An Automatic Extraction of Key Paragraphs Based on Context Dependency

An Automatic Extraction of Key Paragraphs Based on Context Dependency Fumiyo Fukumoto Yoshimi Suzukit Jun'ichi Fukumoto:~ Dept. of Electrical Engineering and Kansai Lab., R & D Group Computer Science, Yamanashi University Oki Electric Industry Co. Ltd.,:~ 4-3-11 Takeda, Kofu 400 Japan 1-2-27 Shiromi, Chuo-ku, Osaka 540 Japan ~fukumoto~skye, ysuzuki~suwa~ }.esi.yamanashi.ac.jp fukumoto~kansai.oki.co.jp Abstract In this paper, we propose a method for ex- tracting key paragraphs in articles based on the degree of context dependency. Like Luhn's technique, our method assumes that the words related to theme in an arti- cle appear throughout paragraphs. Our ex- traction technique of keywords is based on the degree of context dependency that how strongly a word is related to a given con- text. The results of experiments demon- strate the applicability of our proposed method. 1 Introduction With increasing numbers of machine readable doc- uments becoming available, automatic document summarisation has become one of the major research topics in IR and NLP studies. In the field of an automatic summarisation, there are at least two approaches. One is knowledge-based approach with particular subject fields (Reimer, 1988), (Jacobs, 1990). This approach, based on deep knowledge of particular subject fields, is useful for restricted tasks, such as, for example, the construc- tion of 'weather forecasts' summaries. However, when unrestricted subject matter must be treated, as is often the case in practice, the passage retrieval and text summarisation methods proposed have not proven equal to the need, since deep knowledge of particular subject fields is required (Paice, 1990), (Zechner, 1996). The other, alternative strategy is the approach that relies mainly on corpus statistics (Paice, 1990), (Palce, 1993). The main task of this approach is the sentence scoring process. Typically, weights are assigned to the individual words in a text, and the complete sentence scores are then based on the occurrence characteristics of highly-weighted terms (keywords) in the respective sentences. Term weighting technique has been widely inves- tigated in information retrieval and lots of tech- niques such as location heuristics (Baxendale, 1958), rhetorical relations (Miike, 1994), and title informa- tion (Edmundson, 1969) have been proposed. These techniques seem to be less dependent on the domain. However, Salton claims that it is difficult to pro- duce high accuracy of retrieval by using these term- weighting approaches (Salton, 1993). The other term weighting technique is based on keyword frequency (Luhn, 1958). Keyword fre- quency is further less dependent on the domain than other weighting methods and therefore, well studied. Major approaches which are based on keyword fre- quency assume on the fact that the keywords of the article appear frequently in the article, but appear seldom in other articles (Luhn, 1958), (Nagao, 1976), (Salton, 1993), (Zechner, 1996). These approaches seem to show the effect in entirely different articles, such as 'weather forecasts', 'medical reports', and 'computer manuals'. Because each different article is characterised by a larger number of words which appear frequently in one article, but appear seldom in other articles. However, in some articles from the same domain such as …

Content truncated to 500 words