Satoshi Sato’s scientific contributions

This is a preliminary report of our study on automatic generation of web directories. We now concentrate on generation of web directories for specific categories such as aquarium, zoo, and museum. From only a category word (e.g., aquarium), our system generates the directory of the category's instances, in which instances are classified by geographical region, without any human interaction. 1 Introduction The World Wide Web (WWW) is now the popular resource when we look for certain information. Usually we start information hunting from keyword retrieval using a search engine, or from browse of a large web directory such as Yahoo. However, we sometimes cannot find the web page that contains the desired information, because current search engines and directories are not effective enough to navigate the unorganized and chaotic WWW space. A search engine with a large index outputs the list of too many web pages to select appropriate pages. Categorization of web directories that cover wh...

This paper describes two significant natural language processing methods that are used in the automated QA-Pack generation system: the summary-extraction method and the hierarchical text classification method. QA-Pack is the questions-answers package that is generated from the USENET articles; we designed QA-Pack as a substitution of Frequently Asked Questions (FAQs) of the USENET. FAQ is the effective information package that assists us in finding the specific answer in certain domain; FAQ provides us with organized information and the navigation mechanisms. One of the weak points of FAQs are that they do not include the latest information; FAQs' update cycles are usually long, because editing requires timeconsuming works. In contrast, QA-Pack can always provide the latest information that are organized, because the automated generation system updates it in short cycles. 1 Introduction As Internet grows rapidly, the avalanche of electronic information perplexes us. The USENET ne...

In automated editing, text summarization plays an important role when creating the digests of contents; the digest helps users find the desired information quickly. In our automated editing system that makes FAQ-like information pack- ages from the USENET articles, summaries are extracted from original articles. Extracted sum- maries are useful for selecting the desired infor- mation, however, they can be improved into more appropriate summaries to the package. In this paper, we discuss the limitations of the current summary-extraction method, and propose em- ployment of rewriting as a solution.

本稿では, fj. wantedのダイジェストの自動生成を実現する方法について述べる. その中心技術は, ニュース記事からのサマリ抽出法である. この方法は, 言わば「斜め読みを模擬した処理」であり, まず, 表層的な表現を手がかりとして, 42の特徴を抽出し, それらの特徴を用いて, 記事のサマリ (カテゴリとサマリ文) を抽出する. ブラインドデータに対する実験において, 本方法は, カテゴリ判定正解率81%, サマリ文抽出正解率76%という値を示した. 抽出されたサマリはカテゴリ毎に整理され, HTML形式のダイジェストとして出力される. このとき, 元の記事へのポインタは, ハイパーテキストのリンクとして埋め込まれる. 作成されたダイジェストは, WWWのクライアントプログラムによって読むことができる.

リサーチレポート（北陸先端科学技術大学院大学情報科学研究科） Total amount of electronic text is already enormous now and computers provide remarkably good technologies of text editing and storing. However, as for text retrieval, there is no efficient means if the desire to information is not exact enough for the keyword retrieval. A new technology that provides navigative function and helps people acquire desired information is required now. In this study, we have developed a method of automatic generation of digests for the NetNews. A digest is a collection of summaries of original texts. Headlines of a newspaper, a table of contents of a magazine and summary on a back cover of book are all digests and work as navigators. We propose digests are the most appropriate as the navigator for the NetNews with consideration of the fact that readers are guided by above digests almost every time they reach to those printed materials and retrieve the desired information without much effort. Generation of digests for printed materials require certain degree of skills and a lot of time because all summarization, edition, and categorization processes are executed mainly by manual works. By contrast, automatic generation of digests is possible for the NetNews because the NetNews is electronic text. The main characteristics of the NetNews is made good use by automatic generation of digests. We concentrate on making digesting system for a newsgroup, fj.meetings that holds only "meeting announcement" and "call for paper" articles in both Japanese and English. The digesting system has two modules: MekeSummary and EdigDigest. In MakeSummary, we utilized expression patterns and distinctive document styles to extract information. The introduction of utilization of document styles contributes to obtaining high performance on summary extraction tasks without deep analysis of texts. MakeSummary has been evaluated on pre-examined and blind data and the result has demonstrated that the system is powerful enough in ability to accurately extract summaries. Adding more expression patterns could raise the performance easily. We believe that the experience of digesting the NetNews articles can be transplanted to other electronic text information also.