Rapid identification of content structures in a scientific paper is of great importance particularly for those who actively engage in frontier research. This paper presents a multi-classifier approach to identify such structures in terms of classification of rhetorical sentences in scientific papers. The idea behind this approach is based on an observation that no single classifier is the best performer for classifying all rhetorical categories of sentences. Therefore, our approach learns which classifiers are good at what categories, assign the classifiers for those categories and apply only the right classifier for classifying a given category. This paper employsk-fold cross validation over training data to obtain the category-classifier mapping and then re-learn the classification model of the corresponding classifier using full training data on that particular category. This approach has been evaluated for identifying sixteen different rhetorical categories on sentences collected from ACL-ARC paper collection. The experimental results show that the multi-classifier approach can significantly improve the classification performance over multi-label classifiers.
Article first published in “Journal of ICT Research and Applications”, V. 7 (2013), n. 3, pp. 235-249. Reprinted with permission.