Multi-granularity heterogeneous graph attention networks for extractive document summarization.

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Machine Learning

Background:

• Extractive document summarization is crucial in NLP.
• Existing Graph Neural Network (GNN) models struggle with multi-granularity semantic encoding and cross-sentence meta-path capture.

Purpose of the Study:

• To propose MHgatSum, a novel Multi-granularity Heterogeneous Graph ATtention network for extractive document summarization.
• To address limitations in current GNN-based summarization models.

Main Methods:

• Constructing a multi-granularity heterogeneous graph (HetG) with sentence, keyphrase, and topic nodes.
• Employing heterogeneous graph attention networks with hierarchical attention (node-level and semantic-level).
• Incorporating sentence node global importance into local attention mechanisms.

Main Results:

• MHgatSum effectively encodes multi-granularity semantic information.
• The model captures diverse cross-sentence meta-paths (e.g., Sentence-Keyphrase-Sentence, Sentence-Topic-Sentence).
• Empirical experiments on benchmark datasets show MHgatSum outperforms state-of-the-art models.

Conclusions:

• MHgatSum offers superior performance in extractive document summarization.
• The proposed multi-granularity heterogeneous graph approach enhances semantic representation.
• This work advances GNN applications in NLP summarization tasks.