Extractive Summarisation Based on Keyword Profile and Language Model

Human Language Technologies: The 2015 Annual Conference of the North American Chapter of the ACL, pages 123–132, Denver, Colorado, May 31 – June 5, 2015. c⃝2015 Association for Computational Linguistics Extractive Summarisation Based on Keyword Profile and Language Model Han Xu Eric Martin Ashesh Mahidadia School of Computer Science and Engineering UNSW, Sydney, NSW, Australia, 2052 hanx,emartin,ashesh@cse.unsw.edu.au Abstract We present a statistical framework to extract information-rich citation sentences that sum- marise the main contributions of a scientific paper. In a first stage, we automatically dis- cover salient keywords from a paper’s citation summary, keywords that characterise its main contributions. In a second stage, exploiting the results of the first stage, we identify citation sentences that best capture the paper’s main contributions. Experimental results show that our approach using methods rooted in quan- titative statistics and information theory out- performs the current state-of-the-art systems in scientific paper summarisation. 1 Introduction and Motivation Science is not an isolated endeavour, but benefits from and expands on the work of others, with more or less cross fertilisation between disciplines. The interdependent nature of research has naturally re- sulted in a network of scientific areas with dense in- terconnections between related fields. Though re- search is a highly specialised activity, researchers find themselves constantly in need to explore the network further from the core of their research. Tools that can facilitate understanding the key con- tributions of papers in those parts of the network be- ing explored can only prove highly valuable. As an example of such tools, we focus on an application that automatically extracts information- rich sentences describing the main contributions of a given paper. From which corpus the extraction could take place? A natural answer is the abstract of the paper. However, the contributions as perceived by the authors can significantly deviate from those judged extrospectively by the community over time (Mei and Zhai, 2008). Instead, we take as corpus the set of citing sentences to the paper (from other papers). Indeed, those sentences can arguably be deemed as a form of crowd-sourced review of the paper’s main contributions. The set of citing sen- tences is referred to as the citation summary of the target paper. Elkiss et al. (2008) carried out a large- scale study and confirmed that citation summaries contain extra information that does not appear in pa- per abstracts. In addition, they found that the “self- cohesion”, measured as the average cosine similar- ity between sentences, is consistently higher in a pa- per’s citation summary than in its abstract: the for- mer is more focused than the latter in describing pa- pers’ main contributions. This work presents our ef- forts in advancing research along this direction. Section 2 formally defines the problem we aim to solve: summarise scientific papers using the most informative and diversified part of their cita- tion summaries. It surveys several prominent related studies, and introduces the data used in our experi- ments and evaluations. In Section 3, we present our statistical framework built upon quantitative …

Content truncated to 500 words