Phrase Structure Grammar. Chicago: University of Chicago Press. Ross, J. (1967). Constraints on Variables in Syntax. MIT Dissertation, Cambridge, MA. Sag, I., and J. D. Fodor (1994). Extraction without traces. Proceedings of the 13th West Coast Conference on Formal Linguistics, 365-384. Saito, M. (1985). Some Asymmetries in Japanese and their Theoretical Implications. Unpublished Ph.D. Dissertation, MIT. Saito, M. (1992). Long Distance Scrambling in Japanese. Journal of East Asian Linguistics, 1, 69-118. Sakamoto, T. (1995). Transparency between parser and grammar: On the processing of empty subjects in Japanese. In R. Mazuka and N. Nagai (Eds.), Japanese sentence processing. Hillsdale, NJ: Lawrence Erlbaum Associates. Swinney, D. (1979). Lexical access during sentence comprehension: (Re)consideration of context effects. Journal of Verbal Learning and Verbal Behavior 18, 645-659. Swinney, D., J. Nicol, M. Ford, U. Frauenfelder, and J. Bresnan (1987). The time-course of co-indexation during sentence comprehension. Paper presented at the Psychonomic Society Meeting, Seattle, WA. Tada, H. (1993). A/A-bar partition in derivation. MIT Dissertation, Cambridge, MA. Tonoike, S. (1997). On scrambling - Scrambling as a base-generated scopal construction. In S. Tonoike (Ed.), Scrambling (pp. 125-159). Tokyo: Kurosio. Traxler, M. J., and M. Pickering (1996). Plausibility and the processing of unbounded dependencies: An eye-tracking study. Journal of Memory and Language 35, 454-475. Ueyama, A. (1994). Against the A/A'-movement dichotomy. In N. Corver and H. van Riemsdijk (Eds.), Studies on Scrambling: Movement and non-movement approaches to free word-order phenomena . Berlin: Mouton de Gruyter. Waters, G. S., and D. Caplan (1996). The capacity theory of sentence comprehension: a critique of Just and Carpenter . Psychological Review 103, 761-772. Wydell, T. N. (1991). Processes in the reading of Japanese: comparative studies between English and Japanese orthographies. Unpublished doctoral thesis, University of London. Yamashita, H. (1997). The effects of word order and case marking information on the processing of Japanese. Journal of Psycholinguistic Research 26, 163-188. Yamashita, H. (2000). Scrambled Sentences in Japanese: Linguistic properties and motivations for production. Unpublished ms., University of Illinois at UrbanaChampaign, Illinois. Yatabe, S. (1993). The boundedness of scrambling. In S. Choi (Ed.), Japanese/Korean Linguistics (Vol. 3, pp. 204-220). Stanford, CA: CSLI. Yatabe, S. (1996). Long-distance scrambling via partial compaction. MIT Working Papers in Linguistics, 29 (Formal Approaches to Japanese Linguistics 2), 303-317. Zurif, E. S., D., Prather, P., and T. Love (1994). Functional localization in the brain with respect to syntactic processing. Journal of Psycholinguistic Research, 23, 487-497. 48
Although the notions of 'trace' and 'copy' are not, strictly speaking, synonymous (compare Chomsky 1998: 29), we shall continue to use the term 'trace' for ease of exposition. There is evidence that the parser postulates a gap as soon as grammatically possible (compare e.g. Frazier's [1987b] Active Filler Strategy, or De Vincenzi's  Minimal Chain Principle). The term scrambling dates back to Ross (1967). For a detailed overview of scrambling in Japanese, see Nemoto (1999). Within Chomsky's (1995) minimalist framework, several intermediate traces will be postulated between the head and the foot of the chain. This is because the Minimal Link Condition (Chomsky & Lasnik 1993, reprinted in Chomsky 1995: 90) requires that movement proceeds in a successive-cyclic fashion. Given that intermediate traces are considered to be irrelevant to interpretation and are ultimately deleted from the representation, we will ignore all but the trace at the foot of the chain here. Nemoto (1995), however, argues that long-distance scrambling from a control clause is an instance of A-movement. The traditional assumption that the Japanese VP is head-final has been challenged by Kayne (1994), whose Universal Base Hypothesis states that even for so-called headfinal languages, the underlying word order is SVO. If this hypothesis is correct, then we do not expect to find any reactivation effects for scrambled direct objects at the preverbal test position. We shall return briefly to this issue in section 8 below. Within the framework of HPSG, Yatabe (1993, 1996) has proposed an analysis of Japanese long-distance scrambling that does not postulate any movement traces. Here, the term bunsetsu refers to a unit consisting of a content word or stem plus any associated particles and/or affixes, or of a content word alone. In syntactic theory, this idea is reflected in locality constraints on movement such as the Subjacency Condition (Chomsky 1981), which prohibits movement operations that cross more than a certain number of 'bounding nodes'. On the assumption that reaction times in the cross-modal lexical decision task reflect both structurally-mediated antecedent reactivation and the accessibility (or level of activation) of the filler held in short-term memory, it is conceivable that there was a trace reactivation effect in the Low Span Group as well, but that this was overridden by the relatively steeper negative recency priming curve. In other words, the priming effect in the Low Span Group might have been masked by the participants' difficulty to retrieve the antecedent from working memory. We thank Martin Atkinson for pointing this possibility out to us. According to Chomsky (1995), only the verb's internal arguments (i.e., its objects) originate within the inner or core VP; arguments associated with the Agent or Causer thematic role, on the other hand, originate in the specifier of an outer vP 'shell'. No matter what fine structure is assumed for the Japanese verb phrase though, the subject NP in ditransitive structures is normally taken to be the (most) external argument of
the verb, and, given the head-final nature of Japanese, will not be generated immediately adjacent to the verb.
Nakayama (1990, cited in Butler 1994) reports that the number of kanji characters affects visual lexical decision times whereas the number of morae does not. Hence our kanji target words were matched for number of characters rather than morae. But see Babyonyshev & Gibson (1995) for some arguments that Gibson's original theory might be too strong (at least for Japanese). Assuming that a breakdown of automatic processing does not necessarily prevent the reader or listener from understanding a sentence's contents, we cannot completely exclude the possibility of a processor breakdown in our Low Span participants though. Both conscious reanalysis and the use of contextual clues may have allowed participants to reconstruct the message to be conveyed. In this case, comprehension would merely be delayed - which would explain why the Low Span participants were able to produce accurate summaries of the experimental texts. This is consistent with Caplan & Waters (1999) view that working memory is organized in a modular fashion, with different modules assigned to different types of WM task.
copyright ©right 2010-2020。