Mediatory Summary Generation: Summary-Passage Extraction for Information Credibility on the Web

Mediatory Summary Generation: Summary-Passage Extraction for Information Credibility on the Web � Koichi Kaneko, Hideyuki Shibuki, Masahiro Nakano, Rintaro Miyazaki, Madoka Ishioroshi, and Tatsunori Mori Graduate School of Environment and Information Sciences, Yokohama National University 79-7 Tokiwadai, Hodogaya-ku,Yokohama 240-8501, Japan � kaneko, shib, nakano, rintaro, ishioroshi, mori � @forest.eis.ynu.ac.jp Abstract. In this paper, we discuss the summarization for supporting a user’s judgment on the credibility of information on the Web. In general, if a statement contradicts another statement, the credibility of either of the statements decreases. However, these opposing statements may coexist under certain situations, and presenting such situations is helpful for a user’s judgment. Therefore, we focus on the coexistence of these opposing statements, and attempt to develop a system to generate survey reports that contain mediatory summaries, which are defined as passages extracted from Web documents in order to present situations in which these opposing statements can coexist. We describe the outline of the summarization system and describe how to improve the TextRank algorithm from the viewpoint of passage extraction for the system. From experimental results, we confirmed that the methods based on the improved TextRank algorithm can extract significant passages, which are actually considered as significant by human assessors, with higher precision than baseline methods. Keywords: Mediatory Summary, Information Credibility, Passage Extraction, Text Summa- rization 1 Introduction Many pages on the Web contain incorrect or unverifiable information. Therefore, there is a grow- ing demand for technologies that enable us to obtain reliable information. However, it would be almost impossible to automatically judge the accuracy of information presented on the Web. In this case, the second-best approach is to develop a supporting method that judges the credibility of information on the Web. At present, when we would like to judge the credibility of information on the Web, we often read some relevant Web documents retrieved via Web search engines. However, if the content described in some of the documents conflicts with the content in some other documents, Web search engines do not give users any suggestions on how to interpret the conflict. Furthermore, the amount of retrieved documents is too large to read, and the documents may not be ranked in the order of the credibility of information. Therefore, information retrieval is not sufficient to support a user’s judgment on the credibility of information, and therefore, additional techniques are required. Such techniques for supporting a user’s judgment on the credibility of information are classified into three types. The first type involves the extraction of significant descriptions such as statements that may be relevant to a user and the sender of the statements. In this paper, a statement is defined as text such as an opinion, evaluation, or objective fact. The term sender refers to a person or an � This research is partially supported by National Institute of Information and Communications Technology, Japan and a research project of Graduate School of Environment and Information Sciences, Yokohama National University. Copyright 2009 by Koichi Kaneko, Hideyuki Shibuki, Masahiro Nakano, Rintaro …

Content truncated to 500 words