Straight to the point: Discovering themes for summary generation

Straight to the Point: Discovering Themes for Summary Generation Stephen Wan Centre for Language Technology Macquarie University Sydney, Australia swan@ics.mq.edu.au Robert Dale Centre for Language Technology Macquarie University Sydney, Australia rdale@ics.mq.edu.au Mark Dras Centre for Language Technology Macquarie University Sydney, Australia madras@ics.mq.edu.au Cécile Paris CSIRO Mathematical and Information Sciences Locked Bag 17 North Ryde 1670 Sydney, Australia Cecile.Paris@csiro.au Abstract This paper1 presents our approach to the problem of single sentence summarisation. We investigate the use of Singular Value Decomposition (SVD) to guide the generation of a summary towards the theme that is the focus of the document to be summarised. In doing so, the intuition is that the generated summary will more accurately reflect the content of the source document. Currently, we operate in the news domain and at present, our summaries are modelled on headlines. This paper presents SVD as an alternative method to determine if a word is a suitable candidate for inclusion in the headline. The results of a recall based evaluation comparing three different strategies to word selection, indicate that thematic information does help improve recall. 1 Introduction In the midst of a plethora of archived electronic documents, the successful completion of a research task is affected by the ease with which users can quickly identify the relevant documents that satisfy their information needs. To do so, a researcher often relies on generated summaries that reflect the contents of the original document. ������������������������������������������������� ������������������������������������������������������ ������������������������������������������ !� �����"����������������"��������������#� We explore the problem of generating a single sentence summary in the context of single document summarisation. Instead of identifying and extracting the most important sentence, we generate a new sentence from scratch. Thus, the resulting sentence summary may not occur verbatim in the source. As a precursor to single sentence summarisation, we first explore the particular case of headline generation in the news domain, specifically English news. Our system re-uses words from the news article to generate a single sentence summary that resembles a headline. This is done by iteratively selecting a word from the source article and then appending it to previously selected words. This approach has been explored by a number of researchers (eg. see Witbrock and Mittal, 1999; Jin and Hauptmann, 2002). In existing approaches, a word is selected on the basis of two criteria: how well it acts as a summary word, and how grammatical it will be given the preceding summary words that have already been chosen. Our approach uses Singular Value Decomposition (SVD) in the first criterion, as a means of determining if a word is a good candidate for inclusion in the headline. In this paper, we present an overview of our basic summarisation algorithm in Section 2. Section 3 examines limitations of the basic algorithm, illustrating how words can be used out of context, resulting in factually incorrect statements. This is the motivation of our SVD extension which is introduced conceptually in Section 4. Section 5 describes how we generate sentence summaries using SVD. In Section 6, we present our …

Content truncated to 500 words