Cluster Labelling based on Concepts in a Machine-Readable Dictionary

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 1371–1375, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP Cluster Labeling based on Concepts in a Machine-Readable Dictionary Fumiyo Fukumoto Interdisciplinary Graduate School of Medicine and Engineering Univ. of Yamanashi fukumoto@yamanashi.ac.jp Yoshimi Suzuki Interdisciplinary Graduate School of Medicine and Engineering Univ. of Yamanashi ysuzuki@yamanashi.ac.jp Abstract This paper addresses the issue of cluster labeling and presents a method for assign- ing labels by using concepts in a machine- readable dictionary. We assume that salient terms in the cluster content have the same hypernym because hypernymic se- mantic relation represents a generalization that goes from specific to generic. Our ex- perimental results reveal that hypernymic semantic relations can be exploited to in- crease labeling accuracy, as the results of 0.441 F-score improves over the two base- lines. 1 Introduction With the exponential growth of information on the Internet, finding and organizing relevant materials on the Internet is becoming increasingly difficult. Internet directories such as Yahoo! and Google, which classify Web pages into pre-defined hier- archical categories, provide one solution to the problem. Categories in the hierarchical struc- tures are carefully defined by human experts and documents are well-organized. However, manual category-tagging is extremely costly. Moreover, categories on some Internet is often insufficient in finding relevant documents for users. Because these categories tend to have some bias in both defining and classifying documents. Cluster label- ing is one of the techniques to attack the problem. Most of the work on cluster labeling identi- fies salient terms in the cluster content that char- acterize the cluster in contrast to other clusters. Salient terms are extracted by using statistical fea- ture selection, e.g., maximum sum of the indi- vidual term frequencies of documents assigned to a cluster (Cutting et al., 1992), an adapted ver- sions of Information Gain (Geraci et al., 2007), χ2 method (Popescul and Ungar, 2000), and the Jensen-Shannon Divergence (Camel et al., 2009). Other works based on salient terms are frequent phrases by (Osinski and Weiss, 2005), and inte- gration of hierarchical information by (Muhr et al., 2010). However, the suggested terms, even when related to each other, tend to represent different as- pects of the topic underlying the cluster, and it is often the case that a good label does not occur di- rectly in the document. Carmel et. al addressed the issue and presented a method to use Wikipedia as an external knowledge. They showed the effec- tiveness of the method. However, Wikipedia is the free online encyclopedia, and everyone can access and edit the information. Therefore, it is often in- cluded noise information such as categories which do not characterize the cluster in the pages. Chin et. al presented a method to use WordNet (Chin et al., 2006). They used machine learning through extending the given term set with synonyms, hy- pernyms, hyponyms and so on. However, their method needs training data to determine the actual weights. Through supervised training in the label- ing process …

Content truncated to 500 words