Moraicity in Translated versus Native Japanese Text-setting

The syllable as a prosodic unit in Japanese lexical strata: evidence from text-settingRebecca L StarrNational University of Singapore Stephanie S Shih University of California, MercedLabPhon 14National Institute for Japanese Linguistics25 27 July 2014Un- der the sea

su- ba- ra- shii

1The Mora in JapaneseJapanese = prototypical example of a mora-based language e.g., kai.zen = [ka][i].[ze][n]

Rhythmic timing depends on mora units (Homma 1981; Port et al. 1987; cf., Beckman 1982)Phonological/morphological implications: e.g., accent placement, compensatory lengtheningTraditional Japanese poetry = mora counting e.g., haiku 5 7 5 mora form22Syllable usually treated as unnecessary or unimportant in Japanese phonology (Labrune 2012).Kubozono (1999; et seq.): Japanese is (in part) a syllable-based language.accent placement, word formation, etc.

The Syllable in Japanese?33Labrune (2012): no positive psycholinguistic evidence for the cognitive reality of the syllable in Japanese.

The Syllable in Japanese?44Major lexical strata in Japanese (It and Mester 1999):Yamato (native Japanese)Ex: suki (to like)Sino-Japanese (Chinese origin)Ex: ningen (human)Foreign (~85% English origin)Ex: benchi (bench)MimeticEx: fuwafuwa (fluffy)

Lexical Strata in Japanese55Strata characterized by different phonotactics, and phonological rules (It and Mester 1999).e.g., long a never occurs in Sino-Japanese words.

Because Chinese and English are syllable-based, is it possible that the syllable is a more salient unit in the Sino and Foreign strata?Alternatively, does widespread knowledge of English contribute to increased salience of the syllable in just the Foreign stratum?Lexical Strata in Japanese66Text-setting: the pairing of language and music in song.Typologically, text-setting makes use of salient prosodic units particular to a language.English: syllables, lexical and phrasal stresses (Halle & Lerdhal 1993; Shih 2008; Hayes 2009; a.o.)Cantonese: tonal melodies matched in musical melodies (Yung 1991)Similar claims for metrical typology (Hanson and Kiparksy 1996)Evidence from Text-Setting77Japanese text-setting:Claimed to operate as a mora-based system (Kubozono 1999; Hayes and Swiger 2008; cf. Manabe 2009)

= each mora must receive (at least) one note.

e.g., do.ra.go-n bo-o.ru (Dragonball Z theme, 1989) x x x x x x x = 7 notes Evidence from Text-Setting

88BUT, multi-moraic notes show up frequently in modern Japanese songs.Similar findings in poetry (Tanaka 2012)In these settings, it is the syllable that receives at least one note.e.g., do- ra- gon boo-ru x x x x x = 5 notesEvidence from Text-Setting99What constraints govern moraic vs. non-moraic text-setting variation in Japanese?Do Japanese listeners perceive the syllable as an acceptable segmentation unit in text-setting? Is it as acceptable as the mora?

Main Research Questions1010Other factors that contribute to moraic vs. non-moraic setting: style, information density, etc.Moraic vs. syllabic treatments in Japanese text-setting are conditioned by lexical, phonological, and stylistic constraints.Strata in the Japanese lexiconInformation densityPhonological factors (e.g., sonority)Stylistic factors (e.g., musical genre)Japanese listeners fully accept syllable-based text-setting.Particularly when a moraic setting is impractical.

Main Claims1111Corpus studyExperiment (prelim. results)Two Approaches1212Three corpora of Japanese songs compared:

Corpus StudyAnime theme songsnativeLate 1980s 90sDisney songstranslatedLate 1980s 90sChristmas songstranslatedLate 19th c. 21st c.1313Four VariablesCoda-NVi (ai, ui, ei, oi)Long VowelsInter-voiceless-consonant i and u1414Example Moraic Syllabicningen (human)ni-n-ge-nnin-gensekai (world) se-ka-i se-kaihoshii (want)ho-shi-i ho-shiisuki (like)s-ki, s-kis( )ki

Example settings1515Santa appears in moraic and syllabic settings within the same song:

sa n ta no o ji sa n ga

de mo so no san ta wa

(I Saw Mommy Kissing Santa Claus, 1952, trans. 1962)Example Clip

1616Needed: sound filesResults: Difference by lexical stratum

Coda-NViLong VowelsInter-voicelessi and u17SyllabicMoraicResults: Differences by corpus18

Results: Regression modelingGeneralized linear mixed-effect model (glmer).Significantly more syllabic settings in translated song corpora than native corpus. Significantly more syllabic settings in Foreign stratum words.No reliable difference between Sino and Yamato strata words.Other phonological factors also significant for certain variables: e.g., sonority scale for diphthongs (e.g., ai vs. ei; Prince 1983)

19Syllabic text-setting is present in both translated and native corpora, but is more common in translated corpora.Foreign stratum words use significantly less moraic text-setting.Certain phonological contexts encourage syllabic text-setting (e.g., preferred consonant clusters, decreasing sonority).Summary of Results2020ExperimentCorpus study demonstrates that trained composers use both moraic and syllabic settings and prefer to use syllabic settings for Foreign stratum words.Do ordinary Japanese speakers follow the same patterns?What about Japanese learners?Perceptual experiment to test acceptability of different text-setting styles.21Novel methodology: VocaloidCreating stimuli involving multiple sung arrangements presents a challenge.We used Yamahas Vocaloid 3 software, which produces synthesized sung Japanese.

22sakura

FactorsLinguistic variables: Coda-NVi (limited to ai)Strata: Foreign vs. SinoNot possible to make minimal pairs with YamatoSelected near-minimal pairs.Ex: benchi / bengi23FactorsSettings tested:Mora: mi-n-to da-yo

Syllable:min-to da-yo-ne

Split Syllable(melisma):mi-in-to da-yo

Bad 1 (too small):m-in-to da-yo

Bad 2 (too large):mi-i-i-i-intodayo

24PredictionsNative Japanese will prefer Syll. to Split-Syll. Syllabic: (min-to da-yo-ne) 6 moras for only 5 notes. Since there is no moraic solution, listeners will be okay with syllabic.Split-syllabic: (mi-in-to da-yo) 5 moras for 5 notes, but arrangement is non-moraic anyway. Turns short vowel into long vowel. Listeners will prefer both syllable-based settings over the bad settings, which do not correspond to any proposed salient prosodic unit.25MethodologyMotsu-kun is learning to arrange lyrics to a melody. Help him improve by rating his work!

(minto dayone)

Asked to rate on 1-4 Likert scale.

26Participants & Implementation18 native Japanese speakersAsked for frequency of English use10 Japanese learners (Eng & Chi native spkrs)Asked for length of Japanese study Survey conducted online using Qualtrics.Still recruiting participants: http://bit.ly/1oblYNzData analyzed using lmer (linear mixed-model regression) in R.27Findings28

Key FindingsNative speakers rated Syllabic (min-to) just as highly as Moraic setting (mi-n-to).Split-Syll. (mi-in-to) was rated significantly lower than Syll., but significantly higher than Bad1 or Bad2.Conclusions: Native listeners prefer a one-to-one correspondence between note and prosodic unit, whether mora or syllable.When there is room in the melody for a moraic setting, syllable-based is dispreferred, but not totally rejected.29Key FindingsJapanese learners rate Moraic, Syll., and Split-Syll. settings as equally good.Moraic settings, which do not occur in English and Chinese, are just as highly rated as familiar syllable-based settings.Learners must be acquiring familiarity with moraic segmentation as part of the Japanese learning process.But learners do not pick up that Split-Syll (mi-in-to) is dispreferred.Not as sensitive to vowel length?Dont care about one-to-one match between notes and prosodic unit?

30Key FindingsTrend, but no significant differences between Foreign and Sino strata (contrary to corpus findings).No effects of English exposure or whether currently living in Japan. More participants needed.31Conclusions: ExperimentPositive evidence for the cognitive reality of the syllable in Japanese: Japanese listeners fully accept syllable-based segmentation in contexts where moraic segmentation is impractical. When moraic segmentation is available, syllable-based segmentation is dispreferred, but rated more highly than settings which segment along non-salient prosodic boundaries.32Overall ConclusionsBoth syllabic and moraic text-setting styles are prevalent and acceptable in Japanese.Text-setting style is conditioned by factors such as phonological context and (possibly) lexical stratum.Native listeners prefer a one-to-one correspondence between notes and prosodic units, whether mora or syllable.33Acknowledgements to Noriko Manabe, Reiko Kataoka, Roey Gafter, Junko Ito, Mie Hiramoto, Yosuke Sato, Sakiko Kajino, Nala Lee, and Jason Ginsburg for their input and assistance. .contact:rstarr@nus.edu.sgshih@ucmerced.edu

Thank you!

3434Beckman, Mary. 1982. Segment Duration and the 'Mora' in Japanese. Phonetica. 39. 113-135.Halle, John and Fred Lerdahl. 1993. A Generative Text-setting Model. Current Musicology. 55: 3-21.Hanson, Kristin and Paul Kiparsky. 1996. A Parametric Theory of Poetic Meter. Language. 72(2). 287-335. Hayes, Bruce. 2009. Textsetting as Constraint Conflict. In Jean-Louis Aroui and Andy Arleo (ed). Towards a Typology of Poetic Forms: From language to metrics and beyond. Amsterdam: John Benjamins. 43-62. Hayes, Bruce and Tami Swiger. 2008. Two Japanese Childrens Songs. MS. University of California, Los Angeles. 12 pages. Homma, Yayoi. 1981. Durational relationship between Japanese stops and vowels. Journal of Phonetics 9 (3): 273 281.It, Junko & Armin Mester. 1999. Japanese Phonology. In John Goldsmith (ed). The Handbook of Phonological Theory. Oxford: Wiley-Blackwell. 818-838. Kubozono, Haruo. 1999. Mora and Syllable. In Natsuko Tsujimura (ed). The Handbook of Japanese Linguistics. Oxford: Wiley-Blackwell. 31-61. Labrune, Laurence. 2012. Questioning the universality of the syllable: evidence from Japanese. Phonology 29:1, 113 - 152. Manabe, Noriko. 2009.Western Music in Japan: The evolution of styles in childrens songs, hip-hop, and other genres. Ph.D. dissertation, CUNY Graduate Center.Pellegrino, Francois, Coupe, Christophe, and Egidio Marsico. 2011. A Cross-Language Perspective on Speech Information Rate. Language. 87(3). 539 558.Port, Robert. F., Dalby, Jonathan & O'Dell, Michael. (1987). Evidence for mora-timing in Japanese. Journal of the Acoustical Society of America. 81. 1574 1585.Prince, Alan. 1983. Relating to the Grid. Linguistic Inquiry. 11:511-562.Shih, Stephanie. 2008. Text-Setting: a (musical) analogy to poetic meter. Paper presented at New Research Programs in the Linguistics of Literature. University of California, Berkeley.Tanaka, Shinichi. 2012. Syllable Neutralization and Prosodic Unit in Japanese Senryu Poems. Paper presented at Metrics, Music and Mind. Rome, Italy.Yung, Bell. 1991. The Relationship of Text and Tune in Chinese Opera. In J. Sundberg; L. Nord; and R. Carlson (ed). Music, Language, Speech, and Brain. London: Macmillan. 408-418.

Select references353511 songs total:Totoro (1988)Dragonball Z (1989)Kikis Delivery Service (1989)All-Purpose Cultural Cat-Girl Nuku Nuku (1990)Gundam F91 (1991)Bubblegum Crash (1991)Slayers (1995)Sailor Moon (1995)Tenchi Universe (1995)Slayers Next (1996)Boys Before Flowers (1996)Anime Corpus3636Films:The Little Mermaid (1989 translation)Beauty and the Beast (1991)Aladdin (1992)The Lion King (1994)The Little Mermaid (1997 translation)17 songs totalDisney Corpus373710 songs total:Hark the Herald Angels Sing (1888)O Holy Night (1909)Silent Night (1909)Jingle Bells (1958)Santa Claus is Coming to Town (1959)Rudolph the Red-Nosed Reindeer (1959)I Saw Mommy Kissing Santa Claus (1962)Winter Wonderland (1962)The Christmas Song (1996)We Wish You a Merry Christmas (2005)Christmas Corpus3838Excluded:Code-switching involving two or more consecutive words in English with different parts of speech.Ex: Its my dayContent that was spoken rather than sung.Discourse particles like aa.Nonsense words for which a stratum could not be determined.Mimetic stratum words (not enough for analysis).Coding Methodology3939Native speakers LMER

REML criterion at convergence: 422.5Scaled residuals: Min 1Q Median 3Q Max -2.12822 -0.70806 -0.07082 0.59577 2.71389

Random effects: Groups Name Variance Std.Dev. participant (Intercept) 0.2245 0.4739 Residual 0.4958 0.7041 Number of obs: 180, groups: participant, 18

Fixed effects: Estimate Std. Error df t value Pr(>|t|) (Intercept) 1.861e+00 1.703e-01 5.663e+01 10.929 1.33e-15 ***settings2 5.556e-01 1.660e-01 1.570e+02 3.347 0.00102 ** settingb1 9.444e-01 1.660e-01 1.570e+02 5.690 6.05e-08 ***settingb2 1.972e+00 1.660e-01 1.570e+02 11.883 < 2e-16 ***settingm -2.778e-01 1.660e-01 1.570e+02 -1.674 0.09619 Experiment: statistical models4040Learners LMER

REML criterion at convergence: 240.9Scaled residuals: Min 1Q Median 3Q Max -2.3779 -0.7192 0.0632 0.5096 2.3131

Random effects: Groups Name Variance Std.Dev.participant (Intercept) 0.1830 0.4278 Residual 0.5378 0.7334 Number of obs: 100, groups: participant, 10

Fixed effects: Estimate Std. Error df t value Pr(>|t|) (Intercept) 1.7555 0.2375 42.1900 7.391 3.92e-09 ***settingb1 0.7500 0.2319 83.9900 3.234 0.00175 ** settingb2 1.7238 0.2339 84.0200 7.369 1.09e-10 ***settingm -0.1369 0.2324 84.0000 -0.589 0.55744 settings2 0.2131 0.2324 84.0000 0.917 0.36181 stratsino 0.1200 0.1467 83.9900 0.818 0.41560 Experiment: statistical models4141null2795.1042