COPNER: Contrastive Learning with Prompt Guidance for Few‑Shot Named Entity Recognition

Research background : Distance metric learning has become a popular solution for few‑shot named entity recognition (NER), with current methods focusing on learning similarity metrics to measure semantic similarity between test samples and referents.

Limitations of existing methods : (1) They aim to learn a suitable similarity metric rather than optimizing encoder parameters for richer entity representations, severely limiting the model's ability to capture class‑related semantics. (2) Fixed metric referents derived from scarce few‑shot annotations may not adequately represent the semantics of the corresponding entity classes.

Overall framework : COPNER introduces class‑specific words (CWs) as proxies for entities, appending CW prompts to the original input sentence. During training, CW representations serve as token‑level supervision; a contrastive loss pulls together representations of tokens belonging to the same class and their anchor CW, while pushing apart different‑class representations, enabling the sentence encoder to align entity semantics with CW semantics in a unified space. During inference, CW representations act as metric referents for predicting entity classes; because CWs contain generic semantics and are refined through training, they are more suitable and stable than previous referents.

Experimental results :

Using the Few‑NERD dataset (with INTER and INTRA tasks), COPNER markedly outperforms existing state‑of‑the‑art (SOTA) methods on both tasks.

Cross‑label‑space evaluation shows COPNER’s superiority when trained on OntoNotes 5.0 and tested on CONLL’03, WNUT’17, and I2B2’14.

Domain‑transfer experiments with K‑shot settings (K = 5, 10, 20, 50) on CoNLL’03 and MIT Movie datasets demonstrate that COPNER consistently exceeds SOTA performance.