Lost on the garden path: Exploring misinterpretation and good enough language processing

1
Lost on the garden path Exploring
misinterpretation and good enough language
processing

• Kiel Christianson
• Dept. of Educational Psychology
• Beckman Institute

2
Collaborators

• Fernanda Ferreira
• Carrick Williams
• Andrew Hollingworth
• Rose Zacks
• Tim Slattery
• Susan Garnsey
• Laura Matzen
• RAs in my lab (Kent Lee, Jeong Ah Shin, Ji Kim,
  Jung Hyun Lim, Heeyoun Cho)

3
So we dont get lost ourselves,a brief map

• What are garden path sentences?
• And why are they interesting?
• Why worry about interpretation?
• And why havent other psycholinguists until
  recently?
• Basic data
• Christianson, Hollingworth, Halliwell, Ferreira
  (2001)
• Christianson, Williams, Zacks, Ferreira (in
  press)
• Recent data
• Christianson Slattery (2005, in prep)
• Christianson (still running!)
• Some semblance of a conclusion, I hope
• A working definition of good enough
• Parsing, processing, and interpretation
• Implications

4
What relevance to SLA?

• Theoretical Do L2 speakers parse L2 same as L1
  speakers do?
• Pedagogical Misinterpretations can be
  informative wrt mental representations
• You dont know for sure unless you ask!

5
Garden path sentences

• Sentences that lead the human sentence processor
  (HSP) to construct an initial syntactic
  structure, which turns out to be incorrect, and
  thus requires syntactic (and semantic) reanalysis.

6
Example

• While

7

• While Anna

8

• While Anna dressed

9

• While Anna dressed the

10

• While Anna dressed the baby

11

• While Anna dressed the baby spit

12

• While Anna dressed the baby spit up

13

• While Anna dressed the baby spit up on

14

• While Anna dressed the baby spit up on the

15

• While Anna dressed the baby spit up on the bed.

16

• While Anna dressed the baby spit up on the bed.
• the baby ambiguous noun phrase (ambiguous
  region)
• spit up disambiguating verb (disambiguating
  region)

17
Why use sentences like this?

• They induce difficulty and observable slow-downs
  in processing that is normally smooth and fast
• Point is to observe how the system is perturbed,
  and how it recovers
• Not all suffer from mistakes
• Put the book on the shelf in my backpack.

18
Frazier Rayner (1982)

• The garden path theory of syntactic parsing
• Eye-tracking used to measure how people read such
  sentences
• Predictable patterns
• Longer fixations (reading times) on
  disambiguating verb
• Regressive eye movements to ambiguous NP and
  subordinate verb (dressed)
• Serial, modular model
• one parse at a time, just syntax first
• (But this architecture isnt crucial for
  assumptions that follow.)

19
Traditional assumptions (no matter what parsing
model)

• Garden path sentences can be handled one of two
  ways
• Mis-parse is recognized by the HSP, revision is
  undertaken if not successful, processor gives up
  and interpretation is not achieved
• Ambiguity/mis-parse isnt noticed at all person
  just keeps reading

20
Questioning traditional assumptions

• Does the mis-parse HAVE to be reanalyzed
  syntactically?
• Does the interpretation HAVE to be revised?
• Automatic?
• MacDonald et al. (1994) There might be
  situations in which the communicative goals of
  the listener can be achieved with only a partial
  analysis of a sentence, but we view these as
  degenerate cases (p. 686).
• (An assumption made by proponents of both serial
  and parallel models of parsing)

21
Good enough sentence processing

• Ferreira Henderson (1999) Christianson, et al
  (2001) Ferreira, Christianson, Hollingworth
  (2001) Ferreira, Bailey, Ferraro (2003)
  Christianson, et al (in press)
• Loosely defined as processing in which the HSP
  settles for a parse that is in some way
  incomplete or underspecified, resulting in an
  interpretation that is not faithful to the input.

22
So why worry about interpretation?

• The central problem for future theories of
  sentence processing is the development of
  theories of sentence interpretation.
• --Frazier (1998)
• (Besides, isnt the whole point of language to
  derive meaning?)

23
How do we go about studying interpretation?

• Traditionally, we dont.
• comprehension question for every 4th sentence or
  so, just to make sure theyre not zoning out
• While the man hunted the deer that was brown and
  graceful ran into the woods.
• Q Was the deer brown? OR
• Was the deer in the woods?
• Key Q (never asked) Was the man hunting the deer?

24
What happens to the interpretation generated by
the initial mis-parse?

• Does it linger?
• Does it just disappear?
• Can it block a full reanalysis?
• Can it cause interpretive difficulties even after
  the rest of the sentence is read?

25
Christianson, Hollingworth, Halliwell, Ferreira
(2001)

• What happens to that original, incorrect
  interpretation derived from the initial, partial,
  and ultimately incorrect parse?
• If syntax (and, it is generally assumed,
  consequently semantics) fully reanalyzed, it
  should not influence final interpretation
• Major assumption If interpretation is incorrect,
  then full reanalysis has not taken place.
• Syntactic representation remains incomplete, and
  thus the interpretation is incorrect
• Might be too strong Maybe syntax OK, semantics
  never fixed

26
Expt. 1b

• (1a) While Bill hunted the deer (that was brown
  and graceful) ran into the woods.
• (1b) While Bill hunted the deer (that was brown
  and graceful) paced in the zoo. (implausible)
• (1c) While Bill hunted the pheasant the deer
  (that was brown and graceful) ran into the woods.
  (non-GP)

27
How to judge interpretation?

• Radical Just ask.
• Q Did Bill hunt the deer?
• YesINCORRECT NoCORRECT

28
Results Expt. 1b
Also gathered confidence ratings No diff. in
any condition in any expt. VERY confident.
29
Expt. 2

• Maybe no reanalysis at all?
• Maybe just inference (despite the length of
  ambiguous region effect in 1b)?
• (2a) While Bill hunted the brown and graceful
  deer/the deer that was brown and graceful ran
  into the woods.
• (2b) The brown and graceful deer/the deer that
  was brown and graceful ran into the woods while
  Bill hunted.

30
Another question, too

• Did Bill hunt the deer?
• (subordinate clause question)
• OR
• Did the deer run into the woods?
• (matrix clause question)

31
Expt. 2 results
32
Expt. 3

• So far, baseline inference, but syntactic
  manipulations push effect around above and beyond
  inference.
• Conclusion Syntax not fully reanalyzed
• YetWouldnt it be nice to find a syntactic
  structure that, if fully reanalyzed, would NOT
  ALLOW THE INFERENCE?
• Reflexive absolute transitive (RAT) verbs

33
RAT verbs

• While Anna dressed the baby that was cute and
  cuddly spit up on the bed.
• ?If fully reanalyzed, Anna CANNOT be dressing the
  baby must be dressing HERSELF.

34
Results Expt. 3a-b
35
Conclusion

• Good enough sentence processing
• Syntactic parse not fully reanalyzed
• If it is, its not mapped onto semantics
• Processor happy with incomplete analysis as long
  as it is plausible.
• Likely the deer overtly serves as subject of
  matrix clause, remains syntactically present as
  object of subordinate.

36
Older vs. younger readersChristianson, Williams,
Zacks Ferreira (in press, Discourse Processes)

• Perhaps misinterpretation effect larger for older
  readers?
• Caused by decrement in inhibitory control in
  older folks (Hamm Hasher, 1992 Hasher, Zacks,
  May, 1999)
• Older readers might even be worse at inhibiting
  initial incorrect parse.

37
Expt. 1

• OPT verbs Garden path Structure (subordinate-main
  clause order)
• While the man hunted the deer that was brown and
  graceful ran into the woods.
• Non-garden path Structure (main-subordinate
  clause order)
• The deer that was brown and graceful ran into the
  woods while the man hunted.
• Q Did the man hunt the deer?
• RAT Garden path Structure (subordinate-main
  clause order)
• While Anna dressed the baby that was small and
  cute played in the crib.
• Non-garden path Structure (main-subordinate
  clause order)
• The baby that was small and cute played in the
  crib while Anna dressed.
• Q Did Anna dress the baby?

38
Results Expt. 1
39
Reading span correlations
40
Expt. 2

• Maybe olders more likely to infer (Hartmann
  Hasher, 1991)
• OPT verbs allow inference RAT do not
• If so, should see exaggerated effect in plausible
  conditions for older readers
• Also manipulated length of ambiguous region to
  see if longer-held interpretations harder to
  inhibit

41
Sentences

• Long Ambiguous Region -- Plausible/Implausible
• Garden path Structure (subordinate-main clause
  order)
• While the man hunted the deer that was brown and
  graceful ran into the woods/paced in the zoo.
• Non-garden path Structure (main-subordinate
  clause order)
• The deer that was brown and graceful ran into the
  woods/paced in the zoo while the man hunted.
• Short Ambiguous Region -- Plausible/Implausible
• Garden path Structure (subordinate-main clause
  order)
• While the man hunted the deer ran into the
  woods/paced in the zoo.
• Non-garden path Structure (main-subordinate
  clause order)
• The deer ran into the woods/paced in the zoo
  while the man hunted.

42
Results, Expt. 2

• Main effects of ambiguous NP length,
  plausibility, and age
• BUT
• Age did not modulate the effect of plausibility
  (F1lt1 F2lt1) nor did it influence the effect of
  ambiguous NP length (F1lt1 F2lt1).
• As in Expt 1, age didnt interact with sentence
  structure, either.

43
Not inhibition or inference

• As far as we can tell, anyway
• Maybe no inhibition required? Never an
  alternative full parse/interpretation
  constructed?
• Very good enough-y
• If inference not the issue either why better at
  RATs than OPTs?
• Maybe they arent.

44
Expt. 3

• RAT sentences
• While Anna dressed the baby that was cute and
  cuddly spit up on the bed.
• Another question Did Anna dress herself?
• (Answer should be YES!)

45
Results Expt. 3
(nGP The babyspit upwhile Anna dressed.)
46
WM correlations with wrong answer rate in Expt. 3
47
Explanation

• OPT verbs Two ways to answer Did the man hunt
  the deer?
• Recall verbatim and figure out
• Recall propositional content
• The man hunted the deer (initial parse)
• The man hunted SOMETHING unspecified
• Congruent with OPT verbs
• Olders more likely to rely on gist
  (propositional) rather than verbatim content

48
RAT

• Propositional content
• Anna dressed the baby (initial)
• Anna dressed SOMETHING specific
• But what?
• Semantics of RAT verbs dont allow congruency of
  propositional content
• dress doesnt allow unspecified interpretation
• In order to get reflexive reading, must reactive
  syntax to establish government relation and
  co-indexation

49
WM tie-in

• Olders with less WM resources unable to
  reactivate the syntactic structure required to
  get the reflexive reading.
• Processor may settle on good enough
  interpretation, but to answer the question, you
  need more than that
• If not enough working memory available to either
  keep working on structure or recall, recompute,
  and revise, then stuck.

50
Christianson Slattery (2005)

• No one has ever looked to see if garden paths
  affect SUBSEQUENT reading
• Why should they? Recall traditional assumptions.
• If good enough processing takes place, should
  see people moving on to read subsequent text
  before theyve completed a full reanalysis.

51
Method

• Context AFTER garden path sentence (eye-tracking)
• opening region S1 ambiguous NP1
  region
• While the man hunted(,) the deer that was
• disambiguation
• large and brown ran into the woods.
• opening region S2 NP2 region
• The man was hunting a deer (bear) in the
  woods.

52
Results

• Clear classic GP effects
• First pass time
• ME of struct. on NP1 72ms longer when non-GP
• ME of struct. on disambiguation 56ms longer when
  GP
• ME of NP2 on NP2 158ms longer when mismatched

53
New Results

• Go Past time (includes re-fixations after
  leftward regressions)
• ME of struct. on disambiguation 264ms longer
  when GP
• ME of struct. on NP2 86ms longer when GP
• ME of NP2 on NP2 248ms longer when mismatched
• Marg. ME (p .081) by P of struct. on S2
  opening region

54
Summary

• Robust GP effects in early and late measures
• Clear indication that readers moved on to S2
  before structural work on S1 was completed
• Lack of interaction suggests that processes
  related to structural revision and lexical
  content are separate. S1 ambiguity lingers into
  S2 amplified by NP2, irrespective of match.

55
What is good enough processing?

• NOT shallow parsing
• In other words, not just lack of effort
• Confidence ratings downstream effects of GP
  structure
• Results in SOME kind of underspecified
  representation
• Which representation (syntax, semantics, both,
  other)?
• Underspecification likely result of Incomplete
  Processing (good enough)
• Interpretation formed before all sources of
  information are available (some sources slowed by
  computational demands)
• Processor moves on (even if some processes are
  still running)

56
Christianson (in preparation)

• Change detection paradigm
• (Sanford, et al., 2005)
• Memory for text based on representation
  constructed for it.
• Changes to text that are consistent with
  representation should be harder to detect.

57

• The cookout was going well so far. While Tom
  grilled the hot dog that was long and fatty began
  to burn. The burgers sure looked good, though.

58

• The cookout was going well so far. While
• Tom grilled the hot dog that was long and
• fatty it began to burn. The burgers sure looked
  good, though.

59
Conditions

• Garden path vs. non-garden path (comma)
• While Tom grilled, the hot dog that was long and
  fatty began to burn.
• NP-it vs. it-NP
• While Tom grilled it the hot dog that was long
  and fatty began to burn.

60
Results, Expt. 1
Sig. ME of structure order Sig. INTERACTION
61
Summary

• People more sensitive to changes in GP sentences
• NOT shallow processing processor notices the
  ambiguity
• Change acts like question in Christianson, et al
  (in press) and NP2 in Christianson Slattery
• Spurs processor to resolve lingering structural
  problem by some means, because that information
  becomes critical for interpretive task

62

• However, significant interaction (p .018)
  suggests that in GP condition, sometimes the
  partial reanalysis proposed by Christianson et al
  (2001) DOES take place
• Two hot dogs in representation, congruent with
  addition of it in DO position of subordinate
  clause

63
Conclusion

• Good enough processing results in interpretations
  not faithful to the content
• Not previously noticed by researchers because
  right questions not asked
• Not usually noticed by people because usually not
  critical for integration of later material (often
  even incorrect interpretation can be plausibly
  maintained in context)
• Good enough, not just shallow
• Processor actively tries to resolve, but may move
  on because resources are limited, and input is
  not
• (The Life is short! model of sentence
  processing)

64
Implications (psycholinguistic)

• Suggests different mechanisms for parser and
  processor
• Parser worried about getting a licit syntactic
  structure (but might truncate the parse, too)
• Processor worried about getting a plausible,
  contextually consistent interpretation
• Parser might be slowed down by ambiguities
• Processor might run ahead and not check final
  parse unless underspecified representation
  results in an interpretation that doesnt fit in
  context

65
Implications (general)

• Extent to which parser keeps working or processor
  can look back at results probably depends on STM
  capacity
• STM or other individual differences likely
  predictive of eventual interpretation accuracy
• Over-reliance of processor on top-down (semantic,
  discourse) information (perhaps compensatory)
  might accentuate misinterpretations (whether it
  affects syntactic parse or not)
• Older readers, L2 readers, struggling readers,
  young readers

66

• Good enough usually good enough, but not always.
• Misinterpretations informative for theorists
• can be predicted and manipulated consistently
  enough to be exploited in reading research and
  instruction (e.g., to increase meta-linguistic
  awareness)

67

• Thank you!