Template-Free Prompting for Few-Shot Named Entity Recognition via Semantic-Enhanced Contrastive Learning

Area of Science:

• Natural Language Processing
• Machine Learning
• Artificial Intelligence

Background:

• Prompt tuning excels in sentence-level tasks but struggles with token-level tasks like Named Entity Recognition (NER).
• Existing NER methods using N-gram traversal for prompting are computationally expensive.
• Few-shot learning scenarios present challenges for traditional contrastive learning methods.

Purpose of the Study:

• To develop an efficient and effective prompt-based contrastive learning method for few-shot NER.
• To eliminate the need for template construction and manual label word mappings.
• To improve contrastive learning performance in low-resource NLP tasks.

Main Methods:

• Proposed a template-free prompt (TFP) approach using external knowledge to initialize semantic anchors.
• Developed a semantic-enhanced contrastive loss for in-context optimization of prompts and sentence embeddings.
• Enabled contrastive learning in few-shot settings without a large number of negative samples.

Main Results:

• Achieved state-of-the-art (SOTA) performance on label extension, domain adaptation, and low-resource generalization tasks.
• Demonstrated effectiveness across six public datasets under various settings.
• Successfully addressed limitations of conventional contrastive learning in NLP tasks.

Conclusions:

• The proposed prompt-based contrastive learning method offers a significant advancement for few-shot NER.
• The method is efficient, avoids complex setup, and achieves superior results.
• This work provides a robust solution for low-resource NER challenges.