"BANER: Boundary-Aware LLMs for Few-Shot Named Entity Recognition"

BANER teaches LLMs to spot entity boundaries better, making named entity recognition work with just a few examples.

BANER introduces boundary-aware contrastive learning and LoRAHub to enhance LLMs' ability to recognize named entities with minimal training examples.

-----

https://arxiv.org/abs/2412.02228v1

🤖 Original Problem:

Existing two-stage methods for few-shot Named Entity Recognition face challenges with false span detection and misaligned entity prototypes. LLMs, despite their capabilities, haven't been effective at few-shot information extraction.

-----

🔧 Solution in this Paper:

→ BANER introduces a boundary-aware contrastive learning strategy that helps LLMs better understand entity boundaries.

→ The system uses LoRAHub to align information between source and target domains, improving cross-domain classification.

→ A two-stage architecture separates entity span detection from type classification, with each stage optimized for better accuracy.

-----

💡 Key Insights:

→ Boundary-aware contrastive learning significantly improves entity span detection accuracy

→ Domain adaptation through LoRAHub enhances cross-domain performance

→ Two-stage architecture outperforms traditional end-to-end approaches

-----

📊 Results:

→ Achieved 5.2% improvement in F1 score for intra-task scenarios

→ Outperformed baselines by 2.3% in 1-shot and 5.1% in 5-shot cross-domain settings

→ Demonstrated effectiveness across various LLM architectures