Multi-Document Summarization by Capturing the Information Users are Interested in

Proceedings of Recent Advances in Natural Language Processing, pages 77–83, Hissar, Bulgaria, 12-14 September 2011. Multi-Document Summarization by Capturing the Information Users are Interested in Elena Lloret University of Alicante Apdo. de Correos 99 E-03080, Alicante, Spain elloret@dlsi.ua.es Laura Plaza Universidad Complutense de Madrid C/Prof. Jos´e Garc´ıa Santesmases, s/n 28040 Madrid, Spain lplazam@fdi.ucm.es Ahmet Aker University of Sheffield 211 Portobello Sheffield, S1 4DP, UK a.aker@dcs.shef.ac.uk Abstract This paper proposes a method for automat- ically generating summaries taking into account the information in which users may be interested. Our approach relies on existing model summaries from tourist sites and captures from them the type of information humans use to describe places around the world. Relational patterns are first extracted and categorized by the type of information they encode. Then, we apply them to the collection of in- put documents to automatically extract the most relevant sentences and build the sum- maries. In order to evaluate the perfor- mance of our approach, we conduct two types of evaluation. On the one hand, we use ROUGE to assess the information con- tained in our summaries against existing human written summaries, whereas on the other hand, we carry out a human readabil- ity evaluation. Our results indicate that our approach achieves high performance both in ROUGE and manual evaluation. 1 Introduction The amount of information currently available is growing at an exponential rate. Information pre- sented in different formats (text, images, audio, video) needs to be carefully processed in order to allow users to manage it efficiently and effectively. Text summarization (TS) can provide many advan- tages to users, since TS systems are able to gener- ate a brief summary of one or several documents by selection and/or generalization of what is im- portant in the source (Sp¨arck Jones, 2007). However, TS is an especially challenging Nat- ural Language Processing (NLP) task, since the generation of summaries depends on a wide range of issues, such as the summarization input, out- put or purpose. In particular, the type of text or domain we deal with is of great importance in TS, since each domain has its particular fea- tures, and they need to be treated accordingly. For instance, when summarizing newswire text, the reader is mainly concerned about the who, what, when, where and why of the fact reported in the news item; when summarizing a research paper, the reader is mostly interested in the problem be- ing faced, the method proposed to solve it and the results achieved. Therefore, being capable of knowing what a user would like to read in a sum- mary will allow the summaries to be biased to- wards such information. The order in which this information is shown in the source documents is also important (Barzilay et al., 2002), and thus this same order should be kept in the summary. Con- tinuing with the newswire example, the informa- tion in news articles may be presented in chrono- logical order, in a cause-effect manner, etc., so that this logical order …

Content truncated to 500 words