P0L3/CliReNER_v_1_1_28_GOLD

CliReNER Gold: Expert-Annotated Named Entity Recognition for Climate Research

Dataset Summary

CliReNER_gold is a high-fidelity, expert-curated Named Entity Recognition (NER) dataset for the climate change research domain. Designed specifically as a gold-standard evaluation benchmark, it consists of 192 sentences annotated by a panel of 12 domain experts.

While the silver version of this dataset is intended for model training, CliReNER_gold serves as the authoritative ground truth for assessing model performance, ensuring high inter-annotator agreement and consensus through Weighted Expert Voting (WEV).

Dataset Structure

Data Instances

Instances are provided in the standard BIO format, consistent with the CliReNER_silver schema to ensure seamless model evaluation.

{
  "id": "gold_doc_5-sent_12",
  "text": "The implementation of the carbon capture system reduces atmospheric CO2.",
  "tokens": ["The", "implementation", "of", "the", "carbon", "capture", "system", "reduces", "atmospheric", "CO2", "."],
  "ner_tags": [0, 29, 30, 0, 0, 7, 8, 53, 0, 0, 7, 0]
}

Entity Typology (28 Classes)

The dataset utilizes the same 28-class taxonomy as the silver version, covering geographical, biological, chemical, and methodological entities relevant to climate science. (See the associated paper for full definitions).

Dataset Creation

• Source Data: Sentences were selected from the same corpus as the silver dataset using a multi-set multi-cover formulation to ensure that, where possible, each of the 28 entity types appears at least 50 times across the set.
• Expert Annotation Protocol: We recruited 12 experts organized into six thematic pairs (e.g., "Biochemistry, ecology and pathology", "Atmospheric science and geophysics"). Annotators were assigned specific clusters of entity types matching their academic expertise.
• Consensus Methodology (Weighted Expert Voting): To resolve overlapping annotations from multiple experts, we applied a Weighted Expert Voting (WEV) scheme. Annotations from experts with domain-specific knowledge were weighted higher (w_e = 1.1) than non-experts (w_n = 0.9), with a deterministic two-stage voting process to resolve BIO boundary conflicts.
• Agreement: The final inter-annotator agreement among the 12 experts resulted in a Krippendorff's α of 0.6984. The WEV aggregation method showed high convergence with MACE probabilistic models (Cohen's κ = 0.9076).

Limitations

• Dataset Size: As a high-fidelity evaluation set, the sample size (192 sentences) is smaller than the silver training set. It is primarily intended for evaluation rather than training.
• Flat Schema: Like the silver dataset, this follows a flat NER schema, meaning nested and discontinuous entities are excluded to prioritize consistency and reduce complexity.

Usage

You can load the gold dataset for evaluation:

from datasets import load_dataset

# Load the gold-standard evaluation set
dataset = load_dataset("P0L3/CliReNER_v_1_1_28_GOLD")

# Evaluate your model on the test split
print(dataset["test"][0])

Citation Information

@misc{poleksic2026named,
  author       = {Poleksić, Andrija and Martinčić-Ipšić, Sanda},
  title        = {Named Entity Recognition for Climate Change Research},
  year         = {2026},
  howpublished = {Research Square},
  note         = {Preprint}
}