ªLong before shortº preference in the production of a head- nal language - PDF

Description
H. Yamashita, F. Chang / Cognition 81 (2001) B45±B55 B45 COGNITION Cognition 81 (2001) B45±B55 Brief article ªLong before shortº preference in the production of a head- nal

Please download to get full document.

View again

of 11
All materials on our website are shared by users. If you have any questions about copyright issues, please report us to resolve them. We are always happy to assist you.
Information
Category:

Gadgets

Publish on:

Views: 17 | Pages: 11

Extension: PDF | Download: 0

Share
Transcript
H. Yamashita, F. Chang / Cognition 81 (2001) B45±B55 B45 COGNITION Cognition 81 (2001) B45±B55 Brief article ªLong before shortº preference in the production of a head- nal language Hiroko Yamashita*, Franklin Chang University of Illinois at Urbana-Champaign, Urbana, IL, USA Received 25 August 2000; accepted 16 February 2001 Abstract A tendency for speakers to produce short phrases before long phrases has been attributed to the accessibility of short phrases, and thought to re ect universal mechanisms of production. However, recent corpus analyses in Japanese suggest that long phrases tend to be shifted ahead of short ones (Hawkins, J. (1994). A performance theory of order and constituency. Cambridge: Cambridge University Press; Yamashita, in press). Two on-line experiments con rmed that speakers shifted long arguments to earlier positions more than short arguments, exhibiting a ªlong before shortº preference. We reconcile these contradictory data from English and Japanese by a competition between different factors in an incremental production system. q 2001 Elsevier Science B.V. All rights reserved. Keywords: Language production; Phrasal ordering; Scrambling; Japanese 1. ªShort before longº preference in English When people speak, a series of words must be produced in rapid succession. The speaker must access and sequence words within a phrase (i.e. put ªtheº before ªdogº), and furthermore sequence those phrases within a sentence (i.e. put ªthe dogº before ªon the streetº). Given the incredible uency with which people produce sentences, it has been hypothesized that the language production system plans a sentence in an incremental manner, and that decisions about the sequencing of words and phrases depend on the accessibility of these units during the incre- * Corresponding author. 2090A FLB, 707 South Mathews, Urbana, IL 61801, USA. Fax: address: (H. Yamashita) /01/$ - see front matter q 2001 Elsevier Science B.V. All rights reserved. PII: S (01) B46 H. Yamashita, F. Chang / Cognition 81 (2001) B45±B55 mental planning of the sentence (Bock, 1982; Ferreira, 1996). Studies demonstrate that a phrase that is more accessible due to conceptual or lexical factors tends to be produced earlier in the sentence (Bock, 1986; McDonald, Bock, & Kelly, 1993). In this article, we present a phenomenon that contradicts this traditional view of accessibility in word/phrasal ordering and reevaluate the universal character of production mechanisms. One phenomenon that is consistent with the accessibility-based account of word/ phrasal ordering is the ªshort before longº preference in English (e.g. Arnold, Wasow, Losongco, & Grinstrom, 2000; Stallings, MacDonald, & O'Seaghdha, 1998). This preference manifests itself in constructions such as the heavy noun phrase (NP) shift shown in (1). The canonical order in (1a) becomes highly awkward when the NP ªa songº is shifted after the prepositional phrase (PP) ªwith friendsº as in (1b). However, if the NP includes more modi cation, increasing its length as in (1c), the shifted order becomes more acceptable or even preferred over its canonical counterpart. Experimental studies on the construction, both with and without listeners (Arnold et al., 2000; Stallings et al., 1998), suggest that the shift is due to the workings of the production system, rather than a stylistic adaptation for listeners. (1a) Bill sang a song with friends. (1b) # Bill sang with friends a song. (1c) Bill sang with friends a song that was written by a famous guitar player from Texas. While the length in number of words has not been directly linked to accessibility, it seems natural to suggest that short phrases are easier to access than long phrases (Arnold et al., 2000). Short phrases have fewer words to access, and fewer alternative orderings that need to be selected among. Because this account is grounded in the nature of the human sentence processor, the short before long preference might be a universal property of production. 2. ªLong before shortº preference in Japanese Recent corpus analyses in Japanese pose a potential challenge to the universality of an accessibility-based account of phrasal ordering. Japanese is a SOV (subject, object, verb), head- nal language that allows the fronting of arguments (commonly called ªscramblingº). Because of case markers, speakers are allowed to change the order among arguments while keeping semantic content and grammatical relations identical to canonical counterparts. Corpus analyses on written texts (Dryer, 1980; Hawkins, 1994; Yamashita, in press) found that in Japanese, long phrases, such as an argument with a long modi er, tend to shift in front of shorter phrases. These ndings are exactly opposite of what would be predicted if short phrases were more accessible. It is possible that these corpus ndings merely re ect a stylistic bias in written language. We therefore experimentally investigated whether phrase length directly H. Yamashita, F. Chang / Cognition 81 (2001) B45±B55 B47 affects phrasal ordering in Japanese, using a variant of the task that Stallings et al. (1998, Experiment 3) used to study the shift in English. The experiment, which was disguised as a sentence recognition task, asked participants to construct a sentence using sentence components presented on a screen, in any order, and then say the sentence aloud after a cue. 3. Experiment Method Participants Fifty-three Nara University undergraduates were paid for their participation. All were native speakers of Japanese Stimuli Twenty-four sets of sentences were constructed with transitive verbs, a subject NP, and a direct object NP. Three versions of each sentence were created by attaching a modi er to the subject, the object, or leaving the modi er out, yielding long-sub, long-do, and all-short conditions. The modi ers consisted of 5±7 words on average and 10±15 morae. The experimental sentences were combined with 58 ller sentences and 30 stimuli from Experiment 2. Table 1 gives examples of experimental sentences. In addition, when the target sentence was presented, the position of the arguments on the screen was manipulated to avoid screen position effects. In one condition the subject was in a box in the upper corner of the screen and the object was in the lower-left corner; in the other condition their positions were reversed. The verb remained in the upper-left corner throughout. Six lists were created to counterbalance the three versions based on phrase length and two levels of screen positioning Procedure Each participant was tested individually. The participants were told that they were in a memory experiment that measured how quickly they could later recog- Table 1 Example set of experimental items in Experiment 1 Phrase length condition Example All-Short Keezi-ga hannin-o oikaketa. detective-nom suspect-acc chased ªThe detective chased the suspect.º Long-SUB Se-ga takakute gassiri sita keezi-ga hannin-o oikaketa. height-nom tall and big-boned detective-nom suspect-acc chased ªThe detective who is tall and big-boned chased the suspect.º Long-DO Keezi-ga se-ga takakute gassiri sita hannin-o oikaketa. detective-nom height-nom tall and big-boned suspect-acc chased ªThe detective chased the suspect who is tall and big-boned.º B48 H. Yamashita, F. Chang / Cognition 81 (2001) B45±B55 nize the sentences that they uttered. Each trial consisted of three kinds of screens as shown in Fig. 1. Participants rst pressed a return key to initiate a trial, which showed the rst screen containing four boxes with sentence components. Participants prepared a sentence using the sentence components and pressed the return key when they were ready. The screen then went blank for 1500 ms, and then a simple math problem followed. The delay and math problem were designed to encourage subjects to produce their sentences from the meaning, rather than by covert rehearsal or reading from their visual buffer. Immediately afterwards, the third screen presented the verb in the lower-right box as a cue. Participants then produced their sentence aloud using the verb. Because only the verb was given as a cue, the participants had the freedom to order arguments either in a canonical or shifted order. Filler sentences were used to disguise the intent of the study. Forty percent of the ller sentences were presented twice for the recognition task. Participants pressed the spacebar whenever they saw a repeated sentence. To give the impression that phrases were randomly placed on the screen, the location of the arguments and verbs in the llers was randomized. To ensure that the targets were not obvious, however, 50% of the llers had the verb in the upper-left position like the targets. Prior to the main experiment, participants practiced with 10 sentences. All of the sentences that participants uttered were recorded with a Sony Walkman tape recorder (Model WM GX322) using the built-in microphone for later transcription Scoring The data were scored using a coding scheme that marked the order of the subject, object, and verb. Changes of subject to topic were ignored and treated as if they were subjects. The coding also denoted the position of the long phrase and several types of deviations from the original stimuli, as well as various types of errors Results The dependent measure was the percentage of sentences whose object shifted to the sentence-initial position out of all the error-free sentences within each cell. An ANOVA was performed crossing phrase length and location on the screen. As shown in Fig. 2, there was a clear effect of phrase length (F1 2; 104 ˆ38:5; F2 2; 46 ˆ25:4). Although the canonical order was always preferred over shifted order, participants were more likely to front objects when they were long than when they were short (post-hoc comparison of long-object and all-short, F1 1; 104 ˆ58:3; F2 1; 46 ˆ37:3). We con rmed that Japanese speakers tend to shift long arguments ahead of short ones in an on-line task. This tendency was observed in an environment without prior discourse context or an explicit listener, in a similar situation to a previous study in English (Stallings et al., 1998). Note that Japanese allows arguments to shift not only to the sentence-initial position but also to the sentence-internal position when the sentence involves H. Yamashita, F. Chang / Cognition 81 (2001) B45±B55 B49 both indirect and direct objects. The canonical order in (2a) may change by shifting either the indirect or direct object to the sentence-initial position, as in (2b) and (2c). Fig. 1. Sequence of events on each trial of Experiments 1 and 2. All four boxes were lled in presenting the stimuli in Experiment 2. B50 H. Yamashita, F. Chang / Cognition 81 (2001) B45±B55 Fig. 2. Shifting in transitive sentences as a function of phrase length. In addition, as in (2d), the direct object may shift VP internally to the position before the indirect object but after the subject/topic phrase. (2a) Taro-ga/wa Miki-ni ringo-o ageta. Taro-nom/top Miki-dat apple-acc gave (2b) Miki-ni Taro-ga/wa ringo-o ageta. (2c) Ringo-o Taro-ga/wa Miki-ni ageta. (2d) Taro-ga/wa ringo-o Miki-ni ageta. ªTaro gave Miki an apple.º In one of the corpus analyses mentioned earlier (Yamashita, in press), 37% of the shifts were of the internal type, showing that this order does occur. We therefore investigated how participants in an on-line task would position long verb please (VP) internal arguments in Experiment 2. H. Yamashita, F. Chang / Cognition 81 (2001) B45±B55 B51 4. Experiment Participants and procedure Experiment 2 took place interleaved with Experiment 1, so the participants and procedure were identical Stimuli Phrasal length was varied for indirect and direct objects in ditransitive sentences. The all-short condition was the base structure, with a ditransitive verb and two unmodi ed bare NP arguments. In the long indirect object (long-io) condition, the indirect object was modi ed by a relative clause, whereas in the long direct object (long-do) condition, the direct object was modi ed by the same relative clause. The relative clause was normed to insure that it was equally felicitous with each phrase. Furthermore, in order to reduce possible priming effects between this experiment and Experiment 1, the subject phrases in these sentences were marked with the topic marker ªwaº instead of the subject marker ªgaº. Table 2 provides samples of sentences. In presenting the stimulus sentences, the positions of the topic and verb remained the same as in Experiment 1: in lower-right and upper-left boxes, respectively. The positions of the indirect and direct objects were manipulated between upper-right and lower-left. Thirty sets of sentences were constructed varying phrase length and two different presentation positions Scoring The data were scored using a coding scheme that marked the order of the subject, indirect object, object, and verb. Changes of topic to subject were ignored. The coding also denoted the position of the long phrase and the same kind of deviations as in Experiment Results The dependent measure was again the percentage of sentences produced with shifted order within each cell, but now there were three kinds of shifted sentences, IO-shift (IO-topic-DO), DO-shift (DO-topic-IO), and internal-shift (topic-do-io), out of all error-free sentences. An ANOVA was performed crossing the phrase length and box location on the screen. Except where mentioned, box location was not signi cant. For the occurrences of IO-shift (Fig. 3), there was a clear effect of phrase length (F1 2; 104 ˆ23:5; F2 2; 58 ˆ18:4). Participants were more likely to shift indirect objects to the sentence-initial position when they were long than when they were short (post-hoc comparison of long-io and all-short, F1 1; 104 ˆ30; F2 1; 58 ˆ23:3). For the occurrences of DO-shift, a clear effect of phrase length was also found (F1 2; 104 ˆ12:9; F2 2; 58 ˆ11:4). Participants were more likely B52 H. Yamashita, F. Chang / Cognition 81 (2001) B45±B55 to front direct objects to the beginning of the sentence when they were long than when they were short (post-hoc comparison of long-do and all-short, F1 1; 104 ˆ20:2; F2 1; 58 ˆ16:2). There was an interaction of box location and phrase length (F1 2; 104 ˆ15:7; F2 2; 29 ˆ10), such that long elements were more likely to be shifted when they were in the upper-right box, but short elements did not differ as much between the two box locations. For the internal-shift, a robust effect of phrase length was found (F1 2; 104 ˆ109; F2 2; 58 ˆ64:5). Participants were more likely to shift direct objects in front of indirect objects but not all the way to the beginning of the sentence when they were long than when they were short (post-hoc comparison of long-io and all-short, F1 1; 104 ˆ147:7; F2 1; 58 ˆ84:7). In Experiment 2, we once again con rmed the effects of phrase length on ordering. At the same time, we found that length had somewhat different effects on indirect and direct objects. While the shift to the sentence-initial position increased in both conditions when the length was added, long direct objects shifted to the sentence-internal position more frequently than to the sentence-initial position. 5. General discussion: reconciling the directions of the shift in English and Japanese The two experiments con rmed that length directly affected phrasal ordering in production of Japanese: people tended to place long phrases before short ones, consistent with the corpus analyses. Because the short before long preference in English has been argued to be a manifestation of accessibility in an incremental production system, these results in Japanese raise questions about the universality of such processes. We now attempt to account for the difference in the preference direction by situating the shift in the theory of grammatical encoding (Bock & Levelt, 1994; Garrett, 1980). In the theory of grammatical encoding, decisions about word order depend on the properties of different levels in the production system. Some decisions are more Table 2 Example set of experimental items in Experiment 2 Phrase length condition Example All-Short Masako-wa otoko-ni okasi-o todoketa. Masako-top man-dat cake-acc delivered ªMasako delivered the cake to the man.º Long-IO Masako-wa sinbun-de syookai-sarete-ita otoko-ni okasi-o todoketa. Masako-top newspaper-in introduced man-dat cake-acc delivered ªMasako delivered the cake to the man [who was] introduced in the newspaper.º Long-DO Masako-wa otoko-ni sinbun-de syookai-sarete-ita okasi-o todoketa. Masako-top man-dat newspaper-in introduced cake-acc delivered ªMasako delivered the cake [which was] introduced in the newspaper to the man.º H. Yamashita, F. Chang / Cognition 81 (2001) B45±B55 B53 Fig. 3. Shifting in dative sentences as a function of phrase length and shift type. (Note there were no IO shifts in the Long-DO condition.) conceptually driven and others depend more on form (Bock, 1982). In the conceptual arena, there is a tendency to place salience elements earlier in sentences (Osgood & Bock, 1977), while in the form arena, the system is trying select individual lexical items in an incremental manner (Ferreira, 1996). In this theory, long phrases have competing characteristics. They are semantically richer than short phrases, because the extra lexical items enhance the meaning of the head noun, which makes them more salient. This semantic richness increases the overall accessibility of the phrase in the conceptual arena. At the same time, in the form arena, they are longer in terms of lexical content, which increases the number of possible sequences. This might increase the likelihood of simpler phrases being ready before complex phrases, and an overall preference for putting short phrases before long phrases in the form arena. These two contrary biases are delicately balanced against each other, and differences in grammatical constraints in the two languages may upset the balance in favor of one over the other. In Japanese, unlike English, the shifting of phrases all precedes the verb. Early on, the speakers have the most freedom in how they construct their sentence; they are only constrained by their message they are about to convey. In addition, because of the fairly free word order and null pronouns, NPs in Japanese are subject to few B54 H. Yamashita, F. Chang / Cognition 81 (2001) B45±B55 syntactic constraints. For these reasons, the language allows speakers to use word order to mark the saliency of different arguments. In contrast, English has a fairly strict word order that requires all arguments to be overtly present. In such a language, syntactic structure is a crucial constraint in creating grammatical sentences. Furthermore, weight-based shifts including heavy NP shift in English take place after the verb, and there is evidence that verbs exert strong in uence over the phrases that follow them, including their order (Stallings et al., 1998). These syntactic constraints presumably reduce the in uence of conceptual factors, allowing form-related factors to play a larger role in phrasal ordering. This approach explains shifts in both languages as a competition among a variety of factors during incremental production. In Japanese, the extra saliency of long NPs leads to their earlier placement. At the same time, the saliency of long NPs must compete with the accessibility of subjects/topics, or the grammatical tendency to put these rst. If the subject/topic wins at the rst position in the sentence, then the long NP competes with the other object for the next position, where they are more evenly matched in saliency. In English, weight-based shifts seem to be less sensitive to conceptual factors. Stallings et al. (1998) found that animacy did not in uence the shift, but lexical information (verb bias) did. If the shifts in human languages are m
Related Search
Similar documents
View more...
We Need Your Support
Thank you for visiting our website and your interest in our free products and services. We are nonprofit website to share and download documents. To the running of this website, we need your help to support us.

Thanks to everyone for your continued support.

No, Thanks