Multi-document Summarization Techniques for News Summaries

Automated text summarization has various commercial and educational applications. This research is focused on multi-document summarization of news reports – a task, in which multiple reports on the same event need to be summarized in one concise and coherent text. We discussed major challenges of this task and made a review of main techniques used in this field. We have also proposed a hybrid model, in which we combine extractive and abstractive techniques to tackle redundancy and keep generated summaries diverse. We made experiments with a number of popular machine learning models and compared their performance on two datasets, consisting of news reports and their human-written summaries. Our findings suggest that general NLP models do not always do well with multi-document summarization and it is important to develop techniques that address problems, specific to this task.