A Semi-Supervised Method to Learn and Construct Taxonomies Using the Web

Proceedings of the 2010 Conference on Empirical Methods in Natural Language Processing, pages 1110–1118, MIT, Massachusetts, USA, 9-11 October 2010. c⃝2010 Association for Computational Linguistics A Semi-Supervised Method to Learn and Construct Taxonomies using the Web Zornitsa Kozareva and Eduard Hovy USC Information Sciences Institute 4676 Admiralty Way Marina del Rey, CA 90292-6695 {kozareva,hovy}@isi.edu Abstract Although many algorithms have been devel- oped to harvest lexical resources, few organize the mined terms into taxonomies. We pro- pose (1) a semi-supervised algorithm that uses a root concept, a basic level concept, and re- cursive surface patterns to learn automatically from the Web hyponym-hypernym pairs sub- ordinated to the root; (2) a Web based concept positioning procedure to validate the learned pairs’ is-a relations; and (3) a graph algorithm that derives from scratch the integrated tax- onomy structure of all the terms. Comparing results with WordNet, we find that the algo- rithm misses some concepts and links, but also that it discovers many additional ones lacking in WordNet. We evaluate the taxonomization power of our method on reconstructing parts of the WordNet taxonomy. Experiments show that starting from scratch, the algorithm can reconstruct 62% of the WordNet taxonomy for the regions tested. 1 Introduction A variety of NLP tasks, including inference, tex- tual entailment (Glickman et al., 2005; Szpektor et al., 2008), and question answering (Moldovan et al., 1999), rely on semantic knowledge derived from term taxonomies and thesauri such as Word- Net. However, the coverage of WordNet is still lim- ited in many regions (even well-studied ones such as the concepts and instances below Animals and Peo- ple), as noted by researchers such as (Pennacchiotti and Pantel, 2006) and (Hovy et al., 2009) who per- form automated semantic class learning. This hap- pens because WordNet and most other existing tax- onomies are manually created, which makes them difficult to maintain in rapidly changing domains, and (in the face of taxonomic complexity) makes them hard to build with consistency. To surmount these problems, it would be advantageous to have an automatic procedure that can not only augment existing resources but can also produce taxonomies for existing and new domains and tasks starting from scratch. The main stages of automatic taxonomy induc- tion are term extraction and term organization. In recent years there has been a substantial amount of work on term extraction, including semantic class learning (Hearst, 1992; Riloff and Shepherd, 1997; Etzioni et al., 2005; Pasca, 2004; Kozareva et al., 2008), relation acquisition between entities (Girju et al., 2003; Pantel and Pennacchiotti, 2006; Davi- dov et al., 2007), and creation of concept lists (Katz and Lin, 2003). Various attempts have been made to learn the taxonomic organization of concepts (Wid- dows, 2003; Snow et al., 2006; Yang and Callan, 2009). Among the most common is to start with a good ontology and then to try to position the miss- ing concepts into it. (Snow et al., 2006) maximize the conditional probability of hyponym-hypernym relations given certain evidence, while (Yang and Callan, 2009) …

Content truncated to 500 words