An Integrated Approach to Heterogeneous Data for Information Extraction

An Integrated Approach to Heterogeneous Data for Information Extraction Ying Chen, Sophia Y. M. Lee, and Chu-Ren Huang Department of Chinese & Bilingual Studies, The Hong Kong Polytechnic University {chenying3176, sophiaym, churenhuang}@gmail.com Abstract. The paper proposes an integrated framework for web personal information extraction, such as biographical information and occupation, and those kinds of information are necessary to further construct a social network (a kind of semantic web) for a person. As web data is heterogeneous in nature, most of IE systems, regardless of named entity recognition (NER) or relation detection and recognition (RDR) systems, fail to get reliably robust results. We propose a flexible framework, which can effectively complement state- of-the-art statistical IE systems with rule-based IE systems for web data, and achieves substantial improvement over other existing systems. In particular, in our current experiment, both the rule-based IE system, which is designed according to some web specific expression patterns, and the statistical IE systems, which are developed for some homogeneous corpora, are sensitive only to specific information types. Hence we argue that our system performance can be incrementally improved when new and effective IE systems are added into our framework. Keywords: relation extraction, information extraction 1 Introduction Semantic web, which collects and formats different kinds of web knowledge, plays an important role in the development of a new generation of web. One important component of semantic web is to automatically extract different relations existing in web data. Information extraction (IE) can provide such a technology to solve this problem, particularly for a specific named entity. For example, Web People Search1 (WePS) 2009 evaluation (Sekine & Artiles, 2009) tries to extract some personal information, and TREC Entity Track2 plans to find some related information for a product. Web IE is particularly challenging because web data is heterogeneous in nature. Complete or comprehensive IE information necessarily come from many different sources with different formats. For example, “affiliation” and “email” are so different in their own expressions so that they need different extraction approaches. Hence, a homogeneous IE model, regardless of statistical model or rule-based model, often cannot perform effectively for web IE. To overcome this problem, some previous systems (Culotta et al., 2004; Lan et al., 2009; Watanabe et al., 2009) have tried to combine different IE approaches, which are often homogeneous IE, to extract different types of information in web data. Nevertheless, few of them have explored how to effectively utilize or integrate different IE tools for web data. Copyright 2009 by Ying Chen, Sophia Y. M. Lee, and Chu-Ren Huang 1 http://nlp.uned.es/weps/ 2 http://ilps.science.uva.nl/trec-entity/ 82 23rd Pacific Asia Conference on Language, Information and Computation, pages 82–91 In this paper, we propose a framework to integrate heterogeneous IE approaches for web IE. In this framework, we first segment web data according to the expression format. Similar to the genre categories – “formal text” and “informal text” – which were defined in Minkov et al. (2005), a text is either in “formal style” or “informal style.” A “formal style” text …

Content truncated to 500 words