Distant Supervision Relation Extraction via adaptive dependency-path and additional knowledge graph supervision.

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Machine Learning

Background:

• Relation Extraction (RE) systems traditionally require expensive labeled datasets.
• Distant Supervision Relation Extraction (DSRE) reduces costs but suffers from data noise.
• Existing dependency tree methods for DSRE struggle with path selection, leading to suboptimal pruning.

Purpose of the Study:

• To propose a novel DSRE framework, A²DSRE, addressing data noise and improving relation extraction.
• To develop adaptive dependency path selection using graph neural networks.
• To enhance DSRE performance with additional knowledge graph supervision.

Main Methods:

• Introduced A²DSRE framework combining Adaptive dependency-path and Additional KG supervision.
• Employed GeniePath, a graph neural network, for adaptive dependency path extraction by weighting neighbors and exploring correlations.
• Incorporated knowledge graph embedding to reduce data noise by aligning representations.

Main Results:

• The proposed GeniePath effectively prunes irrelevant nodes while retaining crucial ones for relation prediction.
• Adaptive path selection yields more appropriate paths associated with entity relations.
• Additional KG supervision successfully reduces noise in the DSRE data.

Conclusions:

• A²DSRE framework significantly improves relation extraction performance in noisy datasets.
• Adaptive dependency path selection and KG supervision are effective strategies for DSRE.
• The method demonstrates superior performance validated through extensive experiments.