MEG Studies of Lexical Access

1
MEG Studies of Lexical Access

• Liina Pylkkänen
• MIT Mind Articulation Project

KIT/MIT MEG LAB
2
Two basic processes of linguistic computation

• Accessing lexical items from the mental lexicon.

• gakusei
• kashikoi
• sono

sono gakusei-ga kashikoi

• Putting them together.

3
What is the timing and location of lexical access
in the brain?

• What are the cognitive processes involved in
  accessing a word?

Activation and Competition/Inhibition
4
The mental lexicon
sport figure sing door
carry turf turtle gold turk
turkey turn water turbo
turquoise turnip turmoil
5
The mental lexicon
sport figure sing door
carry turf turtle gold turk
turkey turn water turbo
turquoise turnip turmoil
TURN
6
1. Automatic activation
sport figure sing door
carry turf turtle gold turk
turkey water turn turbo
turquoise turnip turmoil
TURN
7
2. Competition/inhibition
sport figure sing door
carry turf turtle gold
turk turkey water turn
turbo turquoise turnip turmoil
TURN
8
3. Task-related processes

• Present in the processing of any stimulus that
  the subject is performing some kind of a task on.
  
• Post-lexical, attentional, nonlinguistic

9
The biggest question in lexical access research

• Is an effect lexical or post-lexical?

the theoretical debate consists of pushing
effects from the access column to the
post-access column and back again. It is quite
easy to find evidence that an effect may be
post-access, but very difficult to find evidence
that it is definitely not. - Kenneth I.
Forster -
10
Goal of this talk

• to use both behavioral reaction time (RT) data
  and MEG data to make an argument about the timing
  of lexical activation.

Do we really need the MEG data for making this
type of an argument?

• Could we just measure behavior in clever ways and
  get the same information?

11
Can we infer the timing of lexical access by
measuring reaction times (RT) only?
Example Semantic priming.
Real word or not?
RT (yes or no)
NURSE
DOCTOR
Time
RT
NURSE
DRIVER
Time
12
Can we infer the timing of lexical access by
measuring reaction times (RT) only?
Is the effect lexical or post-lexical? I.e.
automatic or conscious?
RT
NURSE
DOCTOR
Time
RT
NURSE
DRIVER
Time
13
If lexical ( automatic)
DOCTOR activates NURSE
RT
NURSE
DOCTOR
Time
RT
NURSE
DRIVER
Time
14
If lexical ( automatic)
NURSE is accessed faster due to residual
activation
RT
NURSE
DOCTOR
Time
RT
NURSE
DRIVER
Time
15
If post-lexical ( conscious)
NURSE is responded to faster since it fits the
preceding context (e.g. Neely 1991)
RT
NURSE
DOCTOR
Time
RT
NURSE
DRIVER
Time
16
If post-lexical, effect should dissappear if we
make the preceding context invisible to
conscious recognition
Masking
NURSE

DOCTOR

Time
NURSE

DRIVER

Time
17
If post-lexical, effect should dissappear if we
make the preceding context invisible to
conscious recognition
Effect remains, i.e. is automatic (e.g. Deacon
et al 2000).
RT
NURSE

DOCTOR

Time
RT
NURSE

DRIVER

Time
18
NURSE is accessed faster because DOCTOR already
activated it
activation
activation
RT
NURSE
DOCTOR
nurse
nurse
Time
activation
RT
NURSE
DRIVER
nurse
Time
19
NURSE is accessed faster because DOCTOR already
activated it
When does lexical access occur?
activation
activation
RT
NURSE
DOCTOR
nurse
nurse
Time
activation
RT
NURSE
DRIVER
nurse
Time
20
When does the activation of NURSE occur?
How much can we shorten the interval between the
1st and the 2nd word until the effect dissappears?
activation
activation
RT
NURSE
DOCTOR
nurse
nurse
Time
activation
RT
NURSE
DRIVER
nurse
Time
21
When does the activation of NURSE occur?
How much can we shorten the interval between the
1st and the 2nd word until the effect dissappears?
activation
activation
RT
DOCTOR
NURSE
nurse
nurse
Time
activation
RT
NURSE
DRIVER
nurse
Time
22
When does the activation of NURSE occur?
How much can we shorten the interval between the
1st and the 2nd word until the effect dissappears?
activation
activation
RT
DOCTOR
NURSE
nurse
nurse
Time
activation
RT
NURSE
DRIVER
nurse
Time
23
When does the activation of NURSE occur?
How much can we shorten the interval between the
1st and the 2nd word until the effect dissappears?
200ms
activation
RT
DOCTOR
NURSE
nurse
Time
activation
RT
NURSE
DRIVER
nurse
Time
24
Conclusions (i) the effect on RTs is
lexical.(ii) it takes at least 200 ms for
DOCTOR to activate NURSE (by semantic
association).
200ms
activation
RT
DOCTOR
NURSE
nurse
Time
activation
RT
NURSE
DRIVER
nurse
Time
25
What we cant conclude

• that the activation of the semantic associate
  happens in some specific time window (the
  activation of NURSE by semantic association could
  happen after the onset of the target).
• anything about the activation time of the
  stimulus that the subject is performing the task
  on (except that its faster or slower than in
  some other condition).
• With MEG we can do both, and more...

activation
activation
RT
NURSE
DOCTOR
nurse
nurse
Time
26
MEG allows us to study the timing of activation
directly

• Lexical access the first component of the
  response to NURSE that occurs earlier in the
  related than in the unrelated condition.

RT
NURSE
DOCTOR
Time
RT
NURSE
DRIVER
Time
27
MEG allows us to study the timing of activation
directly

• Studying timing doesnt require priming.
• Experiment 1 (Embick, Hackl, et al) The first
  response component sensitive to stimulus
  frequency occurs at 350ms.

RT
Frequent
ASK
Time
RT
Infrequent
CLAM
Time
28
MEG allows the simultaneous study of multiple
levels of processing

• Experiment 3 350ms response is sensitive to
  stimulus properties affecting the speed of
  lexical access but not to stimulus properties
  affecting the post-access decision.

RT
Stimulus 1 - fast to access - slow to decide on
LINE
Time
RT
Stimulus 2 - slow to access - fast to decide on
PAGE
Time
29
An MEG Study of Word Frequency Effects in Lexical
Decision

• M. Hackl1, D. Embick1,2, J. Schaeffer3, M.
  Kelepir1, A. Marantz1,2
• 1 Dept. of Linguistics and Philosophy, MIT
• 2 JST/MIT Mind Articulation Project
• 3 Dept. of Linguistics, Ben-Gurion University of
  the Negev

30
The frequency effect

• Lexical decisions to frequent words faster than
  decisions to infrequent words.
• Account in activation-based models frequent
  words have a higher resting level.

31
Objective Identification of an MEG component
whose latency varies with the frequency of words,
to be used as an index in further studies of
lexical access and lexical organization.
Primary Result A component in the response to
words at 350ms, m350, varies in latency with the
frequency of words.
32
Stimuli

• Six bins of open-class words, arranged according
  to frequency Cobuild corpus, 320 million words

Category n/Million Log Freq. Example 64 700 2.
8 number 65 140 2.1 ask 66 30 1.4 wheel 67
6 .7 candle 68 1 0 clam 69 .2 -.7 snarl

• Two classes of non-words, pronounceable and
• non-pronounceable ratio of wordsnon-words 11.

33

• Task Lexical Decision.
• Subjects n 9 5F, 4M right-handed native
  speakers of English.
• Analysis Peaks identified based on RMS
  analysis.
• A subset of 17 left-hemisphere sensors were
  used for identification of peaks this set was
  held constant across subjects/conditions.

34
(No Transcript)
35
Three primary components
m170
36
(No Transcript)
37
Two Distinct Components

• Latency of m350 response varies by log
  frequency of words
• (p lt .0001)
• Latency of m250 response does not vary with log
  frequency
• (p .8)

38
Magnetic Field and Contour Map High Frequency
39
Magnetic Field and Contour Map Low Frequency
40
The M350

• Is the first MEG component serving as a predictor
  of the behavioral frequency effect.
• If the M350 indexes lexical access, it is also
  predicted to show priming effects.

41
A neural response sensitive to repetition

• L. Pylkkänen1,2, E, Flagg1,
• A. Stringfellow2, A. Marantz1,2
• 1 Dept. of Linguistics and Philosophy, MIT
• 2 JST/MIT Mind Articulation Project

42
The repetition priming effect

• Words are responded to more quickly on their
  second presentation than on their first.
• After a word has been accessed, its activation
  slowly returns to resting level if the word is
  presented again while there is still residual
  activation, access is facilitated.

43
Objective Identification of an MEG component
whose latency predicts the behavioral repetition
priming effect.
Result A component in the response to words and
pronouceable nonwords at 350ms, M350, occurs
earlier for repeated than for nonrepeated words.
44
Stimuli

• Four categories of 100 stimuli
• Repeated word DOG DOG

DOG
Prime, 500 ms
500 ms
DOG
Target, real word or not?

• Nonrepeated word DOG WIND
• Repeated nonword GULK GULK
• Nonrepeated nonword DOG GULK

45
Analysis

• Only correct trials were analyzed.
• RMS from a minimum of 17 left hemisphere sensors
  showing large responses between 150 and 450 ms.
• The latencies and amplitudes of major RMS peaks
  were recorded using latency and magnetic field
  distribution as criteria for determining whether
  a peak belonged to a certain category of
  responses.

46
Results 3 primary components
M170
M250
M350
RT
47
Effect of repetition on the M350 and RT


n.s
n.s
48
M350 positive signal maximum for repeated and for
nonrepeated words (single subject data)
49
The M350 is the first MEG component that is
sensitive to stimulus properties affecting
lexical access.
Does the M350 index lexical access or
the post-lexical decision process?
50
Separating lexical access from decision an MEG
study
KIT/MIT MEG LAB

• L. Pylkkänen1,2, A. Stringfellow1,2, M. Kelepir1,
  
• A. Marantz1,2
• 1Department of Linguistics and Philosophy,
• KIT-MIT MEG Laboratory,
• Massachusetts Institute of Technology
• 2Mind Articulation Project, International
  Cooperative Research Project, Japan Science and
  Technology Corporation

51
Goal
To study the M350 elicited by stimuli that make
lexical activation fast but postlexical processes
slow. With such stimuli, does the M350 occur
early or late?
RT
Stimulus 1 - fast to access - slow to decide on
Stim1
Time
RT
Stimulus 2 - slow to access - fast to decide on
Stim2
Time
52
Phonotactic probability

• how common the sounds and sequences of the word
  are.

• High probability down, piece, line
• Low probability knife, weight, page

Phonological neighborhood density

• how many similar sounding words a stimulus has in
  the language

• High density down, piece, line
• Low density knife, weight, page

53
High phonotactic probability speeds up activation
while
high neighborhood density slows down decision.
54
Evidence from behavioral measures(Vitevich and
Luce 1997,1999)

• Low-level task Same or different from previous
  stimulus?
• RTs to nonwords with a high phonotactic
  probability are speeded up.

RT
High probability
MIDE
Sublexical frequency effect
RT
YUSH
Low probability

• High-level task Real word? (Lexical decision)
• RTs to nonwords with a high phonotactic
  probability are slowed down!

Why?
RT
High probability
MIDE
RT
Low probability
YUSH
55
High probability stimuli stimuli from high
density neighborhoods

• Stimuli from high density neighborhoods have many
  competitors which slows down decision.

mile mild might migrate mike mime mine mire mind
mite migraine micro
neighborhood activated
RT
MIDE
High probability
neighborhood activated
yuppie yucca yuck yum
RT
Low probability
YUSH
56
In a lexical decision task

• A high probability stimulus is activated fast
• because of a sublexical frequency effect
• but
• responded to slowly
• because of a neighborhood effect.

57
Hypothesis
If the M350 reflects lexical activation,

• a high probability stimulus should elicit a fast
  M350
• because of a sublexical frequency effect
• but
• a slow RT
• because of a neighborhood effect.

58
Stimuli

• Materials of Vitevich and Luce 1999 converted
  into orthographic stimuli.
• Four categories of 70 stimuli

• High and low density words frequency matched.

59

• Task Lexical decision.
• Subjects 9 right-handed (2 F, 7 M),
  English-speaking adults with normal or
  corrected-to-normal vision
• MEG recording

60
Results 3 primary peaks
M170
M250
M350
RT
61
Effect of probability/density (words)


n.s.
n.s.
62
Effect of probability/density (nonwords)


n.s.
n.s.
63
Effect of phonotactic probability on the M350
(positive and negative signal maxima for one
subject)
M350 positive maximum
M350
M350 negative maximum
M350
64
High probability stimuli speed up the M350
because they speed up activation.

mile mild might migrate migraine mike mime
mine mire mind mite micro
High Prob. - fast to access - slow to decide on
MIDE
RT
M350
Time
yuppie yucca yuck yum
Low Prob - slow to access - fast to decide on
YUSH
RT
M350
Time
65
Neighborhood density affects decision latencies
but not M350 latencies.

mile mild might migrate migraine mike mime
mine mire mind mite micro
High Prob. - fast to access - slow to decide on
MIDE
RT
M350
Time
yuppie yucca yuck yum
Low Prob - slow to access - fast to decide on
YUSH
RT
M350
Time
66
The M350 cannot reflect the post-access decision
process.

mile mild might migrate migraine mike mime
mine mire mind mite micro
High Prob. - fast to access - slow to decide on
MIDE
RT
M350
Time
yuppie yucca yuck yum
Low Prob - slow to access - fast to decide on
YUSH
RT
M350
Time
67
Location of the M350 relative to the auditory M100
M350
M100
10 cm
10 cm
M350 locations for the four experimental
conditions for one subject displayed in relation
to the location of the auditory cortex (response
to 1 kHz tone).
68
Conclusion

• The M350
• is the first component sensitive to stimulus
  properties affecting the speed of lexical access.
• independent of the post-access decision process
  of lexical decision

69
Conclusion

• Therefore, the M350
• has the properties we would expect from a
  linguistic component that reflects the automatic
  accessing of the primitive units that enter the
  syntax.
• promises to be an important tool for
  investigating the neural basis of linguistic
  computation.

70
Further predictions currently under investigation

• If the M350 indexes automatic lexical access, it
  should be
• sensitive to semantic as well as phonological
  stimulus properties.
• the first shared component between auditory and
  visual word recognition.

71

• Slides available at
• http//web.mit.edu/liina/www