JHU1 : An Unsupervised Approach to Person Name Disambiguation using Web Snippets

Proceedings of the 4th International Workshop on Semantic Evaluations (SemEval-2007), pages 199–202, Prague, June 2007. c⃝2007 Association for Computational Linguistics JHU1 : An Unsupervised Approach to Person Name Disambiguation using Web Snippets Delip Rao Nikesh Garera David Yarowsky Dept. of Computer Science Johns Hopkins University Baltimore, MD 21218 {delip, ngarera, yarowsky}@cs.jhu.edu Abstract This paper presents an approach to person name disambiguation using K-means clus- tering on rich-feature-enhanced document vectors, augmented with additional web- extracted snippets surrounding the polyse- mous names to facilitate term bridging. This yields a significant F-measure improvement on the shared task training data set. The pa- per also illustrates the significant divergence between the properties of the training and test data in this shared task, substantially skewing results. Our system optimized on F0.2 rather than F0.5 would have achieved top performance in the shared task. 1 Introduction Being able to automatically distinguish between John Doe, the musician, and John Doe, the actor, on the Web is a task of significant importance with ap- plications in IR and other information management tasks. Mann and Yarowsky (2004) used bigograph- ical data annotated with named entitities and per- form fusion of extracted information across multiple documents. Bekkerman and McCallum (2005) stud- ied the problem in a social network setting exploit- ing link topology to disambiguate namesakes. Al- Kamha and Embley (2004) used a combination of attributes (like zipcodes, state, etc.), links, and page similarity to derive the name clusters while Wan et. al. (2005) used lexical features and named entities. 2 Approaches Our framework focuses on the K-means clustering model using both bag of words as features and vari- ous augumented feature sets. We experimented with several similarity functions and chose Pearson’s cor- relation coefficient1 as the distance measure for clus- tering. The weights for the features were set to the term frequency of their respective words in the doc- ument.2 2.1 Submitted system: Clustering using Web Snippets We queried the Google search engine with the target person names and extracted up to the top one thousand results. For each result we also extracted the snippet associated with it. An example is shown below in Figure 2.1. As can be seen the Figure 1: Google snippet for “Dekang Lin” snippets contain high quality, low noise features that could be used to improve the performance of the system. Each snippet was treated as a document and 1This performs better than the standard measures like Eu- clidean and Cosine with K-means clustering on this data. 2We found that using TF weights instead of TF-IDF weights gives a better performance on this task. 199 clustered along with the supplied documents. This process is illustrated in Figure 2. The following example illustrates how these web snippets can improve performance by lexical transitivity. In this hypothetical example, a short test document contains a Canadian postal code (T6G 2H1) not found in any of the training documents. However, there may exist an additional web page not in the training or test data which contains both this …

Content truncated to 500 words