Data compression, generalisation and overgeneralisation in an evolving theory of language development
James Gerard Wolff · 1980
This paper describes a computer model of first language acquisition, program SNPR14, which is a development of an earlier model, MK10/GRAM15, described elsewhere [Wolff, 1978a and b]. Like the earlier model, SNPR14 shows some success in discovering a phrase-structure grammar from an unsegmented, semantics-free, language-like text given only that text as data. SNPR14 is designed to remedy certain weaknesses in MK10/GRAM15. In particular it provides a tentative answer to the problem of how it is that children, in forming syntactic generalisations, can distinguish correct generalisations from overgeneralisations and can eliminate the latter whilst retaining the former. The notion that language development and data compression are intimately related has been extended here: it is suggested that language development is, in part, a process of building a grammar in such a way that the effectiveness of the grammar as a means of compressing cognitive data is maximised for any given size of grammar.