Learning linguistic models from data

Abstract

In this chapter we have proposed two types of linguistic models for both classification and prediction problems. Mass relational methods build conditional models for each class or output focal set using the idea of a mass relation as described in chapter 4. These are then integrated into a Bayesian classification or estimation framework. Linguistic decision trees extend the decision tree formalism to include constraints on attributes defined by label expressions. Classification and prediction is then based on probabilities evaluated using Jeffrey's rule. The efficacy of these models was investigated from the dual persp ectives of predictive accuracy and transparency. Both methods have been shown to give good accuracy across a number of classification and prediction problems with accuracy levels that are comparable with or better than a range of standard machine learning algorithms. In terms of transparency, mass relations generate sets of quantified atomic input-output conditional rules and can be used to evaluate queries or hypotheses expressed as multi-dimensional label expressions. Decision trees have a natural rule representation and the merging algorithm in LID3 allows for the generation of a wider range of descriptive rules than for mass relations. LDTs can also be used for query evaluation although they are slightly more limited than mass relations in this respect because of the difficulty of evaluating rules that are conditional on a constraint on the output attribute. © 2006 Springer Science+Business Media, Inc.