TGIN: Translation-Based Graph Inference Network for Few-Shot Relational Triplet Extraction

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Machine Learning

Background:

• Traditional triplet extraction models require extensive data, hindering performance on long-tail relations.
• Pipeline and joint models face challenges with error propagation and data scarcity.

Purpose of the Study:

• To propose a novel end-to-end model, TGIN, for few-shot triplet extraction.
• To address the limitations of data-hungry models in real-world scenarios.

Main Methods:

• TGIN utilizes a multilayer heterogeneous graph with entity and relation nodes and edges.
• A graph aggregation and update method employs translation algebraic operations for feature mining.
• Label information is propagated from few labeled examples to unlabeled ones.

Main Results:

• TGIN significantly improves triplet extraction accuracy by 2.34%–10.74% over state-of-the-art baselines.
• The model demonstrates enhanced robustness in few-shot settings.
• Extensive experiments on reconstructed datasets validate TGIN's effectiveness.

Conclusions:

• TGIN is the first model to introduce a heterogeneous graph for few-shot relational triplet extraction.
• The proposed approach effectively handles data scarcity and reduces error propagation.
• TGIN offers a promising solution for improving triplet extraction performance in low-data regimes.