Extractive text summarization model based on advantage actor-critic and graph matrix methodology.

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Machine Learning

Background:

• Automatic text summarization faces challenges in identifying informative segments and effective evaluation.
• Deep learning models are mainstream but suffer from exposure bias, limiting performance.
• Existing reinforcement learning models require improvement for better summarization outcomes.

Purpose of the Study:

• To introduce a novel extractive text summarization model using a graph matrix and advantage actor-critic (GA2C) method.
• To address the limitations of exposure bias in deep learning-based summarization.
• To enhance the performance of extractive text summarization through an improved evaluation mechanism.

Main Methods:

• Articles were pre-processed into a graph matrix representation.
• A decision-making network utilized graph matrix states to generate summary segments.
• An evaluation network scored these decisions, with feedback refining the decision-making process via the advantage actor-critic (GA2C) framework.

Main Results:

• The GA2C model demonstrated superior performance on the CNN/Daily Mail dataset compared to the Refresh baseline.
• Significant improvements were observed in Rouge-1 (0.70), Rouge-2 (9.01), and Rouge-L (2.73) metrics.
• Ablation studies confirmed the model's effectiveness and explored optimal configurations for activation functions and evaluation networks.

Conclusions:

• The proposed GA2C model effectively overcomes exposure bias in extractive text summarization.
• The graph matrix and actor-critic approach provide a robust framework for generating high-quality summaries.
• Further research can explore different reward functions and similarity matrices to optimize summarization performance.