Few-Shot Text Classification with Global-Local Feature Information.

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Machine Learning

Background:

• Meta-learning frameworks are established for domain adaptation in computer vision with limited data.
• Text classification using meta-learning remains less explored, presenting a research gap.

Purpose of the Study:

• To propose SumFS, a novel meta-learning framework for text classification.
• To enhance feature representation by integrating extractive summarization and attention mechanisms.
• To address the challenge of limited labeled data in text classification tasks.

Main Methods:

• Developed SumFS with three modules: unsupervised text summarizer, weighting generator with attention, and a meta-learning framework with ridge regression.
• Created a new marine news dataset with limited labeled data for evaluation.
• Utilized THUCnews, Fudan, and the marine news datasets for performance testing.

Main Results:

• SumFS maintains or improves classification accuracy while reducing input features.
• Achieved a reduction of over 50% in training time per epoch.
• Demonstrated effective feature enhancement through summarization and attention weighting.

Conclusions:

• SumFS offers an efficient and effective meta-learning approach for text classification.
• The framework successfully handles limited labeled data scenarios.
• The proposed method shows significant improvements in both performance and computational efficiency.