BERT-PAGG: a Chinese relationship extraction model fusing PAGG and entity location information

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Knowledge Representation

Background:

• Knowledge graph construction relies on accurate relationship extraction.
• Existing models often overlook the significance of entity relative positions.
• Integrating external information improves traditional relationship extraction but can be further optimized.

Purpose of the Study:

• To propose a novel BERT-PAGG relationship extraction model.
• To incorporate entity location information and sentence-level context into relationship extraction.
• To enhance the performance of knowledge graph construction.

Main Methods:

• Developed a BERT-PAGG model combining BERT embeddings with a PAGG module.
• Integrated entity location information using segmented convolution neural networks.
• Employed attention mechanisms for semantic feature capture and gating mechanisms for information flow regulation.

Main Results:

• The BERT-PAGG model achieved state-of-the-art results on two Chinese relation extraction datasets.
• Ablation studies confirmed the effectiveness of the PAGG module in utilizing external information.
• The PAGG module led to a Macro-F1 score increase of at least 2.82%.

Conclusions:

• The proposed BERT-PAGG model effectively utilizes entity location and sentence information for improved relationship extraction.
• The PAGG module enhances the model's ability to process external information.
• This approach offers a significant advancement in knowledge graph construction accuracy.