Abstract

The classification of a text into a category that corresponds with its content is an important task in Natural Language Processing. As we want to support shallow syntactic techniques, it must be possible to establish this category without a semantic text analysis. We therefore propose a statistical approach featuring a set of categories that cover orthogonal concepts: micro-classifiers. They allow for a robust multi-dimensional and multi-label text classification, which reveals to be beneficial in the context of automatic document summarization. The performance of these micro-classifiers is evaluated for four different cut-off thresholds using measures of precision and recall. Furthermore, we examine noun-phrase coreference chains within documents and attempt to find correlations with single and multi-label categorization. The presence of patterns could suggest better ways to enhance automatic document summarization.