Enhancing Cross Document Coreference of Web Documents with Context Similarity and Very Large Scale Text Categorization

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 483–491, Beijing, August 2010 Enhancing Cross Document Coreference of Web Documents with Context Similarity and Very Large Scale Text Categorization Jian Huang Information Sciences and Technology Pennsylvania State University jhuang@ist.psu.edu Pucktada Treeratpituk Information Sciences and Technology Pennsylvania State University pxt162@ist.psu.edu Sarah M. Taylor Lockheed Martin IS&GS sarah.m.taylor@lmco.com C. Lee Giles Information Sciences and Technology Pennsylvania State University giles@ist.psu.edu Abstract Cross Document Coreference (CDC) is the task of constructing the coreference chain for mentions of a person across a set of documents. This work offers a holistic view of using document-level categories, sub-document level context and extracted entities and relations for the CDC task. We train a categorization component with an efficient flat algorithm using thousands of ODP categories and over a million web documents. We propose to use ranked cat- egories as coreference information, partic- ularly suitable for web documents that are widely different in style and content. An ensemble composite coreference function, amenable to inactive features, combines these three levels of evidence for disam- biguation. A thorough feature importance study is conducted to analyze how these three components contribute to the coreference results. The overall solution is evaluated using the WePS benchmark data and demonstrate superior performance. 1 Introduction Cross Document Coreference (CDC) is the task to determine whether Named Entities (NE) from different documents refer to the same underlying identity. CDC enables a range of advanced NLP applications such as automated text summariza- tion and question answering (e.g. list-type ques- tions). CDC has mainly been developed from two perspectives. First, in the Message Understanding Confer- ence (MUC-6), CDC was viewed as an advanced task performed based on a set of Information Extraction (IE) artifacts. IE has been one of the central topics in NLP since the 1970s and gained much success in transforming natural language text to structured text. IE on the Web, however, is inherently very challenging. For one, the Web is comprised of such heterogenous content that IE systems, many of which are developed on tidy and domain-specific corpora, may achieve relatively limited coverage. Also, the content of web documents may not even be in the natural language form. Hence, though IE based features are quite precise, it is rather difficult to achieve good coverage that’s necessary to disambiguate person entities on the Web. Recently, there is significant research interest in a related task called Web Person Search (WePS) (Artiles et al., 2007), which seeks to determine whether two documents refer to the same person given a person name search query. Many systems employed the simple vector space model and word co-occurrence features for this task. Though more robust with better coverage, these methods are more susceptible to irrelevant words with regard to the entity of interest. Rather than relying solely on IE based or word co-occurrence features, this work adopts a holistic view of the different types of features useful for cross document coreference. Specifically, the main features of our proposed CDC approach …

Content truncated to 500 words