lexical access: generation & selection
DESCRIPTION
Lexical Access: Generation & Selection. Main Topic. Listeners as active participants in comprehension process Model system: word recognition. Outline. Speed & Robustness of Lexical Access Active Search Evidence for Stages of Lexical Access Autonomy & Interaction. Outline. - PowerPoint PPT PresentationTRANSCRIPT
Lexical Access:Generation & Selection
Main Topic
• Listeners as active participants in comprehension process
• Model system: word recognition
Outline
1. Speed & Robustness of Lexical Access2. Active Search3. Evidence for Stages of Lexical Access4. Autonomy & Interaction
Outline
1. Speed & Robustness of Lexical Access2. Active Search3. Evidence for Stages of Lexical Access4. Autonomy & Interaction
The mental lexicon
sport figure sing door carry
turf turtle gold turk turkey
turnwater turbo turquoise
turnip turmoil
How do we recognize words?
• The Simplest Theory
– Take a string of letters/phonemes/syllables, match to word in the mental lexicon
– (That’s roughly how word processors work)
• …is it plausible?
Word Recognition is Fast
• Intuitively immediate - words are recognized before end of word is reached
• Speech shadowing at very brief time-lags, ~250ms (Marslen-Wilson 1973, 1975)
• Eye-tracking studies indicate effects of access within 200-300ms
Lexical Access is Robust
• Succeeds in connected speech• Succeeds in fast speech• Survives masking effects of morphological
affixation and phonological processes• Deleted or substituted segments• Speech embedded in noise
But…
• Speed and robustness depends on words in context
sentence --> word context effects
• In isolation, word recognition is slower and a good deal more fragile, susceptible to error
• …but still does not require perfect matching
Questions
• How does lexical access proceed out of context?
• Why is lexical access fast and robust in context?
• When does context affect lexical access?
– does it affect early generation (lookup) processes?– does it affect later selection processes?
Classic Experimental Paradigms
Reaction Time Paradigms
• Lexical Decision• Priming
Looking for Words
• List 1sicklecathartictorridgregariousoxymoronatrophy
• List 2parabolaperiodontistpreternaturalpariahpersimmonporous
Looking for Words
• List 1sicklecathartictorridgregariousoxymoronatrophy
• List 2parabolaperiodontistpreternaturalpariahpersimmonporous
Speed of look-up reflects organization of dictionary
Looking for Words
+
Looking for Words
DASH
Looking for Words
+
Looking for Words
RASK
Looking for Words
+
Looking for Words
CURLY
Looking for Words
+
Looking for Words
PURCE
Looking for Words
+
Looking for Words
WINDOW
Looking for Words
+
Looking for Words
DULIP
Looking for Words
+
Looking for Words
LURID
(Embick et al., 2001)
Looking for Words
• Semantically Related Word Pairsdoctor nursehand fingerspeak talksound volumebook volume
Looking for Words
• In a lexical decision task, responses are faster when a word is preceded by a semantically related word
• DOCTOR primes NURSE• Implies semantic organization of dictionary
Outline
1. Speed & Robustness of Lexical Access2. Active Search3. Evidence for Stages of Lexical Access4. Autonomy & Interaction
Active Recognition
• System actively seeks matches to input - does not wait for complete match
This allows for speed, but …
Cost of Active Search…
• Many inappropriate words activated• Inappropriate choices must be rejected
• Two Stages of Lexical Accessactivation vs. competitionrecognition vs. selectionproposal vs. disposal
The mental lexicon
sport figure sing door carry
turf turtle gold turk turkey
turnwater turbo turquoise
turnip turmoil
The mental lexicon
sport figure sing door carry
turf turtle gold turk turkey
turnwater turbo turquoise
turnip turmoil TURN
Automatic activation
TURN
sport figure sing door carry
turf turtle gold turk turkey
water turn turbo turquoise
turnip turmoil
Lateral inhibition
TURN
sport figure sing door carry
turf turtle gold turk turkey
water turn turbo turquoise
turnip turmoil
What is lexical access?
time
leve
l of a
ctiv
atio
n
resting level
TURN
Stimulus: TURN
TURNIP
TURFTURTLE
Activation Competition Selection/Recognition
(e.g. Luce et al. 1990, Norris 1994)
Cohort
S
songstorysparrowsaunterslowsecretsentryetc.
Cohort
SP
spicespokesparespinsplendidspellingspreadetc.
Cohort
SPI
spitspigotspillspiffyspinakerspiritspinetc.
Cohort
SPIN
spinspinachspinsterspinakerspindle
Cohort
SPINA spinach
Cohort
SPINA spinach
word uniqueness point
Cohort
SPINAspinachspinetspineret
Cross-Modal Priming
Evidence for Cohort Activation
KAPITEIN KAPITAAL
(Marslen-Wilson, Zwitserlood)
Evidence for Cohort Activation
KAPITEIN KAPITAAL
KAPIT…
(Marslen-Wilson, Zwitserlood)
Evidence for Cohort Activation
KAPITEIN KAPITAAL
KAPIT…
BOOT
GELD
(Marslen-Wilson, Zwitserlood)
Evidence for Cohort Activation
KAPITEIN KAPITAAL
KAPIT…
BOOT
GELD
(Marslen-Wilson, Zwitserlood)
Evidence for Cohort Activation
KAPITEIN KAPITAAL
KAPIT…
BOOT
GELD
KAPITEIN
BOOT
GELD
(Marslen-Wilson, Zwitserlood)
Evidence for Cohort Activation
CAPTAIN CAPTIVE
CAPT…
SHIP
GUARD
CAPTAIN
SHIP
GUARD
(Marslen-Wilson, Zwitserlood)
Cohort Model
• Partial words display priming properties of multiple completions: motivates multiple, continuous access
• Marslen-Wilson’s claims
– Activation of candidates is autonomous, based on cohort only– Selection is non-autonomous, can use contextual info.
• How, then, to capture facilitatory effect of context?
Gating Measures
• Presentation of successive parts of words
– S– SP– SPI– SPIN– SPINA…
• Average recognition times
– Out of context: 300-350ms– In context: 200ms
(Grosjean 1980, etc.)
Word Monitoring
• Listening to sentences - monitoring for specific words
– Mean RT ~240ms– Identification estimate ~200ms
• Listening to same words in isolation
– Identification estimate ~300ms
(Brown, Marslen-Wilson, & Tyler)
Cross-Modal Priming
The guests drank vodka, sherry and port at the reception
(Swinney 1979, Seidenberg et al. 1979)
Cross-Modal Priming
The guests drank vodka, sherry and port at the reception
WINE
SHIP
(Swinney 1979, Seidenberg et al. 1979)
Cross-Modal Priming
The guests drank vodka, sherry and port at the reception
WINE
SHIP
(Swinney 1979, Seidenberg et al. 1979)
Cross-Modal Priming
The guests drank vodka, sherry and port at the reception
WINE
SHIP
(Swinney 1979, Seidenberg et al. 1979)
Cross-Modal Priming
The guests drank vodka, sherry and port at the reception
WINE
SHIP
(Swinney 1979, Seidenberg et al. 1979)
Generation and Selection• Investigating the dependence on ‘bottom-up’ information in language
understanding
• ‘Active’ comprehension has benefits and costs
– Speed– Errors– Overgeneration entails selection
• Sources of information for generating candidates
– Bottom-up information (e.g., lexical cohorts)– ‘Top-down’ information (e.g., sentential context)– Questions about whether context aids generation or selection
Cross-modal Priming
• Early: multiple access
• Late: single access
…i.e., delayed effect of context
CMLP - Qualifications
• Multiple access observed– when both meanings have roughly even frequency– when context favors the lower frequency meaning
• Selective access observed– when strongly dominant meaning is favored by context
(see Simspon 1994 for review)
Why multiple/selective access?
• How could context prevent a non-supported meaning from being accessed at all?
(Note: this is different from the question of how the unsupported meaning is suppressed once activated)
• Possible answer: selective access can only occur in situations where context is so strong that it pre-activates the target word/meaning
Tanenhaus & Lucas 1987
Cross-Modal Lexical Access• Seidenberg, Tanenhaus, Leiman, & Bienkowski (1982)
– Cross-modal naming– They all rose vs. They bought a rose Probes: FLOWER, STOOD
– Immediate presentation: equal priming; 200ms delay: selective priming
• Prather & Swinney (1977): similar w/ cross-modal lexical decision• Tanenhaus & Donnenworth-Nolan (1984): similar, w/ extra delay in
presenting target word
Experiment 1
Experiment 1
Experiment 2
cost no cost
Refining the Story
• Frequency in context– eye-tracking in reading– eye-tracking and object recognition
• Electrophysiological measures of multiple access• When can context affect generation?
– strongly supporting contexts– ERP evidence
Evidence for Cohort Activation
CAPTAIN CAPTIVE
CAPT…
SHIP
GUARD
CAPTAIN
SHIP
GUARD
(Marslen-Wilson, Zwitserlood)
Additional Context Effects
• Word context affects phoneme identification…
word --> phoneme context effects
Phoneme Restoration
• The _eel had a broken axleThe _eel on the orange was hard to cut(Warren 1970)
• Phoneme restoration effects are stronger(i) in words than non-words(ii) later in words(iii) in strongly biasing contexts
Phoneme Monitoring
• press the button as soon as you hear a ‘b’
• “in the yard was a large group of twittering birds”
“cat, dog, horse, rabbit …”
• Monitoring is facilitated by context
Perceptual Boundaries
Perceptual Boundaries
DA TA
Perceptual Boundaries
DASK TASK
(Ganong 1980)
Perceptual Boundaries
DASK TASK
(Ganong 1980)
Perceptual Boundaries
DASH TASH
(Ganong 1980)
Perceptual Boundaries
DASH TASH
(Ganong 1980)
B I G A T R
BIG BAT DOG Words
Phonemes
Feedback vs. Decision Bias
Cohort Model• Partial words display priming properties of multiple completions:
motivates multiple, continuous access
• Marslen-Wilson’s claims
– Activation of candidates is autonomous, based on cohort only– Selection is non-autonomous, can use contextual info.
• How, then, to capture facilitatory effect of context…
Cohort
SPINA spinach
Cohort
SPIN
spinspinachspinsterspinakerspindle
Speed of Integration• If context can only be used to choose among candidates generated by
cohort…
– context can choose among candidates prior to uniqueness point
– but selection must be really quick, in order to confer an advantage over bottom-up information
– [… or recognition following uniqueness point must be slow in the absence of context.]
Frequency in Reading
• Rayner & Frazier (1989): Eye-tracking in reading
– measuring fixation durations in fluent reading– ambiguous words read more slowly than unambiguous, when
frequencies are balanced, and context is unbiased– unbalanced words: reading profile like unambiguous words
– when prior context biases one meaning• dominant-biased: no slowdown due to ambiguity• subordinate-biased: slowdown due to ambiguity• contextual bias can offset the effect of frequency bias
– how can context boost the accessibility of a subordinate meaning?
Frequency in Object Recognition
X
“Pick up the be..” (Dahan, Magnuson, & Tanenhaus, 2001)
Frequency in Object Recognition
X
bench
bed
bell
lobster
“Pick up the be..” (Dahan, Magnuson, & Tanenhaus, 2001)
Frequency in Object Recognition
• Timing estimates
– Saccadic eye-movements take 150-180ms to program– Word recognition times estimated as eye-movement times minus
~200ms
Frequency in Object Recognition
(Dahan, Magnuson, & Tanenhaus, 2001)
Frequency in Object Recognition
(Dahan, Magnuson, & Tanenhaus, 2001)
Frequency in Object Recognition
(Dahan, Magnuson, & Tanenhaus, 2001)
Evidence for Cohort Activation
CAPTAIN CAPTIVE
CAPT…
SHIP
GUARD
CAPTAIN
SHIP
GUARD
(Marslen-Wilson, Zwitserlood)
Matches to other parts of words
• Word-ending matches don’t prime
– honing [honey] bij [bee]woning [apartment]foning [--]
Disagreements
– Continuous activation, not limited to cohort, as in TRACE model (McClelland & Elman, 1986)
– Predicts activation of non-cohort members, e.g. shigarette, bleasant
Non-Cohort Competitors
(Allopenna, Magnuson, & Tanenhaus, 1998)
“Pick up the…”
beaker
beetle (onset)speaker (non-onset)carriage (distractor)
Non-Cohort Competitors
(Allopenna, Magnuson, & Tanenhaus, 1998)
“Pick up the…”
beaker
beetle (onset)speaker (non-onset)carriage (distractor)
Non-Cohort Competitors
(Allopenna, Magnuson, & Tanenhaus, 1998)
“Pick up the…”
beaker
beetle (onset)speaker (non-onset)carriage (distractor)
Overview• Reasons to be ‘active’
– Speed• Using highly reliable information (e.g., eleph…; NP-nom…)• Using partially reliable information (e.g., ele…; NP V…; wh NP V…)
– Robustness – Minimize storage
• Reasons to be ‘informationally encapsulated’ (‘modular’)– Architectural constraints– Information availability
• Capturing effects of context– Bottom-up word information proposes candidates for evaluation against context– Bottom-up word information yields activation based on frequency, phonotactic
probability, etc. (assuming multiple access)– Contextual information may prime certain lexical/semantic features, leading to
earlier activation/selection of some words; context can also prime morpho-syntactic features, leading to exclusion of some word candidates
• Context vs. frequency
– The guests drank wine, sherry, and port at the reception.
– The violent hurricane did not damage the ships which were in the port, one of the best equipped along the coast.
Outline
1. Speed & Robustness of Lexical Access2. Active Search3. Evidence for Stages of Lexical Access4. Autonomy & Interaction
M350
(based on research by Alec Marantz, Liina Pylkkänen, Martin Hackl & others)
Lexical access involves
1. Activation of lexical representations• including activation of representations
matching the input, and• lateral inhibition between activated
representations
2. Followed by selection or decision• involving competition among activated
representations that are similar in form
RESPONSE TO A VISUAL WORD Sagittal view
A P
M350
M350
0 200 300 400 Time [msec]
MEG response components elicited by visually presented words in the lexical decision task
RMS analysis of component field patterns.
(Embick et al., 2001)
Neighbors & Competitors
• Phonotactic probability– sound combinations that are likely in English– e.g. ride vs. gush
• Neighborhood density– number of words with similar sounds– ride, bide, sighed, rile, raid, guide, died, tried,
hide, bride, rise, read, road, rhyme, etc.– gush, lush, rush, gut, gull …
RT
Behavioral evidence for dual effects• Same/different task (“low-level”)
RTs to nonwords with a high phonotactic probability are speeded up.
• Lexical decision task (“high-level”)RTs to nonwords with a high phonotactic probability are slowed down!
High probability: MIDE
YUSH RT
RT MIDE
YUSH RT
Low probability:
High probability:
Low probability:
Sublexicalfrequency effect
(Vitevich and Luce 1997,1999)
Competition effect
Stimuli
High probability Low probability
Word BELL, LINE PAGE, DISH
Nonword MIDE, PAKE JIZE, YUSH
• Materials of Vitevich and Luce 1999 converted into orthographic stimuli.
• Four categories of 70 stimuli:
• High and low density words frequency matched.
(Pylkkänen, Stringfellow, Marantz, Brain and Language, 2003)
Effect of probability/density (words)
100
200
300
400
500
600
700
M170 M250 M350 RT
HighProbWord LowProbWord
n.s.
n.s.
**
*
(Pylkkänen, Stringfellow, Marantz, Brain and Language, 2003)
Effect of probability/density (nonwords)
0
100
200
300
400
500
600
700
800
M170 M250 M350 RT
HighProbNonword LowProbNonword
n.s.n.s.
*
**
(Pylkkänen, Stringfellow, Marantz, Brain and Language, 2003)
M350 = 1st component sensitive to lexical factors but not affected by competition
time
leve
l of a
ctiv
atio
n
resting level
TURN
TURNIP
TURFTURTLE
Activation Competition Selection/RecognitionM350
Stimulus: TURN
I wanted to point out a minor difference in your interpretation of Allopenna, Magnuson, & Tanenhaus (1998) and mine. Allopenna et al. is cited on p. 75 as one of the "estimates in the literature [that] the earliest processes involved in lexical access often fall in the 200 ms range". But eye tracking data of the sort we presented actually gives a strikingly different estimate. What we find again and again in studies using the visual world paradigm is that there is an approximately 200-250 ms lag between events in the speech signal and changes in fixation proportions. However, this should not suggest that it takes 200 ms for processes of lexical access to kick in. Rather, given that it takes at least about 150 ms to plan and launch an eye movement to a point of light in a darkened room, this means we can roughly subtract 150 msecs of the lag and attribute it to saccade planning. This leaves us with only about 50 msecs to attribute to the very earliest processes of access that are indexed by the eye movements. (This seems too short by 1-2 dozen msecs, but note that only a very small proportion of trials include such early eye movements, and statistically reliable differences between related and unrelated items emerge another ~25-50 msecs later.)
[Email message, 6/26/07]
Jim Magnuson, UConn
Masked Priming
#######
brother
BROTH
Outline
1. Speed & Robustness of Lexical Access2. Active Search3. Evidence for Stages of Lexical Access4. Autonomy & Interaction
Autonomy
• “…a system [is] autonomous by being encapsulated, by not having access to facts that other systems know about” (Fodor 1983)
• “Autonomy would imply that processing operations at a given level proceed in the same way irrespective of whatever counsel might be deducible from the higher-level considerations” (Boland & Cutler)
Model Implied So Far
• Stage 1: activation based upon cohortsno effect of context at this stage
• Stage 2:selection affected by context
Boland & Cutler
• The debate over interaction/autonomy in lexical access focuses on the generation (activation) stage
• There is broad agreement that context affects lexical choices once multiple candidates have been generated
Cross-Modal Priming
The guests drank vodka, sherry and port at the reception
WINE
SHIP
(Swinney 1979, Seidenberg et al. 1979)
Cross-Modal Priming
The guests drank vodka, sherry and port at the reception
WINE
SHIP
(Swinney 1979, Seidenberg et al. 1979)
Cross-Modal Priming
• How could context prevent a contextually unsupported meaning from being accessed?
Cross-Modal Priming
• Conflicting results over effect of context on multiple access
• Tabossi (1998)
– The violent hurricane did not damage the ships which were in the port, one of the best equipped along the coast.
– Contexts are highly constraining, prime a specific feature of the target meaning.
Active Comprehension
• Distinction between activation and selection applies equally to syntactic comprehension
• Is active comprehension a fully general property of language understanding?
N400
Negative polarity peaking at around 400 ms central scalp distribution
(Kutas & Federmeier 2000)
(Kutas & Federmaier 2000)
‘baseball’ is not at all plausible here, yet it elicits a smaller N400 - why?
Input to left hem. visual system must have privileged access to information about predictions.
Implications
• If Kutas & Federmeier’s results are robust, this implies that
– lexical priming can cause apparentearly context effects
– this implies ‘very active search’– hemispheres are not alike in this regard
+
walk
+
walk
Overview• Reasons to be ‘active’
– Speed• Using highly reliable information (e.g., eleph…; NP-nom…)• Using partially reliable information (e.g., ele…; NP V…; wh NP V…)
– Robustness – Minimize storage
• Reasons to be ‘informationally encapsulated’ (‘modular’)– Architectural constraints– Information availability
• Capturing effects of context– Bottom-up word information proposes candidates for evaluation against context– Bottom-up word information yields activation based on frequency, phonotactic probability, etc.
(assuming multiple access)– Contextual information may prime certain lexical/semantic features, leading to earlier
activation/selection of some words; context can also prime morpho-syntactic features, leading to exclusion of some word candidates
• Hemispheric contrasts - does differential processing reflect deep differences?
Conclusion…
• Word recognition is fast and robust because of use of context
• Speed/robustness is achieved by– active generation of candidates from incomplete input– selection among candidates, based upon context
• Activation ˜ autonomousSelection ˜interactive
Next…
• Syntax
– most issues seen here also apply to syntactic processes– generation stage is much more complex, since syntactic
processing is more than just a lookup/activation process.