Research Article
A Feature-Augmented Deep Learning Model for Extractive Summarization
@INPROCEEDINGS{10.1007/978-3-030-77424-0_23, author={Bui Thi Mai Anh and Nguyen Thi Thu Trang}, title={A Feature-Augmented Deep Learning Model for Extractive Summarization}, proceedings={Industrial Networks and Intelligent Systems. 7th EAI International Conference, INISCOM 2021, Hanoi, Vietnam, April 22-23, 2021, Proceedings}, proceedings_a={INISCOM}, year={2021}, month={5}, keywords={Extractive summarization Sequence to sequence model Recurrent neural networks}, doi={10.1007/978-3-030-77424-0_23} }- Bui Thi Mai Anh
Nguyen Thi Thu Trang
Year: 2021
A Feature-Augmented Deep Learning Model for Extractive Summarization
INISCOM
Springer
DOI: 10.1007/978-3-030-77424-0_23
Abstract
Extractive text summarization can be seen as a classification task in which sentences from the document are labelled with eitherin-summaryornot-in-summary. The most salient sentences (i.e., with highest ranking score) from the original document will be selected to generate the summary. Recent success of deep learning in the field of Natural Language Processing (NLP) has raised a trending research direction for text summarization task. Many neural models have been proposed in which applying recurrent neural network (rNN) for extractive summarization is also becoming increasingly popular. In this paper, we aim to improve the baseline sequence to sequence model proposed by Nallapati et al. by augmenting more sentence features so that the generated summary can benefit from potential features of the whole document. On one hand, the additional sentence-based features enrich the representation vector resulting from the sentence-level RNN of the baseline model. On the other hand, the relevant information from word-level will also be added to the final vector to increase the accuracy of the classification task. The experiment has been conducted for the DailyMail/CNN dataset to evaluate our proposed method and the state of the art works. The empirical results show that the proposed model with augmented features increases about 0.3-0.4 points of ROUGE-1 and ROUGE-2 in comparison with related works.
