Topic session on discourse

Wendy Lehnert (UMass) presented a view of the coreference problem from the perspective of machine learning, i n particular, the idea of treating coreference as classification . To pursue this view, the overall coreference proble m needs to be broken up into sub-problems of the appropriate granularity that are amenable to machine-learnin g paradigms. Once a problem is identified, some of the issues that arise are : finding the right set of features to use in the classifier's training vectors ; determining which features are domain dependent and which are domain independent ; and fording the right size for the feature set, as more features require more training data . Dr. Lehnert presented an example of this approach for learning rules which recognize appositive constructions. UMass created a training interface where all possible appositive constructions are highlighted for a human-in-the-loop, who responds yes/no to each proposal . Some of the domain-independent features used in the training were syntactic in nature ("does the NP start with an indefinite article?"), and others were domain dependent ("does it contain a corporate designator?") .