Unified extractive-abstractive summarization: a hybrid approach utilizing BERT and transformer models for enhanced...

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Machine Learning

Background:

• The increasing volume of digital documents necessitates efficient automated document summarization (ADS).
• Existing summarization methods face challenges in generating dependable summaries that capture salient information accurately.
• Crafting high-quality summaries often relies on extracted features and human-defined parameters.

Purpose of the Study:

• To introduce an intelligent methodology integrating extractive and abstractive summarization techniques.
• To enhance the relevance between source documents and their generated summaries.
• To develop a summarization model producing human-like abstractive summaries.

Main Methods:

• Utilizing BERT for transforming input sentences into vector representations.
• Creating a similarity matrix from sentence representations for extractive summary generation.
• Employing a transformer model with a symmetric objective function for abstractive refinement.
• Integrating extractive and abstractive methods into a hybrid summarization framework.

Main Results:

• The proposed hybrid technique demonstrated superior performance on the CNN/DailyMail and DUC2004 datasets.
• Evaluation using ROUGE metrics confirmed the effectiveness of the integrated approach.
• The method achieved heightened relevance between input documents and generated summaries.

Conclusions:

• The hybrid summarization technique offers a significant advancement over conventional methods.
• The integration of BERT and transformer models yields high-quality, human-like summaries.
• This approach addresses key challenges in automated document summarization.