---
dataset_info:
  features:
  - name: id
    dtype: string
  - name: text
    dtype: string
  - name: tokens
    sequence: string
  - name: ner_tags
    sequence:
      class_label:
        names:
          '0': O
          '1': B-Asset
          '2': I-Asset
          '3': B-Body Part
          '4': I-Body Part
          '5': B-Body of Water
          '6': I-Body of Water
          '7': B-Chemical
          '8': I-Chemical
          '9': B-Disease
          '10': I-Disease
          '11': B-Ecosystem
          '12': I-Ecosystem
          '13': B-Energy Source
          '14': I-Energy Source
          '15': B-Field of Study
          '16': I-Field of Study
          '17': B-Geographical Feature
          '18': I-Geographical Feature
          '19': B-Intellectual Artefact
          '20': I-Intellectual Artefact
          '21': B-Location
          '22': I-Location
          '23': B-Mathematical Expression
          '24': I-Mathematical Expression
          '25': B-Measuring Device
          '26': I-Measuring Device
          '27': B-Meteorological Phenomenon
          '28': I-Meteorological Phenomenon
          '29': B-Method
          '30': I-Method
          '31': B-Natural Disaster
          '32': I-Natural Disaster
          '33': B-Natural Phenomenon
          '34': I-Natural Phenomenon
          '35': B-Organism
          '36': I-Organism
          '37': B-Organization
          '38': I-Organization
          '39': B-Other
          '40': I-Other
          '41': B-Person
          '42': I-Person
          '43': B-Physical Artefact
          '44': I-Physical Artefact
          '45': B-Physical Phenomenon
          '46': I-Physical Phenomenon
          '47': B-Policy
          '48': I-Policy
          '49': B-Quantity
          '50': I-Quantity
          '51': B-Satellite
          '52': I-Satellite
          '53': B-System
          '54': I-System
          '55': B-Time Period
          '56': I-Time Period
  splits:
  - name: train
    num_bytes: 576199
    num_examples: 803
  - name: validation
    num_bytes: 74882
    num_examples: 106
  - name: test
    num_bytes: 83532
    num_examples: 118
  download_size: 256360
  dataset_size: 734613
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: validation
    path: data/validation-*
  - split: test
    path: data/test-*
license: cc-by-4.0
task_categories:
- token-classification
language:
- en
tags:
- climate-change
- ner
- biology
- earth-science
---

# CliReNER Silver: Named Entity Recognition for Climate Research

## Dataset Summary

**CliReNER *silver*** is a fine-grained Named Entity Recognition (NER) dataset designed specifically for the climate change research domain. Recognizing the lack of broad-coverage, expert-annotated resources in Climate NLP, this dataset provides sentence-level annotations across **28 distinct entity types** (27 domain-specific types + `Other`). 

The dataset is derived from 50 full-text peer-reviewed scientific publications covering diverse topics in climate science. Sentences were pre-annotated using GLiNER and rigorously curated by the authors using a flat NER schema (nested entities are excluded).

This ***silver*** dataset consists of 1,027 sentences split into `train` (803), `validation` (106), and `test` (118). It is intended primarily for training and fine-tuning models. For high-fidelity, multi-expert evaluation, please refer to our companion dataset, [**CliReNER<sub>gold</sub>**](https://huggingface.co/datasets/P0L3/CliReNER_v_1_1_28_GOLD) (annotated by 12 domain experts) or pre-expert-annotation version [**CliReNER<sub>gold author annotations</sub>**](https://huggingface.co/datasets/P0L3/CliReNER_v_1_1_28_GOLD_authorannots).

## Dataset Structure

### Data Instances
A typical instance in the dataset represents a single sentence from a scientific paper, tokenized and tagged using the standard BIO (Begin, Inside, Outside) format.

```python
{
  "id": "doc_12-sent_45",
  "text": "The increase in surface water temperature affects the local ecosystem.",
  "tokens": ["The", "increase", "in", "surface", "water", "temperature", "affects", "the", "local", "ecosystem", "."],
  "ner_tags": [0, 0, 0, 1, 2, 2, 0, 0, 0, 11, 0]
}
```

### Data Fields
- `id`: A unique string identifier for the sentence.
- `text`: The original raw string of the sentence.
- `tokens`: A list of strings representing the tokenized sentence.
- `ner_tags`: A list of integers representing the BIO tags for each token. There are 57 tags in total (0 for `O`, and 1-56 representing the `B-` and `I-` tags for the 28 entity types).

## Entity Typology (28 Classes)

The dataset covers a highly granular taxonomy designed for climate change literature:
*   `Geographical Feature`, `Body of Water`, `Meteorological Phenomenon`, `Natural Phenomenon`, `Natural Disaster`, `Physical Phenomenon`.
*   `Organism`, `Ecosystem`, `Body Part`, `Disease`.
*   `Chemical`, `Energy Source`, `Quantity`, `Mathematical Expression`.
*   `Organization`, `Person`, `Location`, `Method`, `Measuring Device`, `Satellite`, `Physical Artefact`, `Intellectual Artefact`, `Policy`, `System`, `Asset`, `Field of Study`, `Time Period`.
*   `Other`.

*(For complete definitions of each entity type, please refer to Appendix B of the associated paper).*

## Dataset Creation

- **Source Data:** Sentences were sampled from a corpus of 50 publications drawn from 25 peer-reviewed journals (curated by [Poleksić and Martinčić-Ipšić, 2024](https://ceur-ws.org/Vol-3747/text2kg_paper9.pdf)).
- **Annotation Process:** We formulated a flat NER schema to reduce cognitive load and model complexity, meaning overlapping or nested entities were resolved to the most relevant span (e.g., in "*gridded rain gauge data*", the full span is marked as `Intellectual Artefact`, subsuming the inner `Measuring Device`). Pre-annotation was performed via [GLiNER](https://huggingface.co/gliner-community/gliner_medium-v2.5), followed by manual author correction in Label Studio.

## Limitations and Biases

- **Class Imbalance:** Due to the natural distribution of scientific text, some classes (e.g., `Chemical`, `Organism`, `Time Period`) are highly frequent, while others (e.g., `Asset`, `Body Part`, `Natural Disaster`) appear sparsely. 
- **Flat Schema:** The dataset constrains entity overlap. Discontinuous and nested entities are not captured, which inherently results in some loss of granular information.
- **Silver Standard:** This specific split (`silver`) was author-annotated. While highly accurate, it does not possess the multi-annotator consensus methodology applied to the `gold` test set.

## Usage

You can easily load the dataset using the Hugging Face `datasets` library:

```python
from datasets import load_dataset

# Load the dataset
dataset = load_dataset("P0L3/CliReNER_v_1_1_28_SILVER")

# View the first training example
print(dataset["train"][0])

# Access the label names (string representations of the integers)
label_names = dataset["train"].features["ner_tags"].feature.names
print(label_names)
```

## Citation Information

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