Summarizing News Articles using Pointer-Generator Networks
NLP | Spring 2018 | Prof. Mausam
Simple sequence-to-sequence models for abstractive summarization suffer from repetition and quite often produce factually incorrect details. To improve upon this, pointer-generator networks with coverage have been proposed. However, these models result in being mostly extractive and very rarely produce novel n-grams (phrases). A modification to the beam search scoring function has been proposed very recently to rectify this. We propose an improved beam search scoring function which results in better abstractive summaries.
Links
Collaborators