Abstract
News outlets are constantly publishing new articles to cover important events all around the globe. This vast amount of news articles gives rise to the topic of multidocument summarization in order to provide quick information about the reported events. Although text summarization in general is an old topic, recent advances have been made on abstractive summarization, which aims to produce a summary that is abstracted (or paraphrased) from the original articles, in contrast to the more traditional extractive summarization which generates summaries containing sentences which are exact copies from the original text. To this end, abstractive multidocument summarization (MDS) systems typically leverage a fine-grained extraction tool to obtain the fine detail of the reported event, and then fuse them together to generate a new sentence not present in the original text. Due to the divide-and-combine nature, abstractive summaries have the advantage of being more concise while maintaining the same level of informativeness, compared to extractive summaries. Furthermore, the fine-grained nature of the processed information provides more flexibility to abstractive MDS systems regarding which specific chunk of information it wants to include in the summary, giving rise to a flexible, personalized summary. The goal of this thesis is to study and improve the state-of-the-art abstractive MDS approaches. In particular, we first investigate the state-of-the-art fact extraction tools that are used by existing abstractive MDS approaches and propose an improvement in terms of the fine-graininess and the semantics of the extraction. Next, we study the fact fusion aspect of abstractive MDS approaches, and improve also their semantics in order to generate more new sentences to be included in the summary. Then, we combine our fine-grained extraction tool and fact fusion approach to build an abstractive MDS technique that can be personalized according to the level of details and the types of information. We also developed a demonstration system for the technique. Finally, our approaches have been evaluated showing that they outperform the baselines.