Predicting Grammatical Classes from Phonological Cues: An Empirical Test

This paper investigates to what extent the grammatical class(es) of a word can be predicted on the basis of phonological and prosodic information only. We report on several experiments with an artificial learning system which has to assign English wordforms to their appropriate grammatical class, using various types of phonological and prosodic information. First of all, we examine several phonological cues which were claimed by Kelly (1996) to be particularly good for distinguishing English nouns from verbs. Our results indicate that these cues are indeed partially predictive for the problem at hand and reveal that a combination of cues yields significantly better results than those obtained for each cue individually. We then show experimentally that ‘raw’ segmental information, augmented with word stress, allows the learning system to improve considerably upon those results . Secondly, we investigate several generalizations of the approach: basic segmental information also proves to be more predictive when the task * Preparation of this paper was supported by a VNC project of FWO NWO (contract number G.2201.96) and by a GOA grant (contract number 98/3). Thanks are due to Walter Daelemans, Frank Wijnen and the participants in the TROPICS conference for interesting discussions of phonological bootstrapping, and to Annick De Houwer for her critical reading of the manuscript.