Investigating Semantic Knowledge for Text Learning

Recent work has made much of using semantic knowledge, derived in particular from domain ontologies, for improving text learning tasks. Semantic knowledge is assumed to capture more in-depth knowledge of the text domain in comparison with conventional statistics-based methods that can only rely on more surface vocabulary-speci c characteristics of a data set. Therefore, using semantic knowledge instead of statistics-based methods should improve performance in text learning tasks signi cantly. We believe that this claim needs careful scrutiny and examine the validity of this assumption in this paper. We explore the usefulness of ontologies for a text classi cation task and the use of feature selection methods to extract terms that can function as candidate ontological concepts for building or extending ontologies. We point to a number of issues that arise when trying to use semantic knowledge for text classi cation. One particularly troublesome issue is that semantic knowledge encoded in ontologies simply may not correspond to the concepts and terms signi cant for text classi cation