Examining the Implicit Prosody Hypothesis in Korean: RC Attachment and Prosodic Phrasing, Papers of Korean Language and Literature

Download Examining the Implicit Prosody Hypothesis in Korean: RC Attachment and Prosodic Phrasing and more Papers Korean Language and Literature in PDF only on Docsity! Default Phrasing and Attachment Preference in Korean Sun-Ah Jun and Sahyang Kim Department of Linguistics University of California, Los Angeles jun@humnet.ucla.edu; sahyang@ucla.edu Abstract This paper tests the validity of the Implicit Prosody Hypothesis (IPH) (Fodor 1998, 2002) based on production and perception experiments on Korean data. IPH states that attachment of a relative clause (RC) in a sentence with a complex noun phrase is influenced by a default prosodic contour of the structure projected in silent reading. It predicts that speakers of a language who prefer high attachment would produce a prosodic break between the RC and the adjacent noun phrase. Results show that Koreans prefer high attachment of an RC, but they do not produce a larger prosodic break between the RC and the adjacent noun phrase. Instead, the most common default phrasing is to produce each word in the same prosodic unit, an Accentual Phrase. Though this does not support IPH, the perception data show some sensitivity to prosody. 1. Introduction Studies have shown that languages differ in their preference of RC (relative clause) attachment when an RC modifies a complex noun phrase. For example, in the sentence ‘Someone shot the servant of the actress who was on the balcony’, the RC (‘who was on the balcony’) can modify the whole NP (‘the servant of the actress’, i.e. the servant), known as high attachment, or the NP immediately adjacent to the RC (‘the actress’), known as low attachment. High attachment is preferred by speakers of Dutch, French, German, Japanese, and Spanish, while low attachment is preferred by speakers of English, Arabic, Norwegian, Romanian, and Swedish (see [1, 2] for more details). This apparent cross-linguistic difference in the resolution of RC attachment has raised some dilemmas for the universalist view of sentence processing [3, 4, 5], which hypothesizes that the human sentence processing mechanism is innate and universal. There have been several attempts to explain the cross-linguistic differences (e.g. Tuning, Two Factor Model, Construal, Attachment-Binding, and Implicit Prosody), and this paper tests the validity of the Implicit Prosody Hypothesis (IPH) proposed by Fodor [6, 7]. The IPH states that a default prosodic contour projected in silent reading favors the syntactic analysis associated with the default prosody for the construction. Since languages differ in their prosody, attachment preferences would differ across languages. This hypothesis is supported by the effect of prosody on attachment resolution. Short RCs tend to attach low and long RCs tend to attach high. [8] found that speakers interpret a prosodic break before an RC (in a sequence of NP1 NP2 RC) as a marker of a stronger syntactic boundary, which prompts high attachment (i.e., RC modifying NP1). This suggests that speakers of a language with high attachment preference would tend to produce a prosodic boundary between the RC and the adjacent noun in the default phrasing of a sentence, but not between the two nouns (i.e. NP1 and NP2). This prediction was tested in production and perception experiments on Japanese data [9]. Since Japanese speakers are known to prefer high attachment [10], it was expected that they would produce a larger prosodic boundary between the RC and the adjacent noun phrase than between the two noun phrases (i.e., RC//NP1-NP2 in Japanese word order). Subjects read sentences where the target phrase, RC NP1 NP2, was located either sentence initially or medially. Target sentences were also varied in the length and the accentedness of the RC, NP1, and NP2. Subjects either briefly skimmed the sentence before reading (Skimming group) or read without skimming (Non-skimming group). Three weeks later, the same subjects answered a questionnaire about the attachment of the RCs for the same target sentences. Results showed that Japanese speakers, both the skimming and non-skimming groups, preferred high attachment (66%), confirming previous findings; and 67.15% of the time, they produced a larger prosodic boundary between the RC and NP1 than between NP1 and NP2 (called ‘early’ boundary), supporting the IPH. The length and the location of the target phrase (RC, NP1, and NP2) influenced their default phrasing as well as their interpretation of RC attachment. ‘Early’ boundary was found more often in long RC sentences and less often in short RC sentences compared to sentences with default length RCs. In perception data, long RCs were interpreted as high attachment more often than default length RCs, but short RCs did not differ in interpretation from the default length RCs. ‘Early’ boundary was found more often when the target phrase was in sentence-initial position than in medial position. But the attachment was not influenced by the location of the target phrase. Finally, though subjects in both groups produced an early boundary most of the time, the subjects in the skimming group had a tendency to produce a ‘neutral’ boundary (RC/NP1/NP2, i.e., no sub-grouping among the three) more often than those in the non-skimming group (26% vs. 24%). In sum, the data showed that the default phrasing of the same structure is quite consistent across subjects, and that there is a correlation between the prosodic pattern and of a structure and the interpretation of the structure. In the current paper, we will examine how native speakers of Korean interpret the same structure, and whether their default phrasing is correlated with the preference of RC attachment, if there is one. Jun [11], based on informal observation, found that native speakers of Korean prefer high attachment, but the attachment preference of an RC by Korean subjects has not been systematically examined. If Korean speakers indeed prefer high attachment, the IPH would predict that most of the target phrases be produced with an ‘early’ boundary as in Japanese. However, this may not be the case because Schafer & Jun [12] found that in the structure of ‘Adjective NP1-possessive NP2’ (e.g., wise baby’s daddy) the most common default phrasing in Korean was to produce each word in one AP. It would be interesting if the same strategy is used in producing the target structure ‘RC NP1 NP2’. The results from Korean would also be interesting because Korean and Japanese differ but also share certain prosodic and morphosyntactic features. Both languages have the same word order (‘Subject Object Verb’ and ‘RC NP1-possessive NP2’ order), but modifiers, such as an RC, are morphologically marked in Korean but not in Japanese. Both languages have an Accentual Phrase (AP), a prosodic unit larger than a Word and smaller than an Intonation Phrase, but Japanese (Tokyo) has lexical pitch accent while Korean (Seoul) does not. Thus, Korean AP tones are purely phrasal tones while Japanese AP tones are influenced by lexical tones [13, 14]. 2. Method 2.1. Subjects Thirty native speakers of Seoul Korean participated in both the production and the perception experiment. They were students at UCLA, and their period of stay in the U.S. varied from 3 days to 8 years. Subjects were randomly divided into two groups (15 speakers each, with 8 females and 7 males): NSkim (no-skimming) and YSkim (yes-skimming). 2.2 Material Four types of sentences were created by varying the length of the relative clause (RC) in the target phrase “RC NP1 NP2” and by varying the location of the target phrase. The target phrase was located at the beginning of a sentence in three types (DEFAULT, LONG, SHORT), and in the middle of a sentence for one type (DEFAULT-MEDIAL). The sentence-initial types differed in the length of the RCs: 6-7 syllables in DEFAULT, 9- 10 syllables in LONG, and 3-4 syllables in SHORT. The sentence medial type (DEFAULT-MEDIAL) had the same target phrases as in the DEFAULT sentence initial type, but differed from the DEFAULT type by locating the target phrase sentence medially. There were eight sentences in each of the four types, creating thirty-two target sentences (these target sentences were chosen after a pretest where 30 speakers in Seoul, Korea rated the degree of bias for NP1 or NP2 in 60 sentences, 15 in each of the four types). The length of the target NP1 and NP2 did not vary across the types: NP1 had 2-3 syllables and NP2 had 3-5 syllables. Twenty four filler sentences were added to the production data, and 32 fillers were added to the perception data. The fillers were made to balance the direction of bias (NP1 or NP2) for perception, and to balance the length and the location of RC for production. 2.2. Procedures 2.2.1. Production The 32 target sentences were pseudo-randomized together with 24 filler sentences. Subjects in the Yskim (Yes-skim) group were instructed to read each sentence after briefly skimming the whole list of sentences. Subjects in the Nskim (No-skim) group were instructed to read each sentence without skimming first. Speakers read each sentence at a normal rate two times in the sound booth at the UCLA Phonetics Lab. As in the experiment on Japanese data, we varied the option of skimming vs. non-skimming in order to match closely the default prosody employed by native speakers in their silent reading during the off-line and on-line sentence processing experiments, respectively. Allowing time to skim the sentences before reading and asking subjects to repeat each sentence two times allowed us to investigate how stable the default prosody was, given the different amounts of semantic information available to them before reading. All utterances were digitized using PitchWorks (Scicon), and the prosodic phrasing was determined by the intonation pattern of the utterance and the degree of juncture between words, following the conventions described in Korean ToBI [14, 15] 2.2.2. Perception The subjects who participated in the production experiment also participated in the perception experiment, an off-line questionnaire experiment. The perception experiment was done right after the production experiment. The questionnaire consisted of the 32 target sentences used in the production experiment. The target sentences were pseudo-randomized with 32 filler sentences. A comprehension question about each sentence was added below the sentence (e.g., who was on the balcony?) and two choices were given as a possible answer (e.g. the actress, the servant). The order of NP1 and NP2 was randomized but balanced so that half of the questionnaire started with NP1 and the other half started with NP2. Subjects were asked to mark an answer on a paper. The production and the perception experiments together took about an hour, and subjects were paid for their participation. 3. Results 3.1. Production The RC, NP1, and NP2 sequence was phrased in several ways but grouped into three types depending on the sub-grouping among the three components. The first type was RC//NP1- NP2 (a larger prosodic boundary between RC and NP1 than that between NP1 and NP2), called ‘early’ boundary. This includes when the prosodic boundary after RC is either an Intonation Phrase (IP) (i.e. {RC}IP{NP1-NP2}) or an Accentual Phrase (AP), defined by a final rising tone (LH) (i.e. {RC}AP{NP1-NP2}) [14]. An example pitch track of this category is shown in Figure 1. The target phrase shown in the figure is from a sentence “the wife of the department chief who won the lottery has decided to travel to Europe next year” (the target phrase is italicized): poGkwEne ‘lottery’, taQcEMtweN ‘win-modifier’, kwajaQnime ‘department chief’s’, puinIN ‘wife-TOP’. Here, the RC is marked by an IP boundary tone (HL%), and NP1 is marked by an AP boundary tone (Ha). taQcEMtweN kwajaQnime puinINwords Ha(L) L HL% (L) Ha L HL%tones 100 150 Hz 350 700 1050 1400 1750 ms poGkwEne IP AP IP RC NP1 NP2 AP Figure 1: Example of ‘early’ boundary -- RC//NP1-NP2 The second type was RC-NP1//NP2 (a larger boundary between NP1 and NP2 than that between RC and NP1), called