Knowledge-enhanced prototypical network with class cluster loss for few-shot relation classification.

Area of Science:

• Natural Language Processing
• Machine Learning
• Artificial Intelligence

Background:

• Few-shot relation classification (FRC) aims to identify relationships between entities using limited labeled data.
• Existing prototype network methods often incorporate external knowledge but struggle with generalization due to complex network structures.
• Current models often neglect intra-class compactness, hindering their ability to manage outlier samples effectively.

Purpose of the Study:

• To propose a novel method for few-shot relation classification that overcomes limitations in model generalization and outlier handling.
• To enhance the representation capabilities of class prototypes in FRC models.
• To improve the discriminative power of learned feature spaces.

Main Methods:

• Introduced a non-weighted prototype enhancement module utilizing feature-level similarity for feature filtering and completion.
• Designed a class cluster loss function that explicitly enforces intra-class compactness and inter-class separability.
• Employed techniques to sample difficult positive and negative instances for robust training.

Main Results:

• The proposed model demonstrated significant effectiveness on the FewRel 1.0 and 2.0 datasets.
• The non-weighted prototype enhancement module improved prototype representation.
• The class cluster loss enhanced the model's ability to learn a highly discriminative metric space.

Conclusions:

• The developed approach offers a more effective solution for few-shot relation classification.
• The proposed methods contribute to better generalization and outlier handling in FRC models.
• The findings highlight the importance of explicit intra-class and inter-class constraints in metric learning for FRC.