The implicit prosody hypothesis and overt prosody in English

Abstract

This study investigates the validity of the Implicit Prosody Hypothesis (IPH) by examining default phrasing in English, a low attachment language, in overt prosody generated by reading aloud sentences where a complex noun phrase serves as the head of a relative clause (NP1 NP2 RC). The prosodic phrasing of 27 sentences collected from 36 speakers was transcribed by three ToBI-trained labellers. Results show that, counter to the predictions of the IPH, the most common prosodic phrasing was (NP1 NP2)//(RC), which would be expected for high attachment preference languages. This default phrasing was found to be influenced by the length of the RC and by syntactically disambiguating properties of the RC verb (i.e., number agreement) only when the RC was short. It was suggested that the prosody generated in silent reading would not necessarily be the same as the prosody generated in reading aloud, especially when reading without skimming the material in advance. Based on the current results and data from previous studies, various ways to access implicit prosody are proposed.

Keywords:

• Implicit prosody
• Overt prosody
• Default phrasing
• Read aloud

Acknowledgements

This work was supported by a UCLA Faculty Research Grant. Many thanks to Michael Wagner and five anonymous reviewers for providing insightful suggestions and helpful comments. I would also like to thank Janet Fodor, Kiwako Ito, Shari Speer, and the audience at the Conference on Prosody and Language Processing held at Cornell University in 2008 and to Bruce Hayes, Pat Keating, and other members of the UCLA Phonetics Lab for their suggestions and feedback. Special thanks to Janet Fodor and Anouschka Bergmann for sending me helpful references; to J'aime Roemer, Molly Shilman, and Chad Vicenik for their help in prosodic transcriptions; to Timothy Arbisi-Kelm, Peter Graff, and Matthew Weitz for data collection; to Sameer ud Dowla Khan for proofreading the manuscript; and to Kristine Yu, Megha Sundara, Amy Schafer, and UCLA Statistics Consulting Group for their help in statistics.

Notes

¹The attachment preference is not fixed even in the same language. The preference varies depending on a variety of factors such as the type of adposition within the complex noun phrase (theta-assigning or not), the type of modifiers (adpositional phrase or relative clause), and the early or late stage of processing (e.g., Carreiras & Clifton, 1993, 1999; De Vencenzi & Job, 1994; Brysbaert & Mitchell, 1996). The attachment preference discussed in this paper is confined to the attachment of the RC in the “NP1 of NP2 RC” structure, at a later stage of processing.

²We did this so that the attachment decisions made by our subjects were not influenced by their familiarity with the sentence materials. However, we later discovered that subjects performed very similarly regardless of how well they remembered the material. Thus, in Jun and Kim's experiment on Korean, the attachment data were collected immediately following the reading task.

³However, a direct mapping between the attachment and prosodic phrasing of an individual sentence was not good (about 63%). Similar results were found in the studies by Bergmann and her colleagues (see the Discussion section). It was suggested in Jun and Koike (2003) that the correlation is the result of group behavior. It may be a by-product of the common default phrasing being (RC)//(NP1 NP2) and high attachment preference in Japanese.

⁴Specifically, the number bias means “syntactically disambiguated towards NP1 or NP2” and the gender bias means “semantically or pragmatically disambiguated towards NP1 or NP2”. In this paper, the term “bias” is used for convenience of naming. I thank the reviewer who pointed out the ambiguity and possible inappropriateness of the term “bias”.

⁵The experiment was not designed to test any effect of bias types on phrasing, so the number of sentences of the two bias types (Number/Gender) is not balanced. The nineteen Number bias sentences include eight Long RC, four Med RC, and seven Short RC sentences, and the eight Gender bias sentences include one Long RC, five Med RC, and two Short RC sentences.

⁶The three NP Bias types were generated from the same sentence by changing the gender or number of relevant head noun or the RC. However, the three RC Length type sentences were not related to each other. That is, Long RC sentences were not created by adding more words to Short RC sentences. Therefore, the data is not designed to compare acoustic durations of the target phrase across NP Bias or RC Length conditions.

⁷In general, high (H) tones were more common than low (L) tones after NP1: H occurred 70% of the time in Labeller 1's data, 85% for Labeller 2, and 91% for Labeller 3. After NP2, however, H occurred 35% of the time in Labeller 1's data, 54% for Labeller 2, and 75% for Labeller 3. A higher rate of disagreement among the labellers after NP2 is likely due to the difficulty of categorising tones in a reduced pitch range. However, within each labeller, the distribution of H and L tones after NP1 and NP2 was similar across the RC Length and NP Bias types.

⁸This method was used in calculating labeller agreement in English ToBI data, labelled by Silverman and colleagues (1992). Agreement was calculated across all possible pairs of transcribers for each word. For example, four labellers (a, b, c, and d) would produce six possible transcriber pairs (ab, ac, ad, bc, bd, and cd). If three of four transcribers (a, b, and c) agree, only three of six pairs will match (ab, ac, and bc; but not ad, bd, and cd), and this would result in 50% agreement. If two out of three labellers (a and b) agree, only one of three pairs will match (ab, but not ac and bc), resulting in 33.3% agreement.

⁹This simplification was introduced because the great majority of break indices were either 1 or 3; the small number of breaks transcribed with particular labels made the χ² test inappropriate for the data in its original form.

¹⁰The item is not included as a random factor because items are not repeated for RC Length.

¹¹Results from the linear mixed effects model showed that the effect of Long RC and its interaction with NP1-bias are significant for Labeller 2, but this is not considered meaningful because this is based entirely on one Long RC sentence.

¹²Quinn et al. (2000) and Lovric et al. (2001) estimated phrasing based on phonetic data (i.e., f0 values and the duration of the local target nouns) instead of defining it phonologically. They also interpreted the relation of f0 values the same when determining the phrasing in the three languages (i.e., French, English, and Arabic). Please see Jun (2003, pp. 222–223) for concerns regarding the methodology.

¹³It is therefore also possible that the higher ratio in the NP1-bias by Gender and Short RC condition (Figure 6) is not because speakers processed the gender-bias meaning but because they were influenced by the syntax factor, i.e., a break before RC. In either way, the result on Gender bias should not be given much significance in the current study because the result is based on very small data.

¹⁴I would like to thank one of the reviewers who directed my attention to the timing of prosodic planning relative to the head noun and bias types and the possibility of prosodic repair.

¹⁵One of the reviewers asked if the break after the RC in Japanese and Korean (from Jun & Kim, 2004; Jun & Koike, 2003) is bigger than the break before RC in English (the current study). In the prosodic structure of both English and Korean, the most common pro sodic break between the RC and the adjacent noun is an Intermediate Phrase (ip) boundary. However, we cannot directly compare the realisations of ip in these two studies because the length and location of the RC is different between the two languages. If we compare the phonetic realisation of the ip in general between the two languages, the English ip has more stable phonetic cues than the Korean ip, as the former normally involves pre-boundary lengthening, phrase accent tone, and pitch reset, while the latter is marked by an optional boundary tone, pitch reset, and minor pre-boundary lengthening. For Japanese, the corresponding break is an Intonation Phrase (IP) boundary in the J_ToBI model, which combines the Intermediate Phrase and Utterance of Beckman and Pierrehumbert's (1986) model. Since the percentage of subcategories (i.e., ip-type or Utterance-type) of Japanese IP is not available in Jun and Koike, we cannot make direct comparisons between English and Japanese, either.

¹⁶Silent/implicit prosody employed in the off-line questionnaire task would be the same as that in the on-line eye-tracking task. But this prosody might be different from that which is employed in the on-line self-paced reading task: in self-paced reading, the reading material is constrained by segmentation chunks, which affect prosody.