Grammatical Trigrams: A Probabilistic Model of Link Grammar

In this paper we present a newclass of language models. This class derives from link grammar, a context-free formalismfor the description of natural language. Wedescribe an algorithm for determining maximum-likelihood estimates of the parameters of these models. The language models which wepresent differ from previous models based on stochastic context-free grammarsin that they are highly lexical. In particular, they include the familiar n-grammodelsas a natural subclass. Themotivation for considering this class is to estimate the contribution whichgrammar can maketo reducingthe relative entropy of natural language.