CRHASum: extractive text summarization with contextualized-representation hierarchical-attention summarization network

• Published in: Neural computing & applications 2020-08, Vol.32 (15), p.11491-11503
• Main Authors: Diao, Yufeng, Lin, Hongfei, Yang, Liang, Fan, Xiaochao, Chu, Yonghe, Wu, Di, Zhang, Dongyu, Xu, Kan
• Format: Article
• Language: English
• Subjects: Artificial Intelligence; Computational Biology/Bioinformatics; Computational Science and Engineering; Computer Science; Context; Data Mining and Knowledge Discovery; Image Processing and Computer Vision; Neural networks; Original Article; Probability and Statistics in Computer Science; Regression; Representations; Sentences
• ISSN: 0941-0643; 1433-3058
• DOI: 10.1007/s00521-019-04638-3

The requirements for automatic document summarization that can be applied to practical applications are increasing rapidly. As a general sentence regression architecture, extractive text summarization captures sentences from a document by leveraging externally related information. However, existing sentence regression approaches have not employed features that mine the contextual information and relations among sentences. To alleviate this problem, we present a neural network model, namely the Contextualized-Representation Hierarchical-Attention Summarization (CRHASum), that uses the contextual information and relations among sentences to improve the sentence regression performance for extractive text summarization. This framework makes the most of their advantages. One advantage is that the contextual representation is allowed to vary across linguistic context information, and the other advantage is that the hierarchical attention mechanism is able to capture the contextual relations from the word-level and sentence-level by using the Bi-GRU. With this design, the CRHASum model is capable of paying attention to the important context in the surrounding context of a given sentence for extractive text summarization. We carry out extensive experiments on three benchmark datasets. CRHASum alone can achieve comparable performance to the state-of-the-art approach. Meanwhile, our method significantly outperforms the state-of-the-art baselines in terms of multiple ROUNG metrics and includes a few basic useful features.