Looking for exemplar effects: testing the comprehension and memory - - PowerPoint PPT Presentation
Looking for exemplar effects: testing the comprehension and memory ry representations of f r' r'duced words in in Dutch le learners of f French LISA MORANO, LOUIS TEN BOSCH, MIRJAM ERNESTUS (2019) 1 LILAS MAUBORGNE EXEMPLAR THEORY IN
LISA MORANO, LOUIS TEN BOSCH, MIRJAM ERNESTUS (2019)
LILAS MAUBORGNE EXEMPLAR THEORY IN LANGUAGE AND SPEECH SCIENCE 01.07.2020
1.
2.
3.
4.
5.
Most experiments investigating exemplars
Native listeners (L1) Almost never second language learners (L2)
Reason for an investigation on exemplars in L2:
If exemplars important in speech comprehension → findings for L1 listeners should be
the same for L2 listeners
Provide information about which acoustic details are stored in exemplars Are exemplars true representation of the acoustic signal? Are they affected by the listener’s knowledge?
L1 phonological filter:
L2 learners perform well in a phoneme categorization task but not always in lexical decision tasks L2 listeners can perform simple low-level tasks in phonetic mode As soon as linguistic processing is needed their L1 phonological filter keeps them from processing
the stimuli like L1 listeners
Exemplars formed before phonological filter applies: L2 exemplars can encode L1 irrelevant
variation
Exemplars formed after phonological filter applies: L2 exemplars encode less L1 irrelevant
variation
Existing exemplar research in L2 provides evidence that L2 listeners can store exemplars Main questions: Are
e L2 L2 intermediate lea earners ab able le to to en encode, in n the form
lars, fin fine lin inguis istic de details s that ar are e no not t rele elevant to to their L1 L1 an and to to use use them for
speech com
Can
Can exemplar effects s in L2 L2 listeners al also so be be fou
hen manipulating reg egular pho phonetic vari ariation ins nstead of
ariation?
Red
eduction = the weakening or deletion of phonemes or even whole syllables
Occurs in informal speech Most other experiments: replaced one allophone with another allophone
Listeners could have stored several abstract representations rather than different exemplars Categorial variation: more difficult to attest for the role of exemplars
Reduction: can’t be stored abstractly
May result in an infinite number of realization All activate same abstract pronunciation variant of the word
Red
eduction = the weakening or deletion of phonemes or even whole syllables
Here: Reduction phenomenon of phrase-medial high vowel devoicing in French “la cité” (the city) /i/: more or less devoiced
→ voicing fails to be re-established in time after devoiced consonant /s/
Phrase-medial high vowel devoicing has never been reported in Dutch Part of sound pattern of French but not of Dutch
Lexical decision task in French for Dutch intermediate French learners Experimental words: High vowel + voiceless consonant
Pronunciation match: prime and target both reduced or not reduced Pronunciation mismatch: prime is reduced, target is not or vice-versa
Expectations:
Faster reaction to a target when it matches with prime L2 participants show exemplar effect Are able to store phonetic information that doesn’t happen in L1
120 Dutch students Studied French in high school (4 to 7 years) 18-29 years old One of the 3 conditions randomly assigned
Ex
Experim imental l wor
ds:
24 most frequent from vocabulary of beginners’ textbook in French for Dutch
learners
Bisyllabic
1st syllable: High vowel: /i/, /y/, /u/ followed by voiceless consonant “la cité” (the city) “le future” (the future) “la pou
Frequency of occurrence: 0.71 per million words to 107.92
Ex
Exis istin ing-word fi fill llers
78 bisyllabic frequent words form the
same textbook
No restriction
Pseu
eudo-word fille fillers
102 created Phonotactically legal syllable added to 1st
syllable of all experimental and existing- word fillers
Stimuli Experimental words Voiced Devoiced Existing-word fillers Pseudo-word fillers
Stimuli recorded with their definite determiners by
first author of the paper
Devoiced stimuli: word’s 1st syllable = completely
devoiced high syllable
Voiced stimuli: word’s 1st syllable = fully voiced high
syllable
Waveforms and spectrograms of target “le silence” (the silence)
Voiced Devoiced
/i/ /i/
Stimuli Experimental words Voiced A B Devoiced A B Existing-word fillers Pseudo-word fillers
Two Tokens for each voicing type:
Voiced and devoiced token A:
On average 805 ms long
Voiced and devoiced token B:
On average 811 ms long
Lexical decision task:
Two blocks of 132 trials One block = 12 experimental words Experimental words repeated
Variant match = both prime and target have
(de)voiced vowels
Variant mismatch = prime is devoiced and target is
voiced or vice-versa
Prime-targets separated by 7-97 trials Remainder of the trials:
36 bisyllabic real-word fillers 48 bisyllabic pseudo-word fillers Block x2 Experimental words x12 Match Prime = voiced Target = voiced Prime = devoiced Target = devoiced Mismatch Prime = voiced Target = devoiced Prime = devoiced Target = voiced Real-word fillers x36 Pseudo-word fillers x48
Lexical decision task:
Practice trials:
6 real-word fillers 6 pseudo-word fillers Same for all participants Similar in frequency of occurrence and phonological structure to stimuli of experiment
5 pseudo-randomization of trials
Four different stimulus lists in each Each experimental word tested for each of four possible matching and
mismatching combinations
Con
Condit ition AB: B:
Different tokens for the primes and the
targets
Match: prime A and target B are different
recordings
Primes and targets: Same pronunciation variant Different durations
Co
Condit itions AA an and BB BB:
AA: only tokens A BB: only tokens B Match: same token and same duration
Average absolute temporal differences between primes and targets per condition
Stimuli Experimental words Voiced A B Devoiced A B Existing-word fillers Pseudo-word fillers
AA BB AB
Participants: Listened to the words Push a yes or no button Say as fast as possible whether the words were real in French or not Participants did not need to know the meaning of the words Experiment lasted less than 30 minutes
“la cité”
Accuracies:
Linear mixed effects model for logistic regression Dependent variable = probability of a correct response
Reaction times = RTs
Measured from noun onset Mixed effects regression models
Before analysis: all RTs and stimulus log-transformed
Dependent variable for linear mixed effect model = log-transformed RT
Crossed random effect factors = item and participant
Predictors of interest:
Voic
ing: categorical predictor indicating whether the first vowel of the stimulus was voiced or devoiced
Con
Condit ition: AB, AA or BB
Tok
Repetitio
ion match: whether the prime and target of the experimental word were of the same pronunciation variant
Control predictors:
Log
Log Sti Stimulus du duration
Tria
ial nu number: position of the trial in the experiment
Control for learning or fatigue effects
Prim
rime/target Di Distance (in number of intervening trial)
Log
Log RT to to the pr previous tria ial
Control local speed effects
Log
Log RT on
prim ime
Participant’s accuracy: relatively high
Overall: 83.92% Pseudo-word fillers: 85.52% Real-word fillers: 75.70%
Lower accuracy on experimental words
Experimental words less frequent than real-word fillers
Fir irst occ
Sensitivity to devoicing manipulation on primes only:
Participants more accurate on the voiced than the devoiced tokens A Simple effect of Voicing
→ Participants sensitive to devoicing for tokens A
No statistical difference for tokens B Significant interaction between Voicing and Token
→ Participants not sensible to devoicing for tokens B
Significant random slope of Voicing on Item Voicing significantly larger for some items than others
Second occ
Participants sensitive to devoicing manipulation (for tokens A)
Investigate whether they were more accurate on matching than on
mismatching targets
For targets with correctly answered primes:
No effect of Repetition match on accuracy Not as main effect Not in interaction with Condition or Token
RT data → priming across all conditions
Participants classify correctly prime and target of experimental words as real words RT target < RT prime (106 ms faster)
RTs in ms from word offset for the experimental primes and targets when both were answered correctly
Analysed statistically RTs to the targets answered correctly (whose primes were also
answered correctly)
Almost all control predictors showed significant effects Participants faster at answering targets when Also answered quickly on previous trial Recognized prime quickly Number of intervening trials between prime and target = low Stimuli short
All factors of interest showed significant effect AB vs AA and BB vs AA: effect of Repetition match differed
Conditions AB and BB patterned together against AA Token of the target better predictor than Condition
Main effect of Voicing without interaction with Token
Participants slower at processing devoiced target for both A and B Contrary to Accuracy data → RT data: participants sensitive to devoicing for both
tokens → L2 listeners sensitive to L1 irrelevant information
Repetition match
Significant interaction with Token Significant interaction with Voicing Three- way interaction not significant
→ Target = devoiced token A: participants faster at answering the target when it matched its prime
Participants faster when voiced targets B mismatched their prime Voiced token A and devoiced token B: main effect of Repetition match not significant
Devoiced primes:
Always shortened participant’s RTs on the targets
Voiced primes:
Never led to any significant differences in RTs between matching and mismatching
target
Devoiced A tokens answered to faster when preceded by devoiced A primes
Exemplar effect for devoiced A targets L2 listeners are able to encode and store in the form of exemplar phonetic variation that
doesn’t occur regularly in their L1
Exemplars formed before phonological filter applies and represent the acoustic signal
faithfully
Same information encoded for native and non-native listeners
Exemplars formed before phonological filter
Are exemplars part of mental lexicon? Hypothesis: Exemplars are stored in episodic memory
Significant interaction Repetition match/Token
Participants used different processes to comprehend B and A tokens AA and BB both used identical tokens but: didn’t pattern in the same way in participant’s RT behaviour on targets BB condition patterned with AB condition In both: no exemplar effect
→ Prime and target identical didn’t lead to exemplar effect
Contrasts with previous studies where exemplar effect also arises when prime are
target are different in the match condition
Why do example effects arise in some situations but not in others?
Time-course hypothesis: Exemplar effects occur when speech processing is slow
Explains the presence versus absence of exemplar effects: participants slower on A than on B
Doesn’t explain mismatch effects: RT slower for voiced B tokens preceded by voiced tokens
Difference in response pattern → subtle acoustic differences between A and B tokens
Voiced and devoiced tokens more different from each other in AA than in BB
Lower accuracies on devoiced A primes and lower RTs on A and B devoiced targets Difficulty of processing of A and B devoiced tokens: participants’ abstract
representation reached higher level of activation than after processing a voiced prime → Devoiced prime + voiced target
Easier to process voiced form High activation of abstract representation
→ RT mismatching target < RT matching target
Lower accuracies on devoiced A primes and lower RTs on A and B devoiced targets Devoiced A primes difficult to comprehend
Stronger individual memory traces for devoiced A than devoiced B primes
Highly activated exemplars: easier to retrieve → Devoiced prime + devoiced target
Easier to use exemplar formed in AA than in BB
→ Only match exemplar effect in AA with devoiced targets
→ Exemplar effects might only appear under testing conditions that encourage participants to rely on their recent memory
Too much variation: B tokens
→ No exemplar stand out from other episodic traces → No exemplar can be reused in matching conditions
Results: support Hanique (2014)
Hanique: Exemplars play a limited role in everyday speech comprehension Here: Exemplar effects arose in very limited conditions Significant mismatch effects in conditions which were expected to trigger exemplar
effects
Exemplar effects can also occur for L2 learners
Even for phonetic information not usual for the learner’s L1 Ex
Exemplars can en encode information that the ph phonological filt filter us usually ly di discards
Exemplars must be formed before phonological filter applies
→ Ex Exemplars s pr probably no not t part part of
Participants showed different response patterns when presented with different
tokens of the same words in the same testing conditions
Questions robustness of exemplar effects → Hanique