| |
Abstract:
An instance-based learning mechanism is described which
produces quality sentence completions. The mechanism acquired
instances (five-word sequences) based on sentences of four levels
of complexity (no embeddings, right-embedded subject or object
relatives, and center-embedded clauses), created from a twenty word
lexicon containing the most common lexical categories. Instances
were linked to lexical representations (initially entirely random).
During training, the instance was retrieved which best matched the
five current words of a new sentence. No external feedback was
given for non-matches; however, the lexical representations were
randomly changed, weighted towards increasing feature matches. At
the end of training, clear categorization of the lexical items
emerged. During testing, instances were retrieved on the basis of
the three most recent words of novel sentences and used to generate
two word continuations. The level of error was in general low
(comparable to an SRN), and implausible continuations such as "a
he" were avoided. Errors for all sentence types were less than for
word lists. Effects of sentence complexity were also found.
Instances can form part of a learning mechanism which, without
external feedback, can form functional categories and predict
grammatical continuations with a reasonable degree of accuracy.
Instance-based models deserve attention, since retrieval from
memory appears to demand less resources. The sort of model
discussed here can explain why measures of processing effort
decrease with proficiency
|
