Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A New Sentence Extraction Strategy for Unsupervised Extractive Summarization Methods

Dehao Tao, Yingzhu Xiong, Zhongliang Yang, Yongfeng Huang

TL;DR

This paper addresses unsupervised extractive summarization by framing it within Information Theory and proposing a uniform graph-based framework. A novel sentence extraction strategy recomputes sentence importances after each selection to reduce mutual information and diversify features. Experiments on CNN/DailyMail, NYT, TTNews, and CLTS with TextRank and PACSUM show ROUGE improvements and more evenly distributed extraction positions, highlighting a trade-off between mutual-information reduction and entropy. The approach is computationally lightweight (roughly $O(n^2)$) and can complement existing methods, with future work aimed at automatic determination of the summary length $k$ via average entropy $\bar{H}$.

Abstract

In recent years, text summarization methods have attracted much attention again thanks to the researches on neural network models. Most of the current text summarization methods based on neural network models are supervised methods which need large-scale datasets. However, large-scale datasets are difficult to obtain in practical applications. In this paper, we model the task of extractive text summarization methods from the perspective of Information Theory, and then describe the unsupervised extractive methods with a uniform framework. To improve the feature distribution and to decrease the mutual information of summarization sentences, we propose a new sentence extraction strategy which can be applied to existing unsupervised extractive methods. Experiments are carried out on different datasets, and results show that our strategy is indeed effective and in line with expectations.

A New Sentence Extraction Strategy for Unsupervised Extractive Summarization Methods

TL;DR

This paper addresses unsupervised extractive summarization by framing it within Information Theory and proposing a uniform graph-based framework. A novel sentence extraction strategy recomputes sentence importances after each selection to reduce mutual information and diversify features. Experiments on CNN/DailyMail, NYT, TTNews, and CLTS with TextRank and PACSUM show ROUGE improvements and more evenly distributed extraction positions, highlighting a trade-off between mutual-information reduction and entropy. The approach is computationally lightweight (roughly ) and can complement existing methods, with future work aimed at automatic determination of the summary length via average entropy .

Abstract

In recent years, text summarization methods have attracted much attention again thanks to the researches on neural network models. Most of the current text summarization methods based on neural network models are supervised methods which need large-scale datasets. However, large-scale datasets are difficult to obtain in practical applications. In this paper, we model the task of extractive text summarization methods from the perspective of Information Theory, and then describe the unsupervised extractive methods with a uniform framework. To improve the feature distribution and to decrease the mutual information of summarization sentences, we propose a new sentence extraction strategy which can be applied to existing unsupervised extractive methods. Experiments are carried out on different datasets, and results show that our strategy is indeed effective and in line with expectations.
Paper Structure (9 sections, 9 equations, 4 figures, 2 tables)

This paper contains 9 sections, 9 equations, 4 figures, 2 tables.

Table of Contents

  1. Introduction
  2. The proposed method
  3. Task Modeling
  4. Decrease Mutual Information at Sentence Extraction Stage
  5. Experiments
  6. Datasets
  7. Deployment details
  8. Result analysis
  9. Conclusion

Figures (4)

  • Figure 1: Framework of graph based unsupervised extractive methods
  • Figure 2: A new strategy to extract sentences
  • Figure 3: ROUGE-1 f1 scores of TextRank(our strategy) with different $\alpha$. $\alpha$ = $\lambda$ = 1 at equivalent point and it means TextRank is equivalent to TextRank(our strategy) there. The left figure is of CNN/DailyMail, and the right figure is of NYT.
  • Figure 4: The position of the extracted sentences in the original text of NYT. The x-axis is the number of sentences in the original text, and the y-axis represents the position of the three sentences. The right two figures are the result of PACSUM, and the left two figures are the result of PACSUM with our strategy. The two figures above show the relative position, and the two figures below show the absolute position.