Query-Based Abstractive Summarization Using Neural Networks

Abstract

Creating short summaries of documents with respect to a query has applications in for example search engines, where it may help inform users of the most relevant results. Constructing such a summary automatically, with the potential expressiveness of a human-written summary, is a difficult problem yet to be fully solved. In this thesis, a neural network model for this task is presented. We adapt an existing dataset of news article summaries for the task and train a pointer-generator model using this dataset to summarize such articles. The generated summaries are then evaluated by measuring similarity to reference summaries. We observe that the generated summaries exhibit abstractive properties, but also that they have issues, such as rarely being truthful. However, we show that a neural network summarization model, similar to existing neural network models for abstractive summarization, can be constructed to make use of queries for more targeted summaries.