Semi-extractive Multi-document Summarization

In this thesis, I design a Maximum Coverage problem with KnaPsack constraint (MCKP) based model for extractive multi-document summarization. The model integrates three measures to detect important sentences including Coverage, rewards sentences in regards to their representative level of the whole document, Relevance, focuses to select sentences that related to the given query, and Compression, rewards concise sentences. To generate a summary, I apply an efficient and scalable greedy algorithm. The algorithm has a near optimal solution when its scoring functions are monotone non-decreasing and submodular. I use DUC 2007 dataset to evaluate our proposed method. Investigating the results using ROUGE package shows improvement over two closely related works. The experimental results illustrates that integrating compression in the MCKP-based model, applying semantic similarity measures to detect Relevance measure and also defining all scoring functions as a monotone submodular function result in having a better performance in generating a summary.