AnnouncementsAnnouncements
Next few lecturesNext few lectures– Require some syntactic knowledgeRequire some syntactic knowledge– Review Chapter 2’s Syntax SectionReview Chapter 2’s Syntax Section
ReadingsReadings– Original Articles Original Articles Greater difficulty level Greater difficulty level– Read in order as stated in syllabus.Read in order as stated in syllabus.– Statistics knowledge?Statistics knowledge?
Sample exam questionsSample exam questions– This week (Friday): I will post a few Qs in This week (Friday): I will post a few Qs in
our discussion forum.our discussion forum.– Next week (Thursday): You will submit Next week (Thursday): You will submit
your Qs into dropboxyour Qs into dropbox
Psy1302 Psy1302 Psychology of LanguagePsychology of Language
Lecture 9Lecture 9Models of Speech RecognitionModels of Speech Recognition
Continuation of Last Continuation of Last Lecture…Lecture…OutlineOutline We are fast at speech recognition.We are fast at speech recognition. How do we achieve speed?How do we achieve speed?
– Parallel ActivationParallel Activation– Constrained by contextual EffectsConstrained by contextual Effects– Terminologies and IdeasTerminologies and Ideas– Two Classic ModelsTwo Classic Models
Cohort ModelCohort Model TRACE ModelTRACE Model
[and many experimental paradigms and [and many experimental paradigms and findings]findings]
Top-down Example 1Top-down Example 1Last time: Shadowing Last time: Shadowing and Correctionsand Corrections
IntendedIntended MispronunciationMispronunciation FeatureFeature narrownarrow marrowmarrow placeplace detrimentaldetrimental tetrimentaltetrimental voicingvoicing perfectionisticperfectionistic berfectionisticberfectionistic voicingvoicing liveslives rivesrives placeplace backback mackmack mannermanner hamperedhampered kamperedkampered place & place &
mannermanner taketake nakenake mannermanner selfself zelfzelf voicingvoicing comfortcomfort vomfortvomfort all threeall three
Bottom-Up vs. Top-Bottom-Up vs. Top-Down ProcessingDown Processing Bottom-up: Bottom-up:
Processing that is Processing that is stimulus or data-stimulus or data-driven.driven.
Top-down: Top-down: Processing that Processing that involves the use of involves the use of knowledge knowledge obtained from obtained from higher-level higher-level sourcessources
Terminologies
Top-down Examples 2Top-down Examples 2Lexical Influence on Phoneme PerceptionLexical Influence on Phoneme Perception
Ganong (1980)Ganong (1980)– Splice speech wavesSplice speech waves
/d/ to /t/ + //d/ to /t/ + /æsk/æsk/ dask-task dask-task /d/ to /t/ + //d/ to /t/ + /æš/æš/ dash-tash dash-tash
– Obtained % of /d/ identificationObtained % of /d/ identification Two possible outcomes:Two possible outcomes:
– No Effect of Lexical KnowledgeNo Effect of Lexical Knowledge– Effect of Lexical Knowledge Effect of Lexical Knowledge
nonword-word: dask-taskword-nonword: dash-tash
% id
enti
fica
tion
as
/d/
short VOT (d) long VOT (t)
100
0
Top-down Examples 2Top-down Examples 2Lexical Influence on Phoneme Lexical Influence on Phoneme PerceptionPerception
Ganong (1980)Ganong (1980)
– Lexical knowledge influence perceptionLexical knowledge influence perception– Only able to shift AMBIGUOUS phones Only able to shift AMBIGUOUS phones
and not those at the ends of continuumand not those at the ends of continuum
Top-down Examples 2Top-down Examples 2Lexical Influence on Phoneme PerceptionLexical Influence on Phoneme Perception
nonword-word: dask-taskword-nonword: dash-tash
% id
enti
fica
tion
as
/d/
short VOT (d) long VOT (t)
100
0
Top-down Examples 3Top-down Examples 3Phoneme Restoration EffectPhoneme Restoration Effect
Warren (1970) & Warren (1970) & Warren & Warren (1970)Warren & Warren (1970):: ““The state governors met with their respective The state governors met with their respective
legilegiSSlatures convening in the capital city”latures convening in the capital city”
– SS replaced with cough or noise and played to listeners replaced with cough or noise and played to listeners
– Then asked listener to figure out where the sound was Then asked listener to figure out where the sound was replaced.replaced.
– What happened?What happened?
Top-down Examples 3Top-down Examples 3Phoneme Restoration EffectPhoneme Restoration Effect
Warren (1970) & Warren (1970) & Warren & Warren (1970)Warren & Warren (1970)::
It was found that the *eel was on the orange.
It was found that the *eel was on the axle.
It was found that the *eel was on the fishing-rod.
It was found that the *eel was on the table.
http://www.asj.gr.jp/2006/data/kashi/index.htmlhttp://www.acsu.buffalo.edu/~bmb/Courses/Old-Courses/PSY341-Fa2003/Exercises/Phon-rest/phon-rest.html
It was found that the *eel was on the shoe.
Gating TaskGating Task(Grosjean 1980)(Grosjean 1980)
Cumulative fragment of speech played.Cumulative fragment of speech played. Measure how much from the onset of Measure how much from the onset of
word participants need to hear before word participants need to hear before identifying it.identifying it.– RECOGNITION POINT = earliest “gate” at RECOGNITION POINT = earliest “gate” at
which the participant picks the correct which the participant picks the correct response and maintains it for the rest of the response and maintains it for the rest of the trials.trials.
50 ms 100 ms 150 ms 200 ms 250 ms 300 ms 367 ms
Top-down Examaple 4Top-down Examaple 4Gating Task Gating Task (Grosjean 1980)(Grosjean 1980)
Compare word in isolation and in context.Compare word in isolation and in context. In isolation: “In isolation: “camel”camel” In context: “The kids went to the zoo and In context: “The kids went to the zoo and
rode on the rode on the camel”camel”
– Recognition Point:Recognition Point: In Isolation ~333 In Isolation ~333 msms
In context ~199 In context ~199 msms
50 ms 100 ms 150 ms 200 ms 250 ms 300 ms 367 ms
50 ms 100 ms 150 ms 200 ms 250 ms 300 ms
Isolation
Context
Top-down Example 5Top-down Example 5Word MonitoringWord Monitoring ((Marslen-Wilson, Brown, & Tyler, 1988) Listening to sentences & Listening to sentences &
monitoring for specific wordsmonitoring for specific words– Word in isolation: ~300 msWord in isolation: ~300 ms– Normal: The boy held the Normal: The boy held the guitarguitar. ~ 240 ms.. ~ 240 ms.
– Discourse Incongruence: ~235 ms.Discourse Incongruence: ~235 ms.– Pragmatic Anomalous: The boy buried the Pragmatic Anomalous: The boy buried the guitarguitar. ~ . ~
268 ms268 ms– Semantic Anomalous: The boy drank the Semantic Anomalous: The boy drank the guitarguitar. .
~291 ms~291 ms– Categorical Anomalous: The boy slept the Categorical Anomalous: The boy slept the guitarguitar. .
~320 ms~320 ms
Speech RecognitionSpeech Recognition How do we achieve speed?How do we achieve speed?
– Parallel searchParallel search I.e. Activation of potential candidates in I.e. Activation of potential candidates in
parallelparallel
– Consult contextual informationConsult contextual information Use of contextual information to select or Use of contextual information to select or
weed out candidates!weed out candidates!
Models that consider Models that consider contextual informationcontextual information Examine 2 influential models of Examine 2 influential models of
speech processingspeech processing(evolved from Forster & Morton’s)(evolved from Forster & Morton’s)– Cohort ModelCohort Model– TRACE ModelTRACE Model
Currently other existing models in Currently other existing models in the literature.the literature.
SubtextSubtext How might psychology experimentsHow might psychology experiments
– inform us of our mental processes inform us of our mental processes – help us create models of our mental help us create models of our mental
representations and of how our mind representations and of how our mind process information?process information?
– be designed to help us distinguish be designed to help us distinguish between models or help us revise an between models or help us revise an existing one?existing one?
SubtextSubtext
In evaluating any model, consider:In evaluating any model, consider:– How well does the model account for How well does the model account for
existing experimental findings?existing experimental findings?– Is the representation depicted in the Is the representation depicted in the
model an intuitively plausible one? model an intuitively plausible one? – Does the model make predictions that Does the model make predictions that
are not in fact borne out by available are not in fact borne out by available empirical (i.e. observational and/or empirical (i.e. observational and/or experimental) evidence?experimental) evidence?
INTEGRATION STAGE(in which the semantic and syntactic
properties of the chosen words are utilized)
SELECTION STAGE(the most likely candidate is chosen from
cohort)
ACCESS STAGE(perceptual representation used to activate lexical items, thus generating a candidate
set of items – the cohort)
Cohort ModelCohort ModelMarslen-Wilson and Welsh (1978)Marslen-Wilson and Welsh (1978)
Input
Cohort Model – Access Cohort Model – Access StageStageMarslen-Wilson and Welsh (1978)Marslen-Wilson and Welsh (1978)
SS
songsong
storystory
sparrowsparrow
sauntersaunter
slowslow
secretsecret
sentrysentry......
(i.e., words beginning w/ the sound heard so far)
Cohort Model – Access Cohort Model – Access StageStageMarslen-Wilson and Welsh (1978)Marslen-Wilson and Welsh (1978)
SPSP
spicespice
spokespoke
sparespare
spinspin
splendidsplendid
spellingspelling
spreadspread
(candidates that no longer fit the incoming stream, are eliminated)
...
Cohort Model – Access Cohort Model – Access StageStageMarslen-Wilson and Welsh (1978)Marslen-Wilson and Welsh (1978)
SPISPI
spitspit
spigotspigot
spillspill
spiffyspiffy
spinakerspinaker
spiritspirit
spinspin...
Cohort Model – Access Cohort Model – Access StageStageMarslen-Wilson and Welsh (1978)Marslen-Wilson and Welsh (1978)
SPINSPIN
spinspin
spinachspinach
spinsterspinster
spinakerspinaker
spindlespindle
Cohort Model – Access Cohort Model – Access StageStageMarslen-Wilson and Welsh (1978)Marslen-Wilson and Welsh (1978)
SPINASPINA spinachspinach
Cohort Model – Access Cohort Model – Access StageStageMarslen-Wilson and Welsh (1978)Marslen-Wilson and Welsh (1978)
SPINASPINA spinachspinach
word uniqueness point
•Note: Some words have no uniqueness point (e.g., “spin”)
Cohort Model – Access Cohort Model – Access StageStageMarslen-Wilson and Welsh (1978)Marslen-Wilson and Welsh (1978)
Uniqueness pointUniqueness point Recognition pointRecognition point Highly Correlated.Highly Correlated.
Support idea of cohort.Support idea of cohort.
Cohort ModelCohort Model
Auditory Lexical Auditory Lexical Decision.Decision.
Uniqueness Uniqueness point + 450 ms point + 450 ms constant for constant for responding responding “NO, It’s not a “NO, It’s not a word.”word.”
INTEGRATION STAGE(in which the semantic and syntactic
properties of the chosen words are utilized)
SELECTION STAGE(the most likely candidate is chosen from
cohort)
ACCESS STAGE(perceptual representation used to activate lexical items, thus generating a candidate
set of items – the cohort)
Cohort ModelCohort ModelMarslen-Wilson and Welsh (1978)Marslen-Wilson and Welsh (1978)
Input
Cohort Model – Access Cohort Model – Access StageStageMarslen-Wilson and Welsh (1978)Marslen-Wilson and Welsh (1978) Selection stage: Making use of Selection stage: Making use of
contextual effectscontextual effects to achieve speed.Contexts: Contexts: – All the information not in the immediate All the information not in the immediate
sensory signal. sensory signal. – E.g., Information from previous sensory E.g., Information from previous sensory
input (prior context) to higher knowledge input (prior context) to higher knowledge sources (e.g., lexical, syntactic, semantic, sources (e.g., lexical, syntactic, semantic, and pragmatic info).and pragmatic info).
One big Q:One big Q:– Which contextual effects are helpful?Which contextual effects are helpful?
Cohort Model – Access Cohort Model – Access StageStageMarslen-Wilson and Welsh (1978)Marslen-Wilson and Welsh (1978)
Another BIG Q : Another BIG Q : When do/can we consider contextual do/can we consider contextual
information?information?– Generation vs. SelectionGeneration vs. Selection
Proposal vs. DisposalProposal vs. Disposal
– Pre-lexical or Post-lexicalPre-lexical or Post-lexical
How do we address the when Q How do we address the when Q experimentally?experimentally?
Zwitserlood (1989)Zwitserlood (1989)
Crazy complicated classic experiment.Crazy complicated classic experiment. Involves 3 separate groups of participantsInvolves 3 separate groups of participants
– Involves Sentence Completion Task.Involves Sentence Completion Task. Determines the Strength of Contextual InformationDetermines the Strength of Contextual Information
– Involves Gating Task.Involves Gating Task. Determines Probe Positions on the PRIME word.Determines Probe Positions on the PRIME word.
– Involves Cross-Modal Priming.Involves Cross-Modal Priming. Determines whether CAPITAIN primes BOAT and Determines whether CAPITAIN primes BOAT and
MONEY (semantically related to CAPITAL) at various MONEY (semantically related to CAPITAL) at various probe positions (i.e. points in time).probe positions (i.e. points in time).
KAPITEIN
BOOT GELD
KAPITEIN KAPITAAL
Cross-Modal PrimingCross-Modal Priming
or
Hear Prime:
Lexical Decision:
“BOAT” “MONEY”
Varying position of when to do lexical decision
What is the strength of the What is the strength of the context? (sentence context? (sentence completion)completion)What’s a good continuation for:What’s a good continuation for: They mourned the loss of their _______.They mourned the loss of their _______. With dampened spirits the men stood With dampened spirits the men stood
around the grave. They mourned the around the grave. They mourned the loss of their _______.loss of their _______.
Classify Responses of Participants into:Classify Responses of Participants into:– Biasing contexts: Biasing contexts:
16%-33% said the prime word and 0% said prime 16%-33% said the prime word and 0% said prime competitor.competitor.
– Neutral contexts: Neutral contexts: 0% said prime word, and 0% said prime 0% said prime word, and 0% said prime
competitor.competitor.
Where to Probe for Where to Probe for Activation? (Gating Task)Activation? (Gating Task)
Isolation Point: 1Isolation Point: 1stst time 50% of the participants time 50% of the participants pick the correct word and sticks with it to the end. pick the correct word and sticks with it to the end.
PROBE POSITIONSPROBE POSITIONS Position 0: Onset of wordPosition 0: Onset of word Position 1: Isolation Point with Biasing ContextPosition 1: Isolation Point with Biasing Context
– (ave. 130 ms after onset)(ave. 130 ms after onset) Position 2: Isolation Point with Neutral ContextPosition 2: Isolation Point with Neutral Context
– (ave. 199 ms after onset)(ave. 199 ms after onset) Position 3: Isolation Point in Carrier PhrasePosition 3: Isolation Point in Carrier Phrase
– The next word is ____. (ave. 278 ms after onset)The next word is ____. (ave. 278 ms after onset) Position 4: Recognition Point w/ Carrier PhrasePosition 4: Recognition Point w/ Carrier Phrase
– (ave. 410 ms after onset)(ave. 410 ms after onset)
WhenWhen does context play does context play a role? a role? (Four Possible (Four Possible Outcomes)Outcomes)
Before word spoken
During lexical access
During selection phase
At post-lexicalintegration stage
TASKTASKHear: CAPTAIN
Lexical Decision: BOAT or MONEY
BOAT – solid lineMONEY – dashed line
GRAPH LEGENDGRAPH LEGEND
Context plays a roleContext plays a roleBEFORE word spokenBEFORE word spoken
C A P T A I N
BOAT
MONEY
Context plays a role Context plays a role DURING lexical accessDURING lexical access
BOAT
MONEY
C A P T A I N
Context plays a role Context plays a role DURING selection DURING selection phasephase
BOAT
MONEY
C A P T A I N
Context plays a roleContext plays a roleAT POST-LEXICAL AT POST-LEXICAL integration integration
BOAT
MONEY
C A P T A I N
Comparing Data to Comparing Data to PredictionsPredictions Zwitserlood’s prediction slides Zwitserlood’s prediction slides
plots level of activation vs. time.plots level of activation vs. time. Her data is in terms of reaction Her data is in terms of reaction
time vs. probe positions (~time).time vs. probe positions (~time). How do we compare the two?How do we compare the two?
– Assumption: Faster reaction = Assumption: Faster reaction = higher level of activationhigher level of activation
ResultsResultsR
eact
ion
Tim
e (
ms)
C A P T A I N
MONEY
BOAT
INTEGRATION STAGE(in which the semantic and syntactic
properties of the chosen words are utilized)
SELECTION STAGE(the most likely candidate is chosen from
cohort)
ACCESS STAGE(perceptual representation used to activate lexical items, thus generating a candidate
set of items; the cohort)
Cohort ModelCohort ModelMarslen-Wilson and Welsh (1978)Marslen-Wilson and Welsh (1978)
Autonomous
Interactive
Interactive
Input
Some TerminologiesSome Terminologies Serial Serial vs.vs. Parallel Parallel Bottom-upBottom-up vs. vs. Top-downTop-down AutonomousAutonomous vs. vs. InteractiveInteractive
– AutonomousAutonomous: stage of processing : stage of processing proceeds independently of information proceeds independently of information from other processing modulesfrom other processing modules
– InteractiveInteractive: stage of processing quickly : stage of processing quickly considers information from other considers information from other processing modules as info comes inprocessing modules as info comes in
IncrementalIncremental: structuring and : structuring and interpreting information as it comes ininterpreting information as it comes in
Terminologies
Problem for Cohort Problem for Cohort ModelModel If you set up the wrong cohort, If you set up the wrong cohort,
how do you recover?how do you recover?– e.g. dragedy for tragedye.g. dragedy for tragedy– Misalignment problemMisalignment problem
The sky is falling!
This guy is falling!or
ThesKyisfalling!
Revised Cohort ModelRevised Cohort Model(Marslen-Wilson (1987)(Marslen-Wilson (1987)
Still set up an initial cohort of candidates. Still set up an initial cohort of candidates. Elimination process is no longer all-or Elimination process is no longer all-or
nothing. Items that do not receive further nothing. Items that do not receive further positive information decay in activation positive information decay in activation rather than being eliminated rather than being eliminated – Allows backtracking for misheard/distorted Allows backtracking for misheard/distorted
wordswords– Context loses some of its power, as it cannot be Context loses some of its power, as it cannot be
used to influence the items that form the initial used to influence the items that form the initial cohort.cohort.
A recognized word has a higher relative A recognized word has a higher relative activation than other words in the cohort. activation than other words in the cohort.
TRACE ModelTRACE Model(McClelland, Elman, (McClelland, Elman, Rumelhart’86)Rumelhart’86)
Model used for other things…Model used for other things…
Connectionist ModelsConnectionist Models
http://www.cheshireeng.com/Neuralyst/nnbg.htm
A NEURONNETWORK OF NEURONS
Connections can be either inhibitory or excitatory.
Digression: Connectionist Networks
Properties of Properties of Connectionist UnitConnectionist Unit
Activation Level = w1*A1 + w2*A2 + ...... + w8*A8
where -1 wn +1
A1 A2 .. .. .. A8
w1
w2
w8
Activation Level
Output Activation
Digression: Connectionist Networks
Squashing/Threshold Squashing/Threshold FunctionFunction
If Activation Level < 0.5Output = 0
If Activation Level 0.5Output = 1
A1 A2 .. .. .. A8
w1
w2
w8
Activation Level
Output Activation
Digression: Connectionist Networks
Network of Network of Connectionist UnitsConnectionist Units
Digression: Connectionist Networks
McClelland (1981)McClelland (1981)
Art
Lance
Ralph
Rick Sam20s
30s
40s
Jet
Shark Sing. Marr. Div.Pusher
Burglar
Bookie
Digression: Connectionist Networks
Inhibitory ConnectionsInhibitory Connections
Art
Lance
Ralph
Rick Sam20s
30s
40s
Jet
Shark Sing. Marr. Div.Pusher
Burglar
Bookie
Digression: Connectionist Networks
Who’s Art?Who’s Art?
Art
Lance
Ralph
Rick Sam20s
30s
40s
Jet
Shark Sing. Marr. Div.Pusher
Burglar
Bookie
Digression: Connectionist Networks
Who’s Art?Who’s Art?
Art
Lance
Ralph
Rick Sam20s
30s
40s
Jet
Shark Sing. Marr. Div.Pusher
Burglar
Bookie
Digression: Connectionist Networks
Content Content Addressability:Addressability:Who is Single and 30-something?Who is Single and 30-something?
Art
Lance
Ralph
Rick Sam20s
30s
40s
Jet
Shark Sing. Marr. Div.Pusher
Burglar
Bookie
Digression: Connectionist Networks
Content Content Addressability:Addressability:Who is Single and 30-something?Who is Single and 30-something?
Art
Lance
Ralph
Rick Sam20s
30s
40s
Jet
Shark Sing. Marr. Div.Pusher
Burglar
Bookie
Digression: Connectionist Networks
Who is Single and 30-Who is Single and 30-something?something?
Art
Lance
Ralph
Rick Sam20s
30s
40s
Jet
Shark Sing. Marr. Div.Pusher
Burglar
Bookie
Digression: Connectionist Networks
Training a Training a Connectionist ModelConnectionist Model All connection weights are initially set All connection weights are initially set
to to random numbersrandom numbers.. Input pattern is applied.Input pattern is applied. Model Produces output. (garbage)Model Produces output. (garbage) Output compared to “desired output”Output compared to “desired output” Connection weights adjusted slightly.Connection weights adjusted slightly. Repeat process with other inputs.Repeat process with other inputs.
==> Memory is in the weights.
Digression: Connectionist Networks
Simple Learning Rule for Simple Learning Rule for a Nodea Node
If Node is ON and is suppose to be OFF:If Node is ON and is suppose to be OFF:– turn turn downdown all connections from nodes passing all connections from nodes passing
activation to it. (w = w - 0.01).activation to it. (w = w - 0.01).
If Node is OFF and is suppose to be ON:If Node is OFF and is suppose to be ON:– turn turn upup all connections from nodes passing activation all connections from nodes passing activation
to it. (w = w + 0.01)to it. (w = w + 0.01)
Digression: Connectionist Networks
TRACE ModelTRACE ModelElman & McClellandElman & McClelland
(note: TRACE preconfigured. Not trained)
Features of the TRACE ModelFeatures of the TRACE Model (in comparison to OLD Cohort (in comparison to OLD Cohort Model)Model)
TRACE can “recover” if a given segment (even TRACE can “recover” if a given segment (even the first one) is missedthe first one) is missed– Does not rely heavily on knowing the left edge of the Does not rely heavily on knowing the left edge of the
wordword
TRACE’s bidirectional connections account for TRACE’s bidirectional connections account for phoneme restoration & other contextual effects phoneme restoration & other contextual effects on speech recognitionon speech recognition
TRACE predicts a lot of top-down information TRACE predicts a lot of top-down information flowflow– Potential problem: Weight given to contextual Potential problem: Weight given to contextual
information may be too strong?information may be too strong?
Cohort vs. Trace
Cohort vs. TRACE?Cohort vs. TRACE?
Do rhymes compete?Do rhymes compete?
Old Cohort Model: onset similarity is primary Old Cohort Model: onset similarity is primary because of the incremental (serial) nature of because of the incremental (serial) nature of speech speech – CatCat activates activates capcap, , castcast, , cattlecattle, , cameracamera, etc., etc.– Rhymes won’t competeRhymes won’t compete
TRACE: global similarity constrained by TRACE: global similarity constrained by incremental nature of speechincremental nature of speech– Cohorts and rhymes compete, but with different Cohorts and rhymes compete, but with different
time coursetime course
Cohort vs. Trace
Eye trackingEye camera
Scene camera
Allopenna, Magnuson & Allopenna, Magnuson & Tanenhaus (1998)Tanenhaus (1998)
“Pick up the beaker”
“Pick up the speaker” (RHYME COMPETITOR!)
Cohort vs. Trace
TRACE predictions match TRACE predictions match eye-tracking dataeye-tracking data
Adapted from Jim Magnuson, “Interaction in language processing: Pragmatic constraints on lexical access”
Cohort vs. Trace
Cohort vs. Trace?Cohort vs. Trace?
Is there lateral inhibition?Is there lateral inhibition? Old Cohort Model: units compete, but Old Cohort Model: units compete, but
don’t necessarily have inhibition built in.don’t necessarily have inhibition built in.
TRACE: within level, units compete and TRACE: within level, units compete and inhibit each other.inhibit each other.
jog job
Cohort vs. Trace
Marslen-Wilson & Warren Marslen-Wilson & Warren (1994)(1994)
jobjob job+ =
jobjog jo(g)b+ =
jobjod jo(d)b+ = (Nonword + Word)
(Word + Word)
(Word + Word)
FAST
MEDIUM
SLOW!!!
TRACE PredictionsTRACE Predictions
Auditory Lexical Decision on Auditory Lexical Decision on spliced & recombined sound spliced & recombined sound waves.waves.
jog jobjo(g)
Cohort vs. Trace
Marslen-Wilson & Warren Marslen-Wilson & Warren (1994)(1994)
Found Found jo(g)b & & jo(d)b equally slow, and equally slow, and slower than slower than job. . No lateral inhibition. No lateral inhibition.
jobjob job+ =
jobjog jo(g)b+ =
jobjod jo(d)b+ = (Nonword + Word)
(Word + Word)
(Word + Word)
FAST
MEDIUM
SLOW
TRACE PredictionsTRACE Predictions
Auditory Lexical Decision on Auditory Lexical Decision on spliced & recombined sound spliced & recombined sound waves.waves.
Cohort vs. Trace
Let’s try a more natural & sensitive measure!
nene(t)(t)tt
nene(k)(k)tt
nene(p)(p)tt
Pick up thePick up the
Dahan, Magnuson, Dahan, Magnuson, Tanenhaus & Hogan Tanenhaus & Hogan (2001)(2001)
netnet net+ =
netneck ne(k)t+ =
netnep ne(p)t+ =
Cohort vs. Trace
beagle
bead
beast
camera
beak
bellneck
net
ring
lobster
Prediction: Delayed target Prediction: Delayed target looks to the net withlooks to the net with
NE(k)T compared to compared to NE(p)T
PredictionsPredictions
Cohort vs. Trace
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
400 600 800 1000 1200 1400time since target onset (in ms)
N3W1
W2W1
W1W1
ne(p)t
ne(k)t Delayed look
netne(p)tne(k)t
ResultsResults
Fixati
on
Pro
port
ion
200
Cohort vs. Trace
Interim SummaryInterim Summary
Newer data are beginning to favor Newer data are beginning to favor the TRACE model over the cohort the TRACE model over the cohort model.model.
Cohort model proposes that Cohort model proposes that access stage is autonomous, but access stage is autonomous, but newer data suggests that there is newer data suggests that there is continuous sensitivity to continuous sensitivity to contextual information.contextual information.
Cohort vs. Trace