Class Label Enhancement via Related Instances

Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, pages 118–128, Edinburgh, Scotland, UK, July 27–31, 2011. c⃝2011 Association for Computational Linguistics Class Label Enhancement via Related Instances Zornitsa Kozareva USC Information Sciences Institute 4676 Admiralty Way Marina del Rey, CA 90292-6695 kozareva@isi.edu Konstantin Voevodski Boston University 111 Cummington St. Boston, MA, 02215 kvodski@bu.edu Shang-Hua Teng University of Southern California 941 Bloom Walk, SAL 300 Los Angeles, CA 90089 shanghua@usc.edu Abstract Class-instance label propagation algorithms have been successfully used to fuse informa- tion from multiple sources in order to enrich a set of unlabeled instances with class labels. Yet, nobody has explored the relationships be- tween the instances themselves to enhance an initial set of class-instance pairs. We pro- pose two graph-theoretic methods (centrality and regularization), which start with a small set of labeled class-instance pairs and use the instance-instance network to extend the class labels to all instances in the network. We carry out a comparative study with state-of-the-art knowledge harvesting algorithm and show that our approach can learn additional class labels while maintaining high accuracy. We conduct a comparative study between class-instance and instance-instance graphs used to propa- gate the class labels and show that the latter one achieves higher accuracy. 1 Introduction Many natural language processing applications use and rely on semantic knowledge resources. Since manually built lexical repositories such as Word- Net (Fellbaum, 1998) cover a limited amount of knowledge and are tedious to maintain over time, re- searchers have developed algorithms for automatic knowledge extraction from structured and unstruc- tured texts. There is a substantial body of work on extracting is-a relations (Etzioni et al., 2005; Kozareva et al., 2008), part-of relations (Girju et al., 2003; Pantel and Pennacchiotti, 2006) and general facts (Lin and Pantel, 2001; Davidov and Rappoport, 2009; Jain and Pantel, 2010). The usefulness of the generated resources has been shown to be valuable to information extraction (Riloff and Jones, 1999), question answering (Katz et al., 2003) and textual entailment (Zanzotto et al., 2006) systems. Among the most common knowledge acquisi- tion approaches are those based on lexical patterns (Hearst, 1992; Etzioni et al., 2005; Kozareva et al., 2008) and clustering (Lin and Pantel, 2002; Davidov and Rappoport, 2008). While clustering can find in- stances and classes that are not explicitly expressed in text, they often may not generate the granularity needed by the users. In contrast, pattern-based ap- proaches generate highly accurate lists, but they are constraint to the information matched by the pattern and often suffer from recall. (Pas¸ca, 2004; Snow et al., 2006; Kozareva and Hovy, 2010) have shown that complete lists of semantic classes and instances are valuable for the enrichment of existing resources like WordNet and for taxonomy induction. There- fore, researchers have focused on the development of methods that can automatically augment the ini- tially extracted class-instance pairs. (Pennacchiotti and Pantel, 2009) fused informa- tion from pattern-based and distributional systems using an ensemble method and a rich set of features derived from query …

Content truncated to 500 words