CU-COMSEM: Exploring Rich Features for Unsupervised Web Personal Name Disambiguation

Proceedings of the 4th International Workshop on Semantic Evaluations (SemEval-2007), pages 125–128, Prague, June 2007. c⃝2007 Association for Computational Linguistics CU-COMSEM: Exploring Rich Features for Unsupervised Web Per- sonal Name Disambiguation Ying Chen Center for Spoken Language Research University of Colorado at Boulder yc@colorado.edu James Martin Department of Computer Science University of Colorado at Boulder James.Martin@colorado.edu Abstract The increasing number of web sources is exacerbating the named-entity ambiguity problem. This paper explores the use of various token-based and phrase-based fea- tures in unsupervised clustering of web pages containing personal names. From these experiments, we find that the use of rich features can significantly improve the disambiguation performance for web per- sonal names. 1 Introduction As the sheer amount of web information expands at an ever more rapid pace, the named-entity am- biguity problem becomes more and more serious in many fields, such as information integration, cross-document co-reference, and question an- swering. Individuals are so glutted with informa- tion that searching for data presents real problems. It is therefore crucial to develop methodologies that can efficiently disambiguate the ambiguous names from any given set of data. In the paper, we present an approach that com- bines unsupervised clustering methods with rich feature extractions to automatically cluster re- turned web pages according to which named en- tity in reality the ambiguous personal name in a web page refers to. We make two contributions to approaches to web personal name disambiguation. First, we seek to go beyond the kind of bag-of- words features employed in earlier systems (Bagga & Baldwin, 1998; Gooi & Allan, 2004; Pedersen et al., 2005), and attempt to exploit deep semantic features beyond the work of Mann & Yarowsky (2003). Second, we exploit some fea- tures that are available only in a web corpus, such as URL information and related web pages. The paper is organized as follows. Section 2 in- troduces our rich feature extractions along with their corresponding similarity matrix learning. In Section 3, we analyze the performance of our sys- tem. Finally, we draw some conclusions. 2 Methodology Our approach follows a common architecture for named-entity disambiguation: the detection of ambiguous objects, feature extractions and their corresponding similarity matrix learning, and fi- nally clustering. Given a webpage, we first run a modified Beau- tiful Soup1 (a HTML parser) to extract a clean text document for that webpage. In a clean text docu- ment, noisy tokens, such as HTML tags and java codes, are removed as much as possible, and sen- tence segmentation is partially done by following the indications of some special HTML tags. For example, a sentence should finish when it meets a “<table>” tag. Then each clean document contin- ues to be preprocessed with MXTERMINATOR (a sentence segmenter), 2 the Penn Treebank to- kenization,3 a syntactic phrase chunker (Hacioglu, 2004), and a named-entity detection and co- reference system for the ACE project4 called EX- 1 http://www.crummy.com/software/BeautifulSoup 2http://www.id.cbs.dk/~dh/corpus/tools/MXTERMINATOR. html 3 http://www.cis.upenn.edu/~treebank/tokenization.html 4 http://www.nist.gov/speech/tests/ace 125 ERT5 (Hacioglu et al. 2005; Chen & Hacioglu, 2006). 2.1 The …

Content truncated to 500 words