A Concept-Based Integer Linear Programming Approach for Single-Document Summarization

Abstract

Automatic single-document summarization is a process that receives a single input document and outputs a condensed version with only the most relevant information. This paper proposes an unsupervised concept-based approach for singledocument summarization using Integer Linear Programming (ILP). Such an approach maximizes the coverage of the important concepts in the summary, avoiding redundancy, and taking into consideration some readability aspects of the generated summary as well. A new weighting method that combines both coverage and position of the sentences is proposed to estimate the importance of a concept. Moreover, a weighted distribution strategy that prioritizes sentences at the beginning of the document if they have relevant concepts is investigated. The readability of the generated summaries is improved by the inclusion of constraints into the ILP model to avoid dangling coreferences and breaks in the normal discourse flow of the document. Experimental results on the DUC 2001-2002 and the CNN corpora demonstrated that the proposed approach is competitive with state-of-the-art summarizers evaluated regarding the traditional ROUGE scores.