Improving few-shot relation classification with multi-scale hierarchical prototype learning.

Area of Science:

• Natural Language Processing
• Machine Learning
• Artificial Intelligence

Background:

• Few-shot relation classification faces challenges with limited annotated data, leading to inaccurate prototypes.
• Existing methods often overlook hierarchical relational information crucial for accurate classification.

Purpose of the Study:

• To propose a novel multi-scale hierarchical prototype (Mario) learning method for few-shot relation classification.
• To enhance the understanding of global semantic information and distinguish subtle class differences.

Main Methods:

• Developed a multi-scale hierarchical prototype learning method (Mario) capturing inter-set, inter-class, and intra-class relational information.
• Incorporated relational descriptive information to mitigate textual expression diversity.

Main Results:

• Achieved high accuracy rates (92.52%/95.33%/85.46%/91.33%) on the FewRel dataset across four few-shot settings.
• Outperformed the strongest baseline by 2.87% and 4.29% in critical 5-way and 10-way 1-shot settings, respectively.

Conclusions:

• The proposed Mario method effectively enhances few-shot relation classification by leveraging multi-scale hierarchical prototypes.
• The model demonstrates superior performance, particularly in challenging low-data scenarios, by better emulating human cognitive processes.