Collocational Grammar

ryone is fighting a fierce rearguard action struggling towards the objective of a discrete syntactic classes, all the while being forced by practical necessity to move closer and closer toward an analysis of relationships whenever they want to get anything done. Everyone is still trying to find relationships of classes rather than classes of relationships. In n-gram models relationships are used, but not classified, and enormous data requirements make their direct interpretation impractical. In Magerman's history based parsing he effects the classification of one half of the infinity of relationships he seeks to use, but the use of relationship classes is implicit rather than explicit. Schuetze derives relationship classes, but only within a lexical not a syntactic context 1. Generalizations are supposed to simplify problems, not create them. The lexical classes of traditional analyses are cognitive classes. They may properly be the objectives of our analyses, they need not be the means. In practice they seem to cause more problems than they solve. It is the thesis of this paper that this is because they are dependent on a more fundamental classification, that of relationships or collocational structure. I have sought to show that it is relationship classes which underlie the success of " data based " language models such as n-gram and statistical parsers, and that the most efficient way of modelling relationship classes is in terms of an analysis which factors out the greatest number of similarities in different token strings. Rather than trying to extrapolate from lexical generality to structural generality I feel 1. Others are similarly led to " relationship classes " as a means of resolving lexical ambiguity, e.g. Yarowsky (1993). The " supertags " of Joshi and Srinivas (1994) seem very close to relationship classes though their formulation is still strongly influenced by concepts of lexical generalization. we should be moving from structural generality to lexical and syntactic generality. We can still have our familiar syntax categories but only in the context of a sub-class of the wider collocational classification. The central issue of NLP becomes, not the efficient classification of parts of speech, but of col-locational regularity, the single most important tool in the analysis of language structure, an effective means of modelling similarities in strings.