Issues in modeling language processing analogically§

Exemplar-based models of language assert that linguistic processing involves analogy to past linguistic experiences stored in the mental lexicon. This study explores how three factors influence the predictions made by exemplar-based simulations of linguistic phenomena. Three questions are posed as they relate to such simulations: (1) Is type frequency or token frequency a better predictor of outcomes? (2) What is the optimal way of aligning the variables in the database so that the most relevant analogs are found? (3) Are there significant differences between representing variables as phonemes versus representing them in terms of distinctive features? Spanish stress assignment and English past tense formation served as the linguistic phenomena on which these issues were tested. The results suggest that type frequency is a better predictor of outcomes, although simulations using token frequency were most successful when only middle frequency words were included. Several methods for aligning variables in the analogical database are discussed. The dual-alignment method has advantages for the English past tense task, but not in predicting Spanish stress. In the Spanish task, strict phonemic representation of words demonstrated no advantage over feature representation. However, phonemic representation produced better results than distinctive features in predicting the English past tense. # 2003 Elsevier B.V. All rights reserved.