The Impact of Supercategory Inclusion on Semantic Classifier Performance

Abstract

It is a known phenomenon that text document classifiers may benefit from inclusion of hypernyms of the terms in the document. However, this inclusion may be a mixed blessing because it may fuzzify the boundaries between document classes [5, 6, 10].

We have elaborated a new type of document classifiers, so called semantic classifiers, trained not on the original data but rather on the categories assigned to the document by our semantic categorizer [1, 4], that require significantly smaller corpus of training data and outperforms traditional classifiers used in the domain.

With this research we want to clarify what is the advantage/disadvantage of using supercategories of the assigned categories (an analogon of hypernyms) on the quality of classification. In particular we concluded that supercategories should be added with restricted weight, for otherwise they may deteriorate the classification performance. We found also that our technique of aggregating the categories counteracts the fuzzifying of class boundaries.