Inducing abstract linguistic representations: human and connectionist learning of noun classes

Noun class information is a crucial component of the interface between the lexicon and the grammar. In order to explain linguistic productivity it is necessary to assume that linguistic rules are defined not over specific words, but classes of word. This is not only true given the classical distinction between lexicon and grammar, but also in 'emergentist' views which see no clear separation between these two systems (Ellis, 1998; Tomasello, 2000). Even though the latter stress the lexical-specificity of many 'grammatical rules', it is still recognised that adult productivity can only be explained if words are grouped into classes, even if those classes do not map neatly onto traditional linguistic categories. The way in which words are grouped into grammatical classes is therefore an important issue in understanding language development, particularly in explaining the leap from lexical learning to grammar learning. Noun classes, such as grammatical gender, are fundamentally abstract, grammatical, notions (Corbett, 1991). However, attempts have been made to uncover subtle phonological and semantic cues that can be used to predict a word's gender (Kelly, 1992). For example, masculine nouns in German are more likely to be monosyllabic, and monosyllabic words that are masculine contain more consonants than those of other classes. In French, feminine nouns tend to end in closed stressed syllables (e.g. personne, tomate, viande), and masculine nouns tend to end in open stressed syllables (e.g. avion, bruit, chapeau, bain). There are also a number of characteristic derivational morphemes associated with each gender (e.g.-eur and-ment are masculine, and-tion,-euse,-iere are feminine). (Sokolik & Smith, 1992) trained a connectionist network to classify French nouns as either masculine or feminine. The network was presented with the orthographic, rather than phonological, forms of the words. They found that it could then indicate the gender for nouns that it had not received during training, although its performance was 2 not perfect (ranging between 73% and 75%). This indicates that there are regularities in the form (in this case spelling) of French words which can to a certain extent predict gender category. Yet there are always words which fall stubbornly outside such generalisations. In the case of French, Carroll (2001) argues that in any case, the kinds of phonological cues that have been appealed to are more subtle than could reasonably be expected to be represented in the lexicon. This is not to say that phonological and semantic cues do not play a …