EXTRACTIVE SUMMARIZATION USING SENTENCE EMBEDDINGS: Automatic summarization of news articles at Blendle

Abstract

In this thesis, we investigate embedding-based extractive summarization techniques, in order to automatically summarize news articles at Blendle. The thesis is comprised of three studies. In the first two studies, we compare existing methods and explore the added value of substituting sentence embeddings. In the third study, we propose a summarization method based on a recurrent neural network (RNN) architecture. This model is an adaptation of the model by Cheng and Lapata (2016). In order to make the RNN training more flexible, we further propose a semi-supervised training framework for this RNN architecture by using unsupervised methods for pre- or co-training the RNN summarizer.