Article summary: "Analogy: A Non-Rule Alternative to Neural Networks"

Royal Skousen

A paper delivered at the 21st Annual Linguistics Symposium, "The Reality of Linguistic Rules", The University of Wisconsin at Milwaukee, 12 April 1992.

Published in Rivista di linguistica 7:2 (1995): 213-231.

ABSTRACT

This paper provides a general introduction to analogical modeling of language, first comparing it with rule approaches, the traditional method of language description. An alternative procedural approach to language description is found in neural networks, but this approach contains a number of serious design defects. The current status of research on analogical modeling is also presented.

OUTLINE

1. Introduction
2. Three Basic Types of Behavior
   • categorical, exceptional/regular, idiosyncratic
3. Problems with Rules
   • leakage across categorical boundaries
   • regular items close to exceptions occasionally behave exceptionally
   • transitions in behavior
4. The Analogical Alternative
   • predicting behavior only for a given context
5. Properties of Analogical Models
   • proximity, gang effect, heterogeneity
6. Natural Statistics
   • principle of minimizing disagreement
   • quadratic, finite measure of disagreement
     • in opposition to logarithmic measures of entropy
   • most powerful statistical test
   • imperfect memory and statistical significance
7. A Prototypical Categorical Rule
   • indefinite article in English
   • leakage in one direction (an >> a)
   • persistent leakage
8. The Competence-Performance Distinction
   • rejection of performance as an overlay to competence
9. Robustness
   • predictability when crucial variables missing or other deviations from expectations
10. Probabilistic Behavior
    • Labov's nondeterministic linguistic behavior
    • accounting for probabilistic rules
11. Multivariate Analysis of Linguistic Variation
    • prediction of Arabic terms of address based on social variables
    • Sankoff's loglinear approach with one freely varying parameter
12. Predicting Unexplained Behavior
    • predicting the past tense in Finnish
    • local determination of variable significance
      • no attempt to determine global significance
    • "unimportant" variables are sometimes crucial
      • o vowel crucial in predicting the past-tense for sorta- 'oppress'
13. A Procedural Alternative: Neural Networks
    • procedural versus declarative approaches
    • neural networks, connectionism: robust characteristics
    • problems: probability matching, training the network
14. Problems with Procedural Approaches
    • variable specification
    • increasing the number of variables: an exponential explosion
    • language in time
      • sequencing of outcomes in time
15. Current Work in Analogical Modeling
    • Derwing and Skousen on past-tense of English
    • Steve Chandler on psychological aspects of analogical modeling