Finding Relevant Concepts for Unknown Terms Using a Web-based Approach

Finding Relevant Concepts for Unknown Terms Using a Web-based Approach Chen-Ming Hung1 and Lee-Feng Chien1, 2 1. Institute of Information Science, Academia Sinica 2. Dept. of Information Management, National Taiwan University Taipei, Taiwan rglly@iis.sinica.edu.tw and lfchien@iis.sinica.edu.tw Abstract. Previous research on automatic thesaurus construction most focused on extracting relevant terms for each term of concern from a small-scale and domain-specific corpus. This study emphasizes on utilizing the Web as the rich and dynamic corpus source for term association estimation. In addition to extracting relevant terms, we are interested in finding concept-level information for each term of concern. For a single term, our idea is that to send it into Web search engines to retrieve its relevant documents and we propose a Greedy-EM- based document clustering algorithm to cluster them and determine an appropriate number of relevant concepts for the term. Then the keywords with the highest weighted log likelihood ratio in each cluster are treated as the label(s) of the associated concept cluster for the term of concern. With some initial experiments, the proposed approach has been shown its potential in finding relevant concepts for unknown terms. 1. INTRODUCTION It has been well recognized that a thesaurus is crucial for representing vocabulary knowledge and helping users to reformulate queries in information retrieval systems. One of the important functions of a thesaurus is to provide the information of term associations for information retrieval systems. Previous research on automatic thesaurus construction most focused on extracting relevant terms for each term of concern from a small-scale and domain-specific corpus. In this study, there are several differences extended from the previous research. First, this study emphasizes on utilizing the Web as the rich and dynamic corpus source for term association estimation. Second, the thesaurus to be constructed has no domain limitation and is pursued to be able to benefit Web information retrieval, e.g. to help users disambiguate their search interests, when users gave poor or short queries. Third, in addition to extracting relevant terms, in this study we are interested in finding concept-level information for each term of concern. For example, for a term “National Taiwan University” given by a user, it might contain some different but relevant concepts from users’ point of view, such as “main page of National Taiwan University”, “entrance examination of NTU”, “the Hospital of NTU”, etc. The purpose of this paper is, therefore, to develop an efficient approach to deal with the above problem. In information retrieval researching area, extracting concepts contained in one text always plays a key role. However, in traditional way, if the text is too short, it is almost impossible to get enough information to extract the contained concepts. In this paper, utilizing the abundant corpora on the World Wide Web, we attempt to find the concepts contained in arbitrary length of topic-specific texts, even only a single term. For a single term, our idea is that to send it into Web search engines to retrieve its relevant documents, and a Greedy-EM- based document clustering algorithm …

Content truncated to 500 words