Combining labeled and unlabeled data for learning cross-document structural relationships

Abstract

Multi-document discourse analysis has emerged with the potential of improving various NLP applications. Based on the newly proposed Cross-document Structure Theory (CST), this paper describes an empirical study that classifies CST relationships between sentence pairs extracted from topically related documents, exploiting both labeled and unlabeled data. We investigate a binary classifier for determining existence of structural relationships and a full classifier using the full taxonomy of relationships. We show that in both cases the exploitation of unlabeled data helps improve the performance of learned classifiers. © Springer-Verlag Berlin Heidelberg 2005.