Hierarchical Human-Like Deep Neural Networks for Abstractive Text Summarization.

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Deep Learning

Background:

• Abstractive text summarization (ATS) aims to generate concise and coherent summaries.
• Deep learning has advanced ATS, but human-like summary generation remains a challenge.
• Human reading cognition is underexplored in current deep neural networks for ATS.

Purpose of the Study:

• To propose a novel Hierarchical Human-like deep neural network for ATS (HH-ATS).
• To mimic human reading cognition stages (rough reading, active reading, postediting) in an AI system.
• To improve the quality and human-likeness of abstractive summaries.

Main Methods:

• Developed HH-ATS, a deep neural network incorporating three components: knowledge-aware hierarchical attention, multitask learning, and a dual discriminator generative adversarial network.
• Modeled the system on human cognitive processes for reading comprehension and summary writing.
• Evaluated performance on benchmark datasets: CNN/Daily Mail and Gigaword.

Main Results:

• HH-ATS demonstrated superior performance compared to existing ATS methods.
• The proposed model consistently achieved substantial improvements on benchmark datasets.
• The hierarchical, human-cognition-inspired approach proved effective for ATS.

Conclusions:

• The HH-ATS model represents a significant advancement in abstractive text summarization.
• Mimicking human reading cognition is a promising direction for developing more sophisticated ATS systems.
• HH-ATS offers a new state-of-the-art for generating high-quality, human-like summaries.