Schema constraint network from feature-applicability judgements

Description

Build a network of pairwise feature constraints from people's applicability judgements, then clamp features and let it settle to reveal default schemas, blends and subschemata.

In the PDP volume's rooms example, 40 descriptors were judged present or absent for imagined offices, living rooms, kitchens, bathrooms and bedrooms, 16 times each. For every pair of descriptors the joint probabilities were computed, and a symmetric weight set as the negative log of discordant over concordant probabilities, with biases from base rates; weights were derived, not learned. Clamping one descriptor on and letting the network climb to a goodness maximum fills in a default: oven yields the default kitchen, bed the default bedroom. Combinations that fit no single category produce blends, and small groups of features act as subschemata. The set of maxima and basins describes the schema structure. The book notes the data collection was informal, with two judges and some adjusted entries, so the procedure is illustrative rather than validated.

Attributes

Loss type
Extinct
Data
Mixed
Effort
Days

Steps

  1. Choose a set of descriptors relevant to the domain, for example 30-50 features.
  2. Ask people to imagine an instance of each category and judge for every descriptor whether it applies, repeating across categories and instances.
  3. Compute for each pair of descriptors the joint probabilities of both on, both off and each on alone.
  4. Set a symmetric weight as the negative log of discordant over concordant joint probabilities, and a bias from each descriptor's base rate.
  5. Clamp one or more descriptors on and update the network until it reaches a goodness maximum.
  6. Read the settled state as the default schema; repeat with other clamps to map maxima, blends and subschemata.

Sources

  • Parallel Distributed Processing, Vol. 2 (McClelland & Rumelhart, 1986)

Also known as

  • Schema constraint network built from feature-applicability judgements
  • Rooms schema network
  • Clamp-and-settle schema analysis
  • Goodness-landscape schema analysis

Categories

  • User Research
  • Cognitive Psychology & Perception