An analysis of probabilistic grammatical tagging methods

Abstract

The probabilistic tagging method introduced in the CLAWS system (see for example Leech -Garside -Atwell 1983, Beale 1985, GarsideLeech 1985, 1987 has proven highly accurate in assigning the correct grammatical category labels or tags to natural language text in the Lancaster-Oslo/Bergen (LOB) Corpus. Very briefly, this method involves assigning probabilities to alternative sequenes of tag assignments, based upon (a) the collocational probabilities of adjacent hypothesized tags, p(t(n + l)|t(n)), and upon (b) the relative tag probabilities, of tags for each word, p(t(n)|w). For example, in the time, context being equal, time would be judged a noun because (a) articles are far more likely to precede nouns than verbs, and (b) time occurs over 1,000 times more frequently as a noun than as a verb.Handling such grammatical category ambiguity is an important problem for natural language processing systems, because it is so widespread. Examining the Brown Corpus gives an idea of just how widespread: about 11% of word types, and 48% of word tokens, occur with more than one category label. The actual extent of categorial ambiguity in English is certainly much higher for several reasons (e. g., the many hapax legomena, which appear to be unambiguous even though a larger corpus or native speaker intuitions may show otherwise).Similar tagging methods have been reported by DeRose (1985DeRose ( , 1988) and by Church (1988), but using dynamic programming methods to achieve linear time and space bounds. These methods also differ in that they forego detailed morphological analysis and idiom processing; Church's method adds a table of third-order transitional probabilities. Results have been generally comparable to those reported for CLAWS.In this article, I report on several investigations I carried out in order to evaluate this class of methods, and the relative effects of certain variations. This work led to my dissertation, Stochastic methods for resolution of grammatical category ambiguity in inflected and uninflected languages, completed in 1989 at Brown University (DeRose 1989). 1