Towards Coherent Single-Document Summarization: An Integer Linear Programming-based Approach

Automatic Text Summarization (ATS) is a viable option to reduce the content of textual documents, e.g., as a possible preprocessing step in many text mining applications. Single-document extractive summarizers have been developed based on different approaches, but many of them have the drawback of producing summaries with low coherence among the selected sentences in the generated summaries. In this paper, we present an unsupervised summarization system as an attempt towards coherent extractive single-document summarization. This system relies on Integer Linear Programming (ILP) as an optimization technique for selecting the smallest subset of sentences of a document maximizing the coverage of relevant concepts. Furthermore, our solution uses a graph-based algorithm for two goals: representing both sentences and concepts and enabling local coherence scoring among the sentences in the generated summaries. The proposed system is evaluated on two single-document benchmark datasets (DUC 2001-2002) using ROUGE measures, and compared with other state-of-the-art summarizers. The achieved results are very competitive.