Extractive Summarization by Aggregating Multiple Similarities

Proceedings of Recent Advances in Natural Language Processing, pages 451–457, Hissar, Bulgaria, Sep 7–9 2015. Extractive Summarization by Aggregating Multiple Similarities Olof Mogren, Mikael K˚ageb¨ack, Devdatt Dubhashi Department of Computer Science and Engineering Chalmers University of Technology, 412 96 G¨oteborg, Sweden mogren@chalmers.se Abstract News reports, social media streams, blogs, digitized archives and books are part of a plethora of reading sources that people face every day. This raises the question of how to best generate automatic sum- maries. Many existing methods for ex- tracting summaries rely on comparing the similarity of two sentences in some way. We present new ways of measuring this similarity, based on sentiment analysis and continuous vector space representations, and show that combining these together with similarity measures from existing methods, helps to create better summaries. The finding is demonstrated with MULT- SUM, a novel summarization method that uses ideas from kernel methods to com- bine sentence similarity measures. Sub- modular optimization is then used to pro- duce summaries that take several differ- ent similarity measures into account. Our method improves over the state-of-the-art on standard benchmark datasets; it is also fast and scale to large document collec- tions, and the results are statistically sig- nificant. 1 Introduction Extractive summarization, the process of select- ing a subset of sentences from a set of documents, is an important component of modern informa- tion retrieval systems (Baeza-Yates et al., 1999). A good summarization system needs to balance two complementary aspects: finding a summary that captures all the important topics of the docu- ments (coverage), yet does not contain too many similar sentences (non-redundancy). It follows that it is essential to have a good way of measur- ing the similarity of sentences, in no way a trivial task. Consequently, several measures for sentence similarity have been explored for extractive sum- marization. In this work, two sets of novel similarity mea- sures capturing deeper semantic features are pre- sented, and evaluated in combination with ex- isting methods of measuring sentence similarity. The new methods are based on sentiment analy- sis, and continuous vector space representations of phrases, respectively. We show that summary quality is improved by combining multiple similarities at the same time using kernel techniques. This is demonstrated us- ing MULTSUM, an ensemble-approach to generic extractive multi-document summarization based on the existing, state-of-the-art method of Lin and Bilmes (2011). Our method obtains state-of- the-art results that are statistically significant on the de-facto standard benchmark dataset DUC 04. The experimental evaluation also confirm that the method generalizes well to other datasets. 2 MULTSUM MULTSUM, our approach for extractive sum- marization, finds representative summaries tak- ing multiple sentence similarity measures into ac- count. As Lin and Bilmes (2011), we formulate the problem as the optimization of monotone non- decreasing submodular set functions. This results in a fast, greedy optimization step that provides a (1 − 1 e) factor approximation. In the original version, the optimization objective is a function scoring a candidate summary by coverage and di- versity, expressed using cosine …

Content truncated to 500 words