SECNet: Unsupervised Text Summarization Using Salience Extractor with a Consistent Network

• Main Authors: Hong-JunYan, 顏宏峻
• Other Authors: Jen-Wei Huang
• Format: Others
• Language: en_US
• Published: 2019
• Online Access: http://ndltd.ncl.edu.tw/handle/3n6548

碩士 === 國立成功大學 === 電腦與通信工程研究所 === 107 === Automated extractive document summarization is an important aspect of natural language processing. Most existing unsupervised summarization models use a graph-based ranking algorithm to evaluate the salience of sentences based on similarity. However, similarity captures only the surface relationships between sentences. In this study, we addressed this issue by exploiting recent developments in the field of unsupervised image generation and attention-based models for image captioning. Our model includes two Seq2Seq models, a salience extractor, and a consistent network. The salience extractor model derives information from documents using an encoder, and generates latent summaries using a decoder. The attention score of latent summaries is used to calculate the importance of each sentence in the document. The use of a consistent network ensures that the latent summaries contain the necessary document information. The proposed model was evaluated using the English summarization benchmark datasets, DUC 2001 and DUC 2002, as well as the large Chinese summarization dataset, LCSTS. We also created a finance news dataset in Chinese and had a group of experts in bank auditing label a summary for use as a reference.