From Information Extraction to Abstractive Summarization

With the vast amount of news articles being reported through different online news outlets and thousands of user comments about the news, it is time consuming for a reader to browse through all of them. Therefore, an automated summarization system for news contents is important. This topic falls into the category of multidocument summarization, which has been mainly solved via extractive methods where summary sentences are verbatim of the original text. However, human summaries typically contains abstraction of facts, which are extracted from the original text, and then combined into a coherent summary. To better human summaries, recently research on abstractive summarization has gained traction in recent years, in which information extraction played a big part in the process. The goal of this research is to study the state-of-the-art work on multi-document summarization leveraging information extraction, to understand the challenges, and to finally propose an improvement. We have developed an Open Information Extraction system that is tailored for the purpose of abstractive summarization. We have also developed an improvement over a state-of-the-art approach which utilizes a fusion graph algorithm, improved with entity disambiguation and verb linking. Our evaluation shows that we have a considerable improvement over the baseline method.