Word frequency effects in sound change as a consequence of - - PowerPoint PPT Presentation

Word frequency effects in sound change as a consequence of perceptual asymmetries: An exemplar-based approach

Paper by S. Todd, J. B. Pierrehumbert, J. Hay Presenter: Sven Kirchner Seminar: Exemplar Theory, Prof. Dr. Bernd Möbius SS 2020

Contents

• Introduction
• Model implementation
• Model description
• Modeling single and two category movement
• Conclusion

Contents

• Introduction
• Model implementation
• Model description
• Modeling single and two category movement
• Conclusion

• People spend more time listening than speaking
• Listened speech stored in memory
• Importance of listener over speaker
• Speaker-based models may overpredict speaker’s avoidance of ambiguity
• Use of listener-turned-speaker model
• Papers deals with so-called regular sound change

• Regular sound change :

Gradual transformation of the phonetic realization of a phoneme over time

• Regular sound change affects phonemes at different rates for different

hypotheses:

• Neogrammarian hypothesis: Change affects all phonemes at the same rate, principle
• f strict modularity
• Frequency Actuation Hypothesis: Word frequency effects different depending on

motivation of change

Frequency Actuation Hypothesis

• Philips (1984): physiologically motivated changes and non-physiologically

motivated changes

• /t/-tapping as physiologically motivated change, reducing articulatory effort
• > high frequency words predicted to change faster
• deletion of glides after coronal stops as non-physiologically motivated

change -> high frequency words predicted to change slower

Three main studies

• /t/-glottaling in

Manchester English Affects all words at the same rate

• > not supported by

FHA

• /t/-tapping in New

Zealand English Affects high- frequency words faster than low- frequency words

• > supported by FHA
• /ɛ/ - raising in New

Zealand English Affects high- frequency words slower than low- frequency words

• > not supported by

FHA

Contents

• Introduction
• Model implementation
• Model description
• Modeling single and two category movement
• Conclusion

/t/-glottaling

• Model as a directed phonetic drift of a single isolated phoneme
• Sound change only affects /t/, unlikely to affect other phonemes
• > single phoneme category subject
• Expect low and high-frequency words to change at same rate

/t/-tapping

• Produced more like an existing phoneme: /d/
• Competition between two phonemes
• Expect model to show faster movement for high frequency words

/ɛ/-raising

• Interaction between phonemes /æ/ and /ɛ/
• Model as system containing two phonemes where one is biased towards the
• ther
• Expect model to show slower movement for high-frequency words

Model desiderata

• 1. Generate movement for each category
• 2. Maintain shape of each category
• 3. Maintain distance between categories (two category model only)
• 4. Maintain overlap of two categories (two category model only)

• A) Intrusive Force
• B) Spreading Force
• C) Repulsive Force
• D) Squeezing Force

• Repulsive force pushes overlapping parts away from each other, enforcing

aversion of acoustic ambiguity

• High frequency words robustly recognized in face of ambiguity, thus less

prone to repulsive force

• Model uses exemplars, episodic traces of experienced instances
• Basis of comparison for categorizing other instances of spoken words
• Misconception of exemplar-based models privileging high-frequency words

Contents

• Introduction
• Model implementation
• Model description
• Modeling single and two category movement
• Conclusion

• Model as a production-perception loop
• Phonemes in question are vowels
• Words as monosyllabic
• Three different representations: category, type, exemplar
• Number of exemplars as type frequency
• Arranged in an exemplar space

Categories

• Phonemes representing an abstract generalization over experienced

instances of that phoneme in words

• E.g map, lab, cat for /æ/

Types

• Type consisting of phonological frame and category
• E.g. word map with phonological frame m_p and category /æ/

Exemplars

• Exemplars as detailed memory trace of instances of that type
• Slightly different realizations of a category
• Simulations include 492 exemplars per category

• Selection of a type weighted by frequency

• Selection of exemplar out of type, independent of previous selections

• Shift target value by a small amount, so-called bias
• Represents influences like reduction of articulatory effort -> Intrusive force

(bias force)

• Parameter β

• Imprecision represents natural variability
• Yields spreading force (Imprecision force)
• Model parameter l

• Transmission connection between speaker and listener
• Creates closed loop in the model

• Activation of specific exemplar with a window
• Overall category activation
• Window size as parameter α

• Identification of type of token
• Not every token identified
• Token must pass future evaluation

• Using acoustic value to measure how likely token is identified as a

realization of a certain category

• Tokens that do not pass discriminability evaluation are rejected and not

stored

• Evaluation as probabilistic, less likely to be identified in overlap areas
• Repulsive force (discriminability force)
• δ as parameter, the highly the harder to pass evaluation

• How good is the chosen token in absolute terms
• Good representations easily recalled in short and long-term memory
• Evaluation as probabilistic, tokens near high activation as more likely to pass
• Squeezing force (typicality force)
• τ as parameter of model, represents activation threshold

• Storing token only if it passed both evaluations
• Overwrites one random exemplar of the same type

Contents

• Introduction
• Model implementation
• Model description
• Modeling single and two category movement
• Conclusion

Single category movement

• Goal of achieving non-distorting category movement
• Balancing parameter
• Parameter tuning in stepwise process

• Frequency shows no effect on sound change
• Bias affects sound change rates, change identical for HF and LF
• Explains lack of word frequency effect for /t/-glottaling in Manchester

English

Two category movement

• Goal of achieving non-distorted category movement
• Additionally, keep distance between categories
• Balance forces by choosing parameters accordingly

• Model successfully predicts sound change while maintaining shape and

distance

• Success attributed to discriminability threshold < 1, non-storing of tokens

that fail discriminability evaluation, typicality evaluation introduces squeezing force

• However: Model generates exact opposite of empirical data (slower rates

for HF in Pusher and higher speed for Pushee)

Enhanced model

• Literature suggests bias towards high-frequency words in perception
• Perception tests showed: ambiguous words were more likely to be identified

as words with high-frequency

• Idea: Privilege high frequency words in model

• Idea to vary discriminability threshold with word frequency
• high-frequency tokens get lower discriminability threshold, thus making it

easier for activation

Contents

• Introduction
• Model implementation
• Model description
• Modeling single and two category movement
• Conclusion

• Model appropriately generates single and two-category movement
• Success in model compared to other models due to balancing of forcing from

speaker and listener

• Listener-based approach as key to fitting to empirical data
• Word frequency-based asymmetries in perception can generate effects on sound

change

• Simultaneously no asymmetries in production