Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
41 tokens/sec
GPT-4o
59 tokens/sec
Gemini 2.5 Pro Pro
41 tokens/sec
o3 Pro
7 tokens/sec
GPT-4.1 Pro
50 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

Few-NERD: A Few-Shot Named Entity Recognition Dataset (2105.07464v6)

Published 16 May 2021 in cs.CL, cs.AI, and cs.LG
Few-NERD: A Few-Shot Named Entity Recognition Dataset

Abstract: Recently, considerable literature has grown up around the theme of few-shot named entity recognition (NER), but little published benchmark data specifically focused on the practical and challenging task. Current approaches collect existing supervised NER datasets and re-organize them to the few-shot setting for empirical study. These strategies conventionally aim to recognize coarse-grained entity types with few examples, while in practice, most unseen entity types are fine-grained. In this paper, we present Few-NERD, a large-scale human-annotated few-shot NER dataset with a hierarchy of 8 coarse-grained and 66 fine-grained entity types. Few-NERD consists of 188,238 sentences from Wikipedia, 4,601,160 words are included and each is annotated as context or a part of a two-level entity type. To the best of our knowledge, this is the first few-shot NER dataset and the largest human-crafted NER dataset. We construct benchmark tasks with different emphases to comprehensively assess the generalization capability of models. Extensive empirical results and analysis show that Few-NERD is challenging and the problem requires further research. We make Few-NERD public at https://ningding97.github.io/fewnerd/.

Few-NERD: A Few-shot Named Entity Recognition Dataset

The paper introduces Few-NERD, a significant contribution in the domain of Named Entity Recognition (NER), specifically tailored for the few-shot learning paradigm. Few-shot learning, a subset of machine learning, aims to train models effectively with a limited number of examples. Few-NERD addresses this challenge by providing a comprehensive and meticulously curated dataset, enriched with both coarse-grained and fine-grained entity types.

Dataset Composition and Design

Few-NERD spans 188,238 sentences sourced from Wikipedia, encompassing 4,601,160 words. Each word is annotated at two hierarchical levels: 8 coarse-grained and 66 fine-grained entity types. This dual-layered architecture is pivotal in facilitating the recognition of both broad and nuanced entity types, a gap often left unaddressed by traditional datasets. The dataset's scale and strategic design arguably make it one of the largest and most granular contributions to the NER community, specifically targeting the few-shot learning context.

Challenges Addressed

The paper identifies notable challenges prevalent in the few-shot NER domain. Traditional datasets are re-purposed into few-shot settings but often lack the granularity needed to effectively train models on unseen, fine-grained entity types. Few-NERD mitigates this issue by incorporating a diverse set of entity types, thus ensuring a more comprehensive training framework for model generalization.

Benchmark Tasks and Experimental Evaluation

Few-NERD introduces three benchmark tasks: Few-NERD (sup), Few-NERD (intra), and Few-NERD (inter). Each task serves to evaluate different dimensions of model generalization:

  • Few-NERD (sup): A standard supervised NER task, providing a broad overview of instance-level generalization across all entity types.
  • Few-NERD (intra): This task assesses type-level generalization, isolating entity types within different coarse-grained categories, thereby testing models on type discrimination without prior exposure.
  • Few-NERD (inter): Here, knowledge transfer is evaluated. The task tests cross-type generalization where entity types within the same coarse category share underlying semantics, thus focusing on coarse-level correlations.

The paper employs models such as BERT-Tagger, ProtoBERT, NNShot, and StructShot to establish baseline performances across these tasks. The empirical results clearly indicate the complexity and challenge posed by Few-NERD, as these state-of-the-art methods attain significantly lower performance compared to standard NER datasets.

Implications and Future Directions

Few-NERD stands as a formidable asset in the landscape of NER datasets due to its meticulous design and emphasis on few-shot learning. Its implications are far-reaching; it provides a rigorous benchmark aiding in the development and refinement of models that can adapt to new, unseen data with minimal supervision. This characteristic is particularly beneficial in fast-evolving domains where annotated data is scarce.

Future work could explore extending the dataset's utility by incorporating cross-domain annotations or even finer-grained hierarchies. Additionally, Few-NERD sets the stage for advancing continual learning paradigms within NER, paving the way for models that dynamically evolve with increasing data complexity.

In conclusion, Few-NERD is a substantial contribution to the NER and few-shot learning communities, offering a robust platform for the development of adaptable and resilient models. The dataset not only addresses present limitations but also anticipates future needs in AI and machine learning research.

User Edit Pencil Streamline Icon: https://streamlinehq.com
Authors (8)
  1. Ning Ding (122 papers)
  2. Guangwei Xu (18 papers)
  3. Yulin Chen (134 papers)
  4. Xiaobin Wang (39 papers)
  5. Xu Han (270 papers)
  6. Pengjun Xie (85 papers)
  7. Hai-Tao Zheng (94 papers)
  8. Zhiyuan Liu (433 papers)
Citations (210)