Effects of a Words Status as a Predictable Phrasal Head on Lexical Decision and Eye Movements

1
Effects of a Words Status as a Predictable
Phrasal Head on Lexical Decision and Eye
Movements

• Adrian Staub and Charles Clifton, Jr.
• University of Massachusetts Amherst
• March 25, 2006

2

• Introduction
• General Question
• Is a word, phrase, or clause processed more
  easily when it is a syntactically predictable
  continuation of the sentence than when the same
  constituent is a legal, but unpredictable,
  continuation of the sentence?

3

• Introduction
• An example (Staub Clifton, 2006)
• Either Linda bought the red car or her husband
  leased the green one.
• Either Linda bought the red car or her husband
  leased the green one.
• The team took either the train or the subway to
  get to the game.
• The team took either the train or the subway to
  get to the game.

first pass reading time 572 ms
471 ms
610 ms
532 ms
4

• Introduction
• An example (Staub Clifton, 2006)
• A significant reduction in reading time on the
  second disjunct when either is present, for both
  S-coordination and NP-coordination.
• Garden-pathing in S-coordination sentences is
  eliminated, as measured by regressive eye
  movements.
• Suggests that when either appears at the
  beginning of the sentence, the parser predicts an
  S-coordination structure.

5

• Introduction
• Another example (Staub, Clifton, Frazier,
  2006)
• The teacher corrected immediately the unusual
  answer the student had given.
• The teacher applauded immediately the unusual
  answer the student had given.
• A significant reduction in reading time on a
  shifted NP when the verb is obligatorily
  transitive (e.g., corrected) compared to when the
  verb is optionally transitive (e.g., applauded)
  (cf. Wasow, 1997).

first pass reading time 756 ms
847 ms
6

• Introduction
• A predictive parser that builds obligatory
  structure in advance?
• (Crocker, 1996 Frazier Fodor, 1978 Gibson,
  1998 Kimball, 1975 Konieczny, 2000 Schneider,
  1999 Lau et al., in press)
• (Also, a bunch of presentations at CUNY 2006,
  e.g., Dikker Indefrey Warren McConnell)

7

• Introduction
• What about predicting the syntactic categories of
  individual words?
• Wright and Garrett (1984)
• lexical decision following a sentence context
• A few strange men devote EXPULSION
• A very large pine forest EXPULSION
• The interesting clock seems very TOLERABLE
• Your visiting friend should enjoy TOLERABLE

8

• Introduction
• LD latency was shorter when the word was a
  predictable phrasal head (though always a legal
  continuation)
• noun targets were faster after a transitive verb
  than after a nominal modifier (devote EXPULSION lt
  forest EXPULSION) and
• adjective targets were faster after a degree
  adverb than after a verb (very TOLERABLE lt enjoy
  TOLERABLE).
• W G suggest the parsing strategy predict
  phrasal heads
• in LD task, participants may check for
  satisfaction of prediction.

9

• Introduction
• Goals of the present experiments
• to rule out artifactual explanations for Wright
  and Garretts results
• oddness of W Gs noun-noun compounds (forest
  expulsion, husband rotation, camera growth,
  etc.)
• global differences between pre-target contexts.
• to test for a cross-categorial difference between
  adjectives and nouns, following a determiner and
• to determine whether these effects can be
  obtained in a more natural eyetracking paradigm
  (or whether they are due to response biases in
  LD).

10

• Experiments 1-3
• 32 items
• a) The supervisor decided that the fancy
  furniture would no longer be produced. (adjective
  noun condition)
• b) The supervisor decided that the porch
  furniture would no longer be produced. (noun noun
  condition)
• a) Eugene found that the quiet area was perfect
  for relaxation.
• b) Eugene found that the beach area was perfect
  for relaxation.

11

• Experiments 1-3
• E1 Lexical decision in context modifier target
  (fancy/porch).
• Prediction faster RT to noun than to adjective.
• E2 Lexical decision in context head noun
  target (furniture).
• Prediction faster RT to head noun following
  adjective than following noun.
• plus two other conditions
• E3 Eyetracking Participants read whole
  sentence.
• Prediction crossover pattern.

12

• Experiments 1-3
• Norming
• Adjectives and nouns in modifier position
  (fancy/porch)
• were matched on length
• did not differ significantly in frequency (p gt
  .9, based on HAL corpus)
• had zero lexical predictability
• sentence continuation norms (N 17) 99 of
  responses were nouns.

13

• Experiments 1-3
• Norming (cont)
• Head noun (furniture)
• had zero lexical predictability following both
  modifiers
• sentence continuation norms (N 20) following
  adjective, 95 noun responses, following noun, 7
  noun responses
• None of the adjective noun or noun noun
  combinations appeared in the Brown Corpus and
• the two versions of the sentences did not differ
  in plausibility (p gt .4), based on a norming
  study (N 20).

14

• Experiments 1-3
• In E1, E2, and E4, RSVP display

FANCY
the
206
that
100
decided
188
supervisor
188
The
188

188
188
1000 ms
15

• Experiment 1
• 32 Participants
• 32 Experimental items mixed with
• 24 items with comprehension questions, and
• 51 other fillers, including 27 with pronounceable
  nonword target.
• Mean accuracy on comprehension questions 94
• Mean accuracy on experimental trials 98

16

• Experiment 1
• Results
• F1(1, 30) 11.26, p lt .01 F2(1, 29) 4.93, p
  lt .05.

milliseconds
17

• Experiment 2
• 32 Participants
• 32 Experimental items mixed with
• 24 items with comprehension questions, and
• 71 other fillers, including 31 with pronounceable
  nonword target.
• 32 fillers were the items from E4
• Mean accuracy on comprehension questions 92
• Mean accuracy on experimental trials 99

18

• Experiment 2
• Two additional conditions
• c) The supervisor decided that the furniture
  (determiner noun condition)
• Should behave like the adjective noun condition,
  if LD latency is being driven by predictability
  of noun as the phrasal head
• d) One two three four five furniture (number
  noun condition)
• Meant to be a neutral baseline (sort of)

19

• Experiment 2
• Results
• adj noun vs. noun noun F1(1, 28) 13.36, p lt
  .01 F2(1, 27) 12.87, p lt .01
• det noun vs. noun noun F1(1, 28) 13.51, p lt
  .01 F2(1, 27) 4.62, p lt .05
• num noun vs. noun noun F1(1, 28) 4.01, p
  .055 F2 (1, 27) 8.37, p lt .01

milliseconds
20

• Experiment 3
• 36 Participants
• Fourward Technologies Dual Purkinje Gen 6 tracker
• 110 filler items (including 32 from E5)

21

• Experiment 3
• Measures
• First fixation duration
• Gaze duration sum of all fixations on the word
  before leaving word
• Go-past time sum of all fixations from first
  fixating word to first leaving it to the right

22

• Experiment 3
• Results (modifier)
• First fix F1(1, 34) 3.60, p .07 F2(1, 29)
  3.00, p .09
• Gaze F1(1, 34) 5.07, p lt .05 F2(1, 29)
  5.40, p lt .05
• Go-past F1(1, 34) 8.53, p lt.01 F2(1, 29)
  9.71, p lt .01

milliseconds
23

• Experiment 3
• Results (head noun)

milliseconds
24

• Experiment 3
• Results (Gaze - interaction of condition and
  region)

F1(1, 34) 4.56, p lt .05 F2(1, 29) 7.60, p
.01
milliseconds
25

• Experiments 1-3
• Summary
• In E1, LD latency was shorter to a noun than to
  an adjective after a determiner
• in E3, this effect was replicated in gaze
  duration and go-past time.
• In E2, LD latency was shorter to a head noun
  after an adjective than after a noun
• this effect was not significantly replicated in
  E3.
• Likely explanation Spillover processing from
  the adjective (see Rayner Duffy, 1986 Rayner
  et al., 1989 Weingartner Frazier, in
  preparation).

26

• Experiments 1-3
• Summary, cont
• In E2, LD was also shorter in the determiner-noun
  and number-noun conditions than in the noun-noun
  condition.
• Noun is a predictable head after both a
  determiner and an adjective
• Suggestion of inhibition in the noun-noun
  condition, compared to the neutral number-noun
  condition.

27

• Experiments 4-5
• E4
• Lexical decision in context
• 32 items
• 32 participants
• manipulated presence vs. absence of degree
  adverb, and adjective vs. nonword target.
• a nonword control was necessary to rule out main
  effect of context type

28

• Experiments 4-5
• Materials are modified versions of W Gs
• (a) The ducks in the campus pond tend to eat
  the totally SLIMY
• (b) The ducks in the campus pond tend to eat
  the totally SPINT
• (c) The ducks in the campus pond tend to eat
  the SLIMY
• (d) The ducks in the campus pond tend to eat
  the SPINT
• Prediction adjective LD faster when degree
  adverb is present.

29

• Experiments 4-5
• E4 results
• Interaction F1 (1, 28) 7.53, p lt .02 F2 (1,
  31) 4.04, p .05.

milliseconds
30

• Experiments 4-5
• E4 results, cont
• Adjective LD faster when degree adverb present
  (635 ms vs. 665 ms)
• difference between nonwords not significant (690
  ms vs. 678 ms).

31

• Experiments 4-5
• E5
• Eyetracking
• 32 items
• 32 participants
• four different degree adverbs very, totally,
  slightly, and fairly
• when adverb absent, additional word inserted
  early in sentence, to control linear position of
  critical word
• determiner and degree adverb matched in length,
  to control launch position

32

• Experiments 4-5
• Examples
• (a) The auto mechanic used some flexible plastic
  to fix the problem.
• (b) The mechanic used some very flexible plastic
  to fix the problem.
• (a) The mountain guide pointed out numerous
  dangerous trails in the area.
• (b) The guide pointed out numerous slightly
  dangerous trails in the area.

33

• Experiments 4-5
• E5 results (adjective)

milliseconds
34

• Experiments 4-5
• E5 results (noun)
• First fix F1(1, 30) 4.99, p lt .05 F2(1, 30)
  2.80, p .11.
• Gaze F1(1, 30) 30.49, p lt .01 F2(1, 30)
  6.93, p lt .02.
• Go-past F1(1, 30) 4.38, p lt .05 F2(1, 30)
  2.75, p .11.

milliseconds
35

• Experiments 4-5
• Summary of E4 and E5
• In E4, LD was faster to an adjective after a
  degree adverb than when the adjective directly
  followed a determiner.
• In E5, nonsignificant effect on adjective
• but significant gaze duration effect on
  subsequent noun.

36

• Experiments 4-5
• Summary of E4 and E5, cont
• Plausible spillover explanation, again last
  fixation before the adjective was significantly
  longer (273 ms vs. 259 ms) when the preceding
  word was a degree adverb.
• This may have counteracted effect on adjective,
  leaving only a modulation of the adjectives
  spillover effect on subsequent noun.
• Alternate story compositional semantic
  interpretation takes place on the head noun, not
  the adjective (cf. Kamp Partee, 1995)
• so prediction of an adjective benefits processing
  primarily at this later point.

37

• General Discussion
• Wright and Garretts findings are not
  artifactual E2 and E4 replicated their basic
  findings (with better-controlled materials).
• LD to a noun faster after an adjective or a
  determiner than after another noun
• LD to an adjective faster after a degree adverb
  than after a determiner.

38

• General Discussion
• E1 (and E3) also showed that nouns are processed
  more quickly than adjectives, after a determiner.
• The pattern of results in the eyetracking
  experiments was complicated by the presence of
  spillover processing, so.

39

• General Discussion
• E6 (cf. E3) (in progress, N 22)
• in collaboration with Sandy Pollatsek and Jukka
  Hyönä
• eyetracking
• 48 items 24 lexicalized noum compounds, 24 novel
  compounds
• a) The maid caught the elevator mechanic who was
  goofing off.
• b) The maid scolded the elevator mechanic who
  was goofing off.
• Because elevator must be a modifier in (b),
  mechanic is a predictable phrasal head. Since
  the preceding lexical item is identical, maybe no
  spillover effect, in contrast to E3?
• Preliminary Results First fixation on mechanic
  284 ms in (a), 271 in (b). A very early effect!

40

• General Discussion
• Taken together, the results do support the idea
  that the parser predicts phrasal heads.
• When grammatically obligatory? Role of
  transitional probability between syntactic
  categories?
• e.g., 99 nouns after determiner in cloze task
• but see E6, where noun-noun sequence is constant
• Is the strategy to predict the phrasal head as
  the very next lexical item? Or a less
  determinate prediction?

41

• General Discussion
• Likely an effect on syntactic integration, not
  lexical access
• Classic cross-modal priming work by Tanenhaus and
  colleagues
• but see recent eyetracking studies by Folk and
  Morris (2003).
• When a word is a predictable phrasal head, it may
  be slotted into a pre-built structure (as in a
  left-corner parser).

42

• General Discussion
• Methodological conclusion
• Eyetracking gives the real story so
• watch out for artificially slow presentation
  rates (cf. SPR, ERP), or artificially long SOA
  before critical word, which can give extra time
  for various factors to influence processing.
• Especially when prediction is involved!

43

• General Discussion
• One Future Direction ERP
• What is the ERP waveform associated with the
  contrast between getting a word thats a
  predictable head, and getting one that isnt?
• An attenuated version of the P600 or LAN?

44

• Thanks to
• Lyn Frazier
• Keith Rayner

45

• Thanks!