Automatic text summarization using a filter-based approach

Abstract

[eng] Recent advances in deep neural networks (DNNs) have revolutionized the field of natural language processing (NLP) with promising results in automatic summarization of short texts. However, automatic text summarization of long texts remains challenging, especially when multiple sub-topics are present in the text. In this work, we present QuBART1 : a coupled DNN architecture which allows automatic summarization of one or various topics of a text selected by the user. This architecture is a two stage DNN. The first stage consists in extracting a subset of the input text based on keywords introduced by the user. This extraction is based on the distances of each sentence with respect to the user input in the latent space of a large language model. The second stage is an abstractive summarization of the previous extraction obtaining the final output, the summary. This approach generates state of the art results while giving the user the ability of controlling the desired output.