Modeling Reading Development From First Grade Text

1
Modeling Reading Development From First Grade Text

• Michael W. Harm, CMU
• Mark S. Seidenberg, Wisconsin/Madison

2
Why Construct Computational Models?

• Can complement empirical studies
• Allow manipulation of multiple factors that would
  be prohibitively difficult in a classroom setting
• We can look under the hood and see what kinds
  of internal representations are formed during
  learning

3
Computational Models of Reading

• Have a long history
• Rumelhart McClelland 1981, Seidenberg
  McClelland 1989, Plaut Hinton, 1991, Coltheart
  et. al, 1993, Plaut et. al 1996, Zorzi 1999, etc.
• Historically, have been aimed at explaining
  adult performance
• Architecture of the reading system
• Acquired dyslexia (due to brain damage)
• Have traditionally not examined issues of reading
  development

4
Modeling Development Was Difficult With Earlier
Models

• Earlier models did not learn about phonology
• But phonology is important in learning to read!
• Sampled large corpus of (adult) text in toto
• Did not allow for manipulation of actual sequence
  or nature of word exposures
• Prohibited examination of effects of different
  basals, interventions

5
The Harm Seidenberg 1999 Model of Reading
Begin by modeling pre- literate phonological
knowledge that children have
Can vary the strength and consistency of this
knowledge
and simulate the different degrees of
phonological ability children bring to bear
learning to read
6
Reading Uses this Phonological Knowledge
The nature of the phono representations
influences what is learned during reading
Core result the phonologically impaired model
learns differently
7
Two New Applications
We have applied the Harm Seidenberg 1999 model
to two novel applications

• Simulation of effects of different instructional
  basals
• Simulation of an reading intervention scheme

8
Simulating Effect of Different Reading Basals

• Some early reading texts are more tuned to
  overlap in spelling/sound
• Others emphasize variety in text exposure to
  wider range of words

Q How does this interact with reading
impairments?
9
Results Nonword Reading

• Basal 1 More tuned to spelling/sound
  correspondences
• Basal 2 Less systematic text
• For both normal and impaired model, Basal 1
  better than Basal 2.

10
Results Word Reading

• For normal models, large effect of basal
• For impaired models, floor effect smaller
  difference

11
Summary Simulating Effects of Basals

• We can explore the impact of reading materials,
  and differential effects on normal and impaired
  reading
• Holds promise for more sophisticated
  explorations/manipulations
• and direct ties to more fine grained properties
  of basals (see other talks in this session)

12
Simulating Reading Interventions

• There is extensive evidence that one cause of
  poor reading development is a phonological
  impairment
• However, interventions targeted at auditory
  phonology generally are not very effective
• Interventions aimed at spelling/sound
  representations have greater success

Why?
13
Analysis of the Model
Phonology
Spelling
So effective interventions must target the
relationship between spelling and sound
14
The McCandliss et al. Intervention

• Use lessons based on the Beck word-building
  scheme
• Break words apart when errors are made
• Emphasizes componential structure of words
• and its relation to components of sound

15
Performed Simulation of this Intervention

• Used phonologically impaired simulation from Harm
  Seidenberg 1999
• Simulated intervention using actual items from
  lessons
• Brought about improvements in nonword reading
• Analyzed internal representations of words

16
Normal and Impaired Simulation
17
Representations Cluster Better Due to Intervention
Remediated Simulation
18
Conclusions

• Computational simulations are now poised to
  explore more detailed aspects of childrens
  experience
• With reading basals
• And with interventions
• Opens up a promising new line of research linking
  behavioral experimentation with computational
  analysis

19
With thanks to...
Mark S. Seidenberg Bruce McCandliss
20
fin