Avoiding the Garden Path Eye Movements in Context

Document Sample
Avoiding the Garden Path Eye Movements in Context Powered By Docstoc
					Avoiding the Garden
Path: Eye Movements
in Context
Gerry T. M. Altmann, Alan Garnham
& Yvette Dennis
Journal of Memory and Language 31, 685-712
    Garden path theory vs.
    Incremental interactive theory
   Garden Path theory: (Clifton & Ferreira 1989)
      Initial processing: purely syntactic, context and other info
       affect reanalysis
      Principles: Minimal attachment, Late Closure
      Ambiguity resolution: favor the simpler structure


   Incremental Interactive theory: (Altmann & Steedman
    1988)
      Initial processing: contexts influence the first pass
        analysis
      Principles: referential hypothesis, principle of parsimony
      Ambiguity resolution: whether a referent can be found in
        the felicitous context
            Incremental interactive
            theory (Altmann & Steedman 1988)
    Referential Hypothesis (The principle of referential support):
     An NP analysis which is referentially supported will be favored over
     one that is not

The psychologist told the woman that he was having trouble with to visit him again.
(relative-supporting context) A psychologist was counseling two women. He was
      worried about one of them but not about the other.

    Principle of parsimony:
     A reading which carries fewer unsupported presuppositions will
     be favored over one that carries more

Did you see the man who just walked past the window?
        Garden path theory vs.
        Incremental interactive theory
1) The fireman told the woman that he had risked his life for
   to install a smoke detector.
2) The fireman told the woman that he had risked his life for
   many people in similar fires.


 GP: (1) is always garden-pathed
 Interactive: garden path could be reduced
  with referential context
          Methodology Issues
   Eye movement measures:
       First pass reading times per character in a region
       Total pass reading times per character in a region
       The number of regressions out of a region

   However…
       Effects in second pass reading is not necessarily a
        first pass parsing effect.
       Initial analysis process: first pass reading
       Need appropriate measures to make comparisons
        across different experimental conditions
        Goals of this study
   Previous studies: no unambiguous control, no
    good referential contexts, lack of eye
    movement data to support the context effect

 To test the referential hypothesis
 Use eye movements to demonstrate the
  context effect
 Provide an analysis that could reflect the
  initial process (regression contingent analysis)
            Predictions of Context effect:
                                         Null context   Felicitous context
S1) The fireman told the woman                 Rel.   Comp.
  that he had risked his life for
  to install a smoke detector.      GP G(S1)/N G(S1) NG(S2)
   (ambiguous relative)                G (S2)
S2) The fireman told the woman
  that he had risked his life for   RH G(S1)/N NG(S1) NG(S2)
  many people in similar fires.         G(S2)
   (ambiguous complement)
S3) The fireman asked the                With supporting context
  woman that he had risked his
                                         Ambiguous RC       unambiguous
  life for to install a smoke
  detector. (unambiguous control)
                                    GP          G                 NG
                                    RH         NG                 NG
Experiment 1: method

 Subject: 42 paid subjects
 Apparatus: infrared limbus eye-
  tracking system, sampled every 5 ms
 Material: 36 experimental + 34 filler
  passages, followed by comprehension
  questions, sentence by sentence
  presentation, 3 targets x 2 contexts,
  block design for context conditions
  (null & referential)
Limbus tracking
   The limbus is the boundary between the white sclera
    and the dark iris of the eye. Due to the fact that the
    sclera is (normally) white and the iris is darker, this
    boundary can easily be optically detected and tracked.
    This technique is based on the position and shape of
    the limbus relative to the head, so either the head must
    be held quite still or the apparatus must be fixed to the
    user's head.
   Due to the more or less occasional covering of the top
    and bottom of the limbus by the eyelids, "it is probably
    fair to regard limbus tracking as suitable for precise
    horizontal tracking only" (Scott & Findlay 1993).

(From:http://www.diku.dk/~panic/eyegaze/node9.html)
Experiment 1: material
Experiment 1: possible
material problems
 Supporting context for complement
  clause?
 Structural priming?
 Only one verb ‘told’ is used for
  ambiguous stimuli, 4 verbs for
  unambiguous
 Not clear about the null context or
  what’s the felicitous context for
  unambiguous sentences
       Predictions of Exp 1

                      Null        Felicitous context
                      context
Disambiguating   GP   RC: G       RC: G
region           RH   RC: G       NG
                                  RC=control
                                  Residual between comp & RC
                                   construction complexity
Ambiguous             RC=Comp     RC faster with supporting
                      Control    context compared to control
region
                      longer
Experiment 1: Results (first pass
reading time)
Experiment 1: Discussion
   Garden path effect in the null context
   Unambiguous control has longer reading time in the
    ambiguous region  because use of a relative clause
    presupposes things, whereas complement clauses do
    not. Longer reading times in RC signals the violation of
    presuppositions. Context eliminates this processing
    complexity.
   Evidence for subcategorization: if ignored, both null
    and referential context should have longer reading
    time in the ambiguous region. But no differences found
    in the referential context.
Experiment 1: Discussion

   Context has different effects on different
    regions of the ambiguous relative.
       Disambiguating region: slower reading times
        in the null context  no garden path in the
        felicitous context
       Ambiguous region: slower reading times in
        the null context  complement is read more
        slowly in the null context, because under the
        felicitous context, complement clause
        repeats info contained in the context or it is
        easier to integrate new info.
Experiment 1: Results (first
pass regression)
          Experiment 1: Discussion
   Regression data:
      More likely to have regression out of the disambiguating region for
       the ambiguous relative than the unambiguous control.
      The first pass reading time: garden path has been eliminated by
       the provision of referential context, the regression data: still some
       increased processing complexity for relatives.
         • Still garden pathed even with context, but it’s easier to recover (first
           pass reading time may not be a good garden path indicator)
         • If regression data indicates garden path in the referential context,
           then need explanations for why majority of cases (66%) do not have
           first pass regression out of the disambiguating region?
         • The discrepancy between the reading time and regression data
           suggests the regression data reflect a minority cases where people
           still garden pathed. The effect is overshadowed in the overall reading
           time.
Experiment 1: regression
contingent analysis (absence of
a regression)
Experiment 1: regression
contingent analysis (prior to a
regression)
          Experiment 1: Discussion

   Regression contingent analysis:
       Referential context only eliminates a minority
        of garden paths. Garden path still occurs in the
        felicitous context.
         • Subjects are not concentrate enough, failed to build
           the correct representations, lose track of
           information…
         • Subjects may give one of the referents more
           prominence than the other
Experiment 2

 Whether people is biased to attend
  the referential contexts
 All the procedure is the same as exp1,
  except the comprehension questions.
Experiment 2: Results (first
pass reading for exp1 &2)
Experiment 2: Results (first
pass reading)
Experiment 1&2: first pass
reading comparison
Experiment 2: Results
Experiment 2: discussion

 In exp 2, a residual difference in the
  disambiguating region between RC
  and unambiguous control was found.
 Is this contributed by minority cases?
Experiment 2: Results (first
pass without regression)
Experiment 2: Results (prior
to regression)
Experiment 1&2: prior to
regression comparison
Conclusion

   Contexts help to avoid garden path in
    most of the cases.

   Residual differences in reading times
    between the relative and the other two
    targets were due entirely to minority
    cases

				
DOCUMENT INFO