The resolution of ambiguity in text a computational model

1

• The resolution of ambiguity in texta
  computational model
• Stefan L. Frank1,2 Mathieu Koppen1 Leo G.M.
  Noordman1,2 Wietske Vonk3

1 Nijmegen Institute for Cognition and
Information, Radboud University Nijmegen
2 Discourse Studies, Tilburg University
3 Max Planck Institute for Psycholinguistics,
Nijmegen
Results The models outcome is compared to
reaction time and error rate data from
experiments on pronoun resolution.
Example Bob liked Joe because he was friendly
Abstract The DSS model simulates how background
knowledge is used to disambiguate text. It
extends Frank et al.s (2003) DSS model of
knowledge-based inference by adding a process
that chooses among the texts alternative
readings. Comparing the models results to data
from experiments on the resolution of ambiguous
pronouns shows that the model correctly predicts
the effects of processing depth and context
informativeness on reading times and error rates.
Text
X(t1) Bob likes Joe
Situations

• Who is friendly? Infer either X1(t) or X2(t)
• Gravity pulls X(t) towards X1(t) and X2(t). The
  end point of X(t) represents the ambiguitys
  reading.
• Foregrounding moves initial position of X(t)
  towards X1(t) (first-mention effect)
• Context and world knowledge push X(t) towards
  possible causes of X(t1), such as X2(t).

Inference the DSS model The Distributed
Situation Space model (Frank et al., 2003)
simulates knowledge-based inferencing for story
comprehension.
Congruency agreement between foregrounding and
context, with regard to the pronouns referent
(the example is incongruent).
The effect of sentence congruency on reading
times depends on processing depth and on anaphor
type (ambiguous pronoun or unam-biguous name).
Left experimental data (Stewart et al., 2000).
Right model results (Frank et al., 2005).

• Representation Story situations are points in
  150-dimensional situation space.
• Inference These points move through situation
  space, according to microworld knowledge. A
  processing-depth parame-ter controls the
  criterion for stabilization.
• Result The end points represent the original
  situations plus inferred infor-mation. The
  processing time required predicts sentence
  reading times.
• Disambiguation the DSS model
• Foregrounding Initial interpretation of the
  ambiguity affects its position in situation space
• Gravity in situation space leads to forced choice
  among alternative situations (the readings of the
  ambiguity)

X2(t)
X(t)
correct
three dimensions from situation space
Effects of context informativeness and congruency
on error rates. Left experimental data (Leonard
et al., 1997). Right model results (Frank et
al., 2005).
possible paths of X(t)
References Frank, S.L., Koppen, M., Noordman,
L.G.M., Vonk, W. (2003). Modeling
knowledge-based inference in story comprehension.
Cognitive Science, 27, 875-910. Frank, S.L.,
Koppen, M., Noordman, L.G.M., Vonk, W. (2005).
Resolving referential ambiguity in text a
computational model. Manuscript submitted for
publication. Leonard, C.L., Waters, G.S.,
Caplan, D. (1997). The influence of contextual
information on the resolution of ambiguous
pronouns by younger and older adults. Applied
Psycholinguistics, 18, 293-317. Stewart, A.J.,
Pickering, M.J., Sanford, A.J. (2000). The time
course of the influence of implicit causality
information Focusing versus integration
accounts. Journal of Memory and language, 42,
423-443.
error
Foregrounding and gravity are knowledge
independent. Adding the inference process from
DSS makes disambiguation knowledge based.
X1(t)