Automatic Generation of Natural Language Summaries

Automatic text summarization has gained much interest in the last few years, since it could, at least in principle, make the process of information seeking in large document collections less tedious and time-consuming. Most existing summarization methods generate summaries by initially extracting the sentences that are most relevant to the user’s query from documents returned by an information retrieval engine. In this thesis, we present a new competitive sentence extraction method that assigns relevance scores to the sentences of the texts to be summarized. Coupled with a simple method to avoid selecting redundant sentences, the resulting summarization system achieves state-of-the-art results on widely used benchmark datasets. Moreover, we propose two novel sentence compression methods, which rewrite a source sentence in a shorter form, retaining the most important information. The first method produces extractive compressions, i.e., it only deletes words, whereas the second one produces abstractive compressions, i.e., it also uses paraphrasing. Experiments show that the extractive method generates compressions better or comparable, in terms of grammaticality and meaning preservation, to those produced by state-of-theart systems. On the other hand, the abstractive method produces more varied (due to paraphrasing) and slightly shorter compressions than the extractive one. In terms of grammaticality and meaning preservation, the two methods have similar scores. Finally, we propose an optimization model that generates summaries by jointly se-