An Abstract Generation System for Social Scientific Papers

An Abstract Generation System for Social Scientific Papers Michio Kaneko Dongli Han Graduate School of Integrated Basic Sciences Department of Information Science, Nihon University, Tokyo, JAPAN College of Humanities and Sciences m-kaneko@chs.nihon-u.ac.jp Nihon University, Tokyo, JAPAN han@chs.nihon-u.ac.jp Abstract Abstracts are quite useful when one is trying to understand the content of a paper, or conducting a survey with a large number of scientific documents. The situation is even clearer for the domain of social science, as most papers are very long and some of them don’t even have any abstracts at all. In this work, we narrow our attention down to the social scientific papers and try to generate their abstracts automatically. Specifically, we put weight on three points: important keywords, readability as an abstract, and features of social scientific papers. Experimental results show the effectiveness of our method, whereas some problems remain and will need to be solved in the future. 1 Introduction Abstracts are expected to help readers who are trying to understand the outline of a paper, or conducting a survey with a large number of scientific documents. The situation is even clearer for the domain of social science, as most papers in this area tend to be very long and some of them don’t even have any abstracts at all. There have been many methods proposed for Japanese summarization (e.g., Ochitani et al. 1997; Hatakeyama et al., 2002; Mikami et al., 1999; Ohtake et al., 1999; Hatayama et al., 2002; Tomita et al., 2009; Fukushima et al. 2011). However, most existing proposals are made towards general text summarization instead of abstract generation for scientific papers. Here, it is important to distinguish between a summary and an abstract. According to a Japanese dictionary, an abstract contains the most important stuffs or the important matter that has been stated in a document, and a summary is a short text transformed from a long text containing all the important points in the original text (Umesao et al., 1995). With the difference between summaries and abstracts in mind, we attempt to propose a new method to generate abstracts for social scientific papers in this paper. Specifically, we put weight on three points: important keywords, readability as an abstract, and features of social scientific papers. In this paper, we first describe our proposal in Section 2, 3, 4 and 5. Specifically, Section 2 gives a brief introduction on the necessary language resources for the development of the subsequent modules. Section 3, 4 and 5 describe the sentence processing, importance degree estimation, and abstract generation respectively. Finally, we discuss some experiments conducted to evaluate the effectiveness of our approach in Section 6. 2 Necessary Language Resources In this work, in order to perform textual analysis and importance degree estimation for words or phrases, we create the following five lexicon-files beforehand. <Adverb Lexicon>: created from (Nitta, 2002) containing adverbs describing degrees (like emphasis). <Sentence-End Expression Lexicon>: extracted from (Morita and Matuki, 1989) containing all expressions functioning similarly to auxiliary verbs in Japanese. <Conjunction …

Content truncated to 500 words