computing segmental and suprasegmental information in lexical … · 2020-01-26 · garcia (mcgill)...

IntroductionMethods

Results & Discussion

Computing segmental and suprasegmentalinformation in lexical decision

Guilherme D. Garcia

McGill University

24th mfm

Garcia (McGill) Suprasegmental information and lexical decision

http://www.guilherme.ca

IntroductionMethods


Today

Lexical decision & stress

Speakers take longer to recognize words if stress coin-cides with the segmental point of recognition.

Ï Hypothesis: when crucial segmental and suprasegmentalinformation need to be processed at (nearly) at the same time,reaction time slows down



IntroductionMethods


IntroductionRole of stress?Portuguese stress

MethodsLexical decision taskStimuliSpeakersHypothesis

Results & DiscussionDataAnalysisConclusion



IntroductionMethods


Role of stress?Portuguese stress

Known segmental effects on lexical accessCutler et al. 1987, Sommers 1996, Luce and Pisoni 1998, Vitevitch and Luce 1998

Ï Word frequencyMore frequent words → faster RTs

Ï Neighbourhood densityWords with sparse phonological neighbourhoods → faster RTs(cf. Vitevitch and Rodríguez 2005 for Spanish)

Ï PhonotacticsMore frequent patterns → faster RTs

Ï Stress...?Word-initial stress → faster RTs (Vitevitch et al. 1997)

Garcia (McGill) Suprasegmental information and lexical decision 4 of 19


IntroductionMethods



Stress & lexical access

Word-initial stress in En → faster RTs (Vitevitch et al. 1997)

Ï Confound: positional bias in English (Cutler and Carter 1987)

∴ The relationship between stress per se and RT is unclear

Today

Lexical decision task in Portuguese+ Unlike English, no bias towards word-initial stress



IntroductionMethods



Portuguese stressGeneral patterns

In Pt, a heavy (H) σ has a coda or a diphthong:

Ï Final stress if H]PWd . Else, penult stress (a)Ï Antepenult stress considered ‘irregular/idiosyncratic’ (b)

Bias: penult stress (default)

(a) caracól (‘snail’), caválo (‘horse’)(b) patético (‘pathetic’), fantástico (‘fantastic’)



IntroductionMethods


StimuliSpeakersHypothesis

Auditory lexical decision taskMethods: stimuli

Trisyllabic words (n = 360)Different syllable shapes and segmental qualitiesReal (n = 180) and nonce words pseudorandomly presented

Ï Real word → Final onset → Nonce word∴ Point of recognition (PoR): onset of final syllable

E.g. moletom (‘sweater’) → molerom

Ï Stimuli: article + noun



IntroductionMethods



Auditory lexical decision taskMethods: stimuli

Ï Recorded by a phonetically trained Portuguese speaker (f)Ï No phonetic manipulationÏ Response button only available after each stimulus

Stimuli preceded by question:‘Is this a real word in Portuguese?’



IntroductionMethods



Auditory lexical decision taskMethods: participants & variables

Participants (n = 51): native Portuguese speakersCut-off point: 80% accuracy (n = 37)

Ï Variables examined:StressBigram probability, neighbourhood density, frequency

Ï RTs scaled and centred by speaker (z-score)



IntroductionMethods



1 Hypothesis: when crucial segmental and suprasegmentalinformation need to be processed at (nearly) at the same time,reaction time slows down

2 Are phonetic cues to stress available early on?



IntroductionMethods


DataAnalysisConclusion

Neighbourhood densityMore neighbours = faster RTs

APU PU U

-1

0

1

2

0 2 4 6 8 0 2 4 6 8 0 2 4 6 8

Neighbourhood density (z-score)

Reac

tion

time

(z-sc

ore)

Ï Unlike English (Sommers 1996)

Ï But consistent with Spanish (Vitevitch and Rodríguez 2005)



IntroductionMethods



StressEarlier stress = faster RTs

Real word Nonce word

-0.2

-0.1

0.0

APU PU U APU PU UStress

Reac

tion

time

(z-sc

ore)

Ï Real words = faster RTs, consistent with previous studies



IntroductionMethods



OverviewMultilevel linear regression estimates (fixed effects)

Word: nonce

Stress: APU

Stress: U

Frequency

N. density

Bigram prob.

-0.05 0.00 0.05 0.10

← U slower than PU

← APU faster than PU



IntroductionMethods



StressBayesian estimation (Kruschke 2013)

APUµ1 vs. PUµ2

-2 -1 0 1 2 3

0.0

0.4

0.8

y

p(y

)

Data Group 1 w. Post. Pred.

N1 = 4327

-2 -1 0 1 2 3

0.0

0.4

0.8

yp

(y)


N2 = 6474

Normality

log10(ν)0.34 0.35 0.36 0.37 0.38 0.39 0.40

95% HDI0.343 0.394

mode = 0.369

Group 1 Mean

μ1

-0.32 -0.30 -0.28 -0.26

95% HDI-0.318 -0.289

mean = -0.304

Group 2 Mean

μ2

-0.32 -0.30 -0.28 -0.26

95% HDI-0.283 -0.259

mean = -0.27

Difference of Means

μ1 − μ2

-0.05 -0.04 -0.03 -0.02 -0.01 0.00

95% HDI-0.052 -0.0156

mean = -0.0336100% < 0 < 0%

Group 1 Std. Dev.

σ1

0.34 0.35 0.36 0.37 0.38 0.39 0.40

95% HDI0.355 0.383

mode = 0.367

Group 2 Std. Dev.

σ2

0.34 0.35 0.36 0.37 0.38 0.39 0.40

95% HDI0.361 0.386

mode = 0.373

Difference of Std. Dev.s

σ1 − σ2

-0.02 -0.01 0.00 0.01

95% HDI-0.0206 0.0103

mode = -0.00427

Effect Size

(μ1 − μ2) (σ12 + σ2

2) 2

-0.15 -0.10 -0.05 0.00

95% HDI-0.14 -0.042

mode = -0.0904100% < 0 < 0%

PUµ1 vs. Uµ2

-2 -1 0 1 2 3

0.0

0.4

0.8

y

p(y

)Data Group 1 w. Post. Pred.

N1 = 6474

-2 -1 0 1 2 3

0.0

0.4

0.8

y

p(y

)


N2 = 2148

Normality

log10(ν)0.36 0.38 0.40 0.42

95% HDI0.363 0.423

mode = 0.395

Group 1 Mean

μ1

-0.30 -0.25 -0.20 -0.15

95% HDI-0.28 -0.255

mean = -0.267

Group 2 Mean

μ2

-0.30 -0.25 -0.20 -0.15

95% HDI-0.215 -0.169

mean = -0.192

Difference of Means

μ1 − μ2

-0.10 -0.08 -0.06 -0.04 -0.02 0.00

95% HDI-0.101 -0.0497

mean = -0.075100% < 0 < 0%

Group 1 Std. Dev.

σ1

0.36 0.38 0.40 0.42 0.44 0.46

95% HDI0.367 0.394

mode = 0.38

Group 2 Std. Dev.

σ2

0.36 0.38 0.40 0.42 0.44 0.46

95% HDI0.398 0.44

mode = 0.418

Difference of Std. Dev.s

σ1 − σ2

-0.06 -0.05 -0.04 -0.03 -0.02 -0.01

95% HDI-0.0592 -0.016

mode = -0.0384

Effect Size

(μ1 − μ2) (σ12 + σ2

2) 2

-0.25 -0.20 -0.15 -0.10 -0.05 0.00

95% HDI-0.251 -0.124

mode = -0.184100% < 0 < 0%



IntroductionMethods



Phonetic cues

1 Hypothesis: when crucial segmental and suprasegmentalinformation need to be processed at (nearly) at the same time,reaction time slows down

2 Are phonetic cues to stress available early on?Ï Given the relative nature of stress:

How early can speakers perceive APU stress?



IntroductionMethods



Phonetic cuesDuration

Main phonetic correlate in Pt is duration: (Major 1985)

APU PU U

0.15

0.20

0.25

0.30

APU PU U APU PU U APU PU U

Syllable

Dur

atio

n (m

s)

Ï Overall, stressed syllables are considerably longer



IntroductionMethods



Phonetic cuesIntensity

Besides, note amplitude in APU σs:

APU PU U

45

50

55

60

APU PU U APU PU U APU PU U

Syllable

Ampl

itude

(dB

)

∴ APU stress is phonetically cued early



IntroductionMethods



Conclusion

Ï Results are consistent with the hypothesis:U stress ‘coincides’ with PoR → slower RTStimuli contain sufficiently robust phonetic cues to stress

Ï Stress plays a clear role in lexical access in these dataÏ Distance matters: segment vs. syllable?



IntroductionMethods



Next steps

Ï Prediction: whenever stress is aligned with PoR → slower RTsThus, if PoR is varied within each stress pattern:

APU PU U

APU ↗PU ↗U ↗



References I

Cutler, A. and Carter, D. M. (1987). The predominance of strong initialsyllables in the English vocabulary. Computer Speech & Language,2(3-4):133–142.

Cutler, A., Mehler, J., Norris, D., and Segui, J. (1987). Phoneme identificationand the lexicon. Cognitive Psychology, 19(2):141–177.

Kruschke, J. K. (2013). Bayesian estimation supersedes the t test. Journal ofExperimental Psychology: General, 142(2):573–603.

Luce, P. and Pisoni, D. (1998). Recognizing Spoken Words: The NeighborhoodActivation Model. Ear and Hearing, 19(1):1–36.

Major, R. C. (1985). Stress and rhythm in Brazilian Portuguese. Language,61(2):259–282.

Sommers, M. S. (1996). The structural organization of the mental lexicon andits contribution to age-related declines in spoken-word recognition.Psychology and aging, 11(2):333.

Garcia (McGill) Suprasegmental information and lexical decision References


References II

Vitevitch, M. S. and Luce, P. A. (1998). When words compete: Levels ofprocessing in perception of spoken words. Psychological science,9(4):325–329.

Vitevitch, M. S., Luce, P. A., Charles-Luce, J., and Kemmerer, D. (1997).Phonotactics and syllable stress: Implications for the processing of spokennonsense words. Language and speech, 40(1):47–62.

Vitevitch, M. S. and Rodríguez, E. (2005). Neighborhood density effects inspoken word recognition in Spanish. Journal of Multilingual CommunicationDisorders, 3(1):64–73.

Garcia (McGill) Suprasegmental information and lexical decision References


Acknowledgments

Thanks to Heather Goad, Morgan Sonderegger, Natália Brambatti Guzzo,and the Montreal Language Modelling Lab (MLML).

Thank you!

Slides on guilherme.ca

Garcia (McGill) Suprasegmental information and lexical decision f

http://mlmlab.org/people/



computing segmental and suprasegmental information in lexical … · 2020-01-26 · garcia (mcgill)...

Documents