Morphology Learning

Regular vs. Irregular Morphology

• In (probably) every language with bound morphology, children have to learn both "rules" and exceptions to the "rules".
• There may also many categories of regulars and degrees of regularity.
• Children typically exhibit three overlapping stages (the details are extremely controversial)
  • They memorize very frequent forms.
  • They begin to learn "rules", applying them incorrectly to irregular forms (over-regularization).
  • They correctly sort out the regulars and irregulars.
• Dual-mechanism hypothesis (Pinker, etc.)
  • Morphological rules are learned by a symbolic device.
  • Exceptions are learned by a rote pattern associator, for example, a feedforward connectionist net.
  • Assumes there is a default, regular process for every grammatical function (for example, -ed for English past tense)
• Single-mechanism hypothesis (Rumelhart and McClelland, Marchman and Plunkett, MacWhinney and Leinbach, Seidenberg)
  • Regular and irregular morphology is learned by a single pattern associator.
  • The models exhibit over-regularization errors.
  • The approach is well-suited to handling degrees of regularity.

A Connectionist Model of Morphology Acquisition (Gasser)

• Some phenomena
  • Children learn to recognize and produce novel combinations of familiar morphemes for a wide variety of types of morphological "rules"
  • Morpheme recognition is, roughly speaking, a supervised task.
  • Morpheme production is not a supervised task but must build somehow on morpheme recognition.
• Some questions
  • What sort of architecture could achieve this performance, given plausible training? Does it require pre-wired linguistic constraints?
  • How is the task to be realized?
• An approach
  • Form: sequences of phones (or "acoustic" or "articulatory" segments)
  • Morphemes: constant patterns representing root and grammatical morphemes
  • Architecture

    

    • Recognition
      • Elman net
      • Horizontal modularity (separate networks for root and grammatical morphemes)
    • Production: two components mediated by "phonological" representations derived from recognition learning
      • Morphemes -> phonology: Elman/Jordan net
      • Phonology -> "articulation": Elman/Jordan net (though this should eventually not be supervised)
• Performance
  • For both recognition and production (morphemes -> phonology), generalizes for many kinds of morphological "rules": suffixation, prefixation, infixation, mutation, templates, deletion, but not reduplication and metathesis
  • Phonological representations also support "articulation"
  • Generates predictions about relative difficulty of "rule" types and error types

[L645 Syllabus]