Enhanced Heterogeneous Graph Attention Network with a Novel Multilabel Focal Loss for Document-Level Relation...

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Information Extraction

Background:

• Document-level relation extraction aims to identify all relationships between entities within a document.
• Existing methods often use holistic entity representations, potentially losing information from fine-grained mentions.
• A shift towards mention-level analysis is proposed to ground relation prediction in specific entity instances.

Purpose of the Study:

• To propose a novel two-stage, mention-level framework for document-level relation extraction.
• To enhance the modeling of intra-sentential and inter-sentential relations using fine-grained entity mentions.
• To improve relation inference by considering long-distance semantic dependencies and coreference information.

Main Methods:

• A two-stage framework utilizing local and global mention representations.
• An enhanced heterogeneous graph attention network for modeling inter-sentential relations.
• An entity-coreference path-based strategy for relation inference.
• A novel cross-entropy-based multilabel focal loss function to handle class imbalance and multilabel prediction.

Main Results:

• The proposed mention-level framework significantly outperforms existing methods in document-level relation extraction.
• The enhanced heterogeneous graph attention network effectively captures long-distance semantic relationships.
• The novel loss function successfully addresses class imbalance and multilabel prediction challenges.

Conclusions:

• Grounding relation prediction in specific entity mentions is crucial for effective document-level relation extraction.
• The proposed framework offers a significant advancement in accurately identifying relations across multiple sentences.
• Future work can build upon this mention-level approach for more sophisticated information extraction tasks.