Abstract:Summarization aims at extracting the salient information from a document and presenting the extracted information in a condensed form. Most existing methods for extractive text summarization generate a summary from a document using a two-stage process. In the first stage, the sentences are ranked based on their saliency scores and, in the second stage, the summary generation process starts with the top-ranked sentence and selects the next sentences one by one from the ranked list. To improve summary diversity, a sentence is included in the summary if the sentence is sufficiently dissimilar from the already selected sentences. Sentence selection is continued until the summary of the desired length is reached. The second stage is greedy in nature and it uses a predefined similarity threshold value to check the dissimilarity of a sentence with the already selected sentences. Due to this fixed similarity threshold which is manually tuned, in most cases, this approach fails to manage the diversity in a summary. This article proposes a summarization approach that uses a neural network-based learning model that learns to include a sentence in a summary by taking into account both the saliency of the sentence and the diversity in the summary. For this purpose, the model is trained using two types of features—saliency features and diversity features. We have evaluated the proposed approach using two open benchmark datasets—the DUC dataset and the Daily Mail dataset. Experimental results show that the proposed neural summarization approach is effective in producing better non-redundant informative summaries and outperforms many existing summarization approaches to which it is compared.

Neural sentence fusion for diversity driven abstractive multi-document summarization

Automatic Document Summarization Via Deep Neural Networks

Abstractive Multi-Document Summarization Via Joint Learning with Single-Document Summarization.

Towards a Neural Network Approach to Abstractive Multi-Document Summarization.

Large-Scale Multi-Document Summarization with Information Extraction and Compression

Abstractive Document Summarization via Neural Model with Joint Attention

FusionSum: Abstractive Summarization with Sentence Fusion and Cooperative Reinforcement Learning

Adapting Neural Single-Document Summarization Model for Abstractive Multi-Document Summarization: A Pilot Study.

A hybrid machine learning model for multi-document summarization

Abstractive Sentence Summarization with Guidance of Selective Multimodal Reference.

A New Method for Extractive Text Summarization Using Neural Networks

Multi-Document and Multi-Lingual Summarization Using Neural Networks

A Joint Sentence Scoring and Selection Framework for Neural Extractive Document Summarization.

Modeling Endorsement for Multi-Document Abstractive Summarization

Neural Document Summarization by Jointly Learning to Score and Select Sentences

Multi-Document Summarization Based On Two-Level Sparse Representation Model

Multi-Document Abstractive Summarization Using Chunk-graph and Recurrent Neural Network

Neural Summarization by Extracting Sentences and Words

Learning to Consider Relevance and Redundancy Dynamically for Abstractive Multi-document Summarization

An Unsupervised Multi-Document Summarization Framework Based on Neural Document Model.

Analyzing Sentence Fusion in Abstractive Summarization