[Paper Review] PromptNER: A Prompting Method for Few-shot Named Entity Recognition via k Nearest Neighbor Search

Mozhi Zhang, Hang Yan+2 more|arXiv (Cornell University)|May 20, 2023

Topic Modeling4 citations

TL;DR

PromptNER proposes a prompting-based few-shot NER method that leverages k-nearest neighbor (k-NN) search to optimize label prototypes using only support set examples, enabling effective fine-tuning in low-resource target domains. It achieves state-of-the-art performance on Few-NERD and CrossNER by combining prompt engineering, a position-aware biaffine span detector, and a prompt-based classifier with k-NN refinement.

ABSTRACT

Few-shot Named Entity Recognition (NER) is a task aiming to identify named entities via limited annotated samples. Recently, prototypical networks have shown promising performance in few-shot NER. Most of prototypical networks will utilize the entities from the support set to construct label prototypes and use the query set to compute span-level similarities and optimize these label prototype representations. However, these methods are usually unsuitable for fine-tuning in the target domain, where only the support set is available. In this paper, we propose PromptNER: a novel prompting method for few-shot NER via k nearest neighbor search. We use prompts that contains entity category information to construct label prototypes, which enables our model to fine-tune with only the support set. Our approach achieves excellent transfer learning ability, and extensive experiments on the Few-NERD and CrossNER datasets demonstrate that our model achieves superior performance over state-of-the-art methods.

Motivation & Objective

• To address the limitation of traditional prototypical networks in few-shot NER, which cannot be fine-tuned on novel support sets due to fixed label prototypes.
• To enable transfer learning in few-shot NER by allowing model optimization using only the support set during inference.
• To improve span detection and entity typing by integrating k-NN search with prompt-based classification.
• To provide a simple yet effective baseline for few-shot NER that generalizes across domains and entity types.

Proposed method

• Constructs natural language prompts that embed entity category information to guide pre-trained language models (PLMs) in entity extraction.
• Employs a position-aware biaffine module to detect candidate spans from the input sentence.
• Uses a prompt-based classifier to perform span-level metric learning, replacing traditional prototype-based similarity computation.
• Introduces k-NN search to retrieve ground-truth entity representations from the support set for refining span predictions.
• Optimizes label prototype embeddings directly on the novel support set without requiring query sets, enabling end-to-end fine-tuning.
• Applies a novel reranking strategy to filter false positive spans by measuring their distance to nearest label prototypes.

Experimental results

Research questions

• RQ1 Can a prompting-based method improve few-shot NER performance by enabling fine-tuning on only the support set, without relying on query sets?
• RQ2 How effective is k-NN search in enhancing label prototype representation learning when only limited support examples are available?
• RQ3 Can prompt-based classification and span detection outperform traditional prototype-based metric learning in few-shot NER?
• RQ4 Does the proposed method generalize across different domains and entity types in low-resource settings?
• RQ5 Can the integration of k-NN and prompt engineering reduce false positive span predictions in few-shot NER?

Key findings

• PromptNER achieves state-of-the-art performance on the Few-NERD and CrossNER benchmarks, outperforming existing few-shot NER methods.
• The model demonstrates strong transfer learning ability, enabling effective fine-tuning on novel support sets without requiring query data during adaptation.
• The k-NN search component significantly improves prediction accuracy by retrieving relevant support examples for span refinement.
• The proposed reranking strategy effectively filters false positive spans, especially for unseen entity types.
• The prompt-based classifier with k-NN integration leads to more robust and generalizable entity typing compared to standard prototype-based approaches.
• The method maintains high performance across multiple N-way K-shot settings, confirming its robustness in low-resource scenarios.