A talk by John Henderson
October 16th, 1997. Shaffer 3.
The machine learning community provides many strategies for automatically constructing compact representations of categorical concepts (classifiers) from large repositories of classified data. The natural language processing community has discovered that systems in which linguistic bias is automatically constructed from data are often more accurate and definitely more scalable than systems in which it is encoded manually. A recent computational learning theory result [Freund & Shapire, '95], which has proven successful in practice, shows how to create a diverse and accurate ensemble of classifiers one at a time. This talk will describe that result, and show some methods for using it in a few natural language processing tasks. I will not be presenting results, just descriptions and proposals attempting to induce discussion.