📄 NETL: A System for Representing and Using Real-World Knowledge

Summary

This entry covers Stuart C. Shapiro's review, published in American Journal of Computational Linguistics (Vol. 6, No. 3-4, 1980), of Scott Fahlman's book NETL, a lightly revised version of his 1977 PhD thesis. NETL stores real-world knowledge as a network of simple parallel elements. Nodes hold marker bits, and links spread those markers across the network. Fahlman claims this lets searches and common inferences run in roughly constant time regardless of knowledge-base size, but only on dedicated parallel hardware. The reviewer finds the proposed hardware design weak and judges the representation on its own merits. NETL is built around a property-inheritance hierarchy of "typical member" nodes, with individuals treated as virtual copies of their types. Shapiro credits Fahlman with the most serious attempt yet at this technique. He also shows that Fahlman's own findings, especially copy-confusion, reveal its limits. Shapiro recasts the constructs in predicate-calculus terms and criticizes the book for inventing jargon instead of engaging with logic. He still considers it essential reading for inheritance-hierarchy research.

Key points

• Book details: MIT Press, 1979, 285 pp., $17.50. Each node unit stores about a dozen marker bits.
• Core node and link types:
  • *INDV-nodes represent individuals and sets.
  • *TYPE-nodes represent typical members.
  • *VC (virtual-copy) links connect individuals to their types, e.g., CLYDE → TYPICAL-ELEPHANT, with ELEPHANT-SET as the set node.
• Parts and roles are handled with role-nodes attached by EXISTENCE-links. *MAP and *TMAP nodes, with map-wires and owner-wires, store exceptions about specific parts such as Clyde's left front leg.
• Copy-confusion: collapsing *VC chains can merge distinct roles, for example confusing Clyde's weight with his trunk's, or making Clyde his own father. Fahlman weighs four fixes and adopts the fourth, accepting that parallel inference is incomplete.
• Shapiro's logical reading:
  • A *TYPE-node acts as a universally quantified variable.
  • A role-node is existentially quantified, with a Skolem-style arc.
  • The restriction on *EVERY-nodes rules out statements quantifying over two variables.