GeoCorpus-3v2 / README.md
ronunes's picture
Update README.md
4df2ef7 verified
metadata
configs:
  - config_name: holdout
    data_files:
      - split: train
        path: holdout/train.json
      - split: dev
        path: holdout/dev.json
      - split: test
        path: holdout/test.json
  - config_name: fold0
    data_files:
      - split: train
        path: 5folds/fold0/train.json
      - split: dev
        path: 5folds/fold0/dev.json
      - split: test
        path: 5folds/fold0/test.json
  - config_name: fold1
    data_files:
      - split: train
        path: 5folds/fold1/train.json
      - split: dev
        path: 5folds/fold1/dev.json
      - split: test
        path: 5folds/fold1/test.json
  - config_name: fold2
    data_files:
      - split: train
        path: 5folds/fold2/train.json
      - split: dev
        path: 5folds/fold2/dev.json
      - split: test
        path: 5folds/fold2/test.json
  - config_name: fold3
    data_files:
      - split: train
        path: 5folds/fold3/train.json
      - split: dev
        path: 5folds/fold3/dev.json
      - split: test
        path: 5folds/fold3/test.json
  - config_name: fold4
    data_files:
      - split: train
        path: 5folds/fold4/train.json
      - split: dev
        path: 5folds/fold4/dev.json
      - split: test
        path: 5folds/fold4/test.json
task_categories:
  - token-classification
language:
  - pt
tags:
  - geological
  - oil and gas

Description

This dataset is an updated version of the GeoCorpus-3, originally described in Portuguese word embeddings for the oil and gas industry: Development and evaluation , where duplicates have been removed and annotation errors have been corrected. The updated corpus is thoroughly described in the paper "An Evaluation of Large Language Models for Geological Named Entity Recognition", which will be presented at the ICTAI 2024 conference. The preprint of the paper is available here.

Dataset Structure

The dataset is divided into two main subsets:

  1. Holdout: A fixed train, dev, and test split, created using stratified sampling according to the methodology described here.
  2. Cross-Validation: A 5-fold cross-validation setup, with each fold also created using stratified sampling, following the same methodology described here.

To access these subsets, you can use the following code:

from datasets import load_dataset

# Load the holdout set
dataset_holdout = load_dataset("ronunes/GeoCorpus-3v2", name="holdout")

# Load a specific fold
dataset_fold0 = load_dataset("ronunes/GeoCorpus-3v2", name="fold0")
dataset_fold1 = load_dataset("ronunes/GeoCorpus-3v2", name="fold1")
# And so on for the other folds

Additional Resources

The code used for training and further experiments can be found in the GitHub repository associated with the paper. The repository also includes the corpus in CoNLL format for easy integration with various NLP tools.

BibTeX

Please cite the dataset using the following BibTeX entry:

@inproceedings{nunes2024geoner,
  author = {Nunes, Rafael and Spritzer, Andre and Balreira, Dennis and Freitas, Carla and Carbonera, Joel},
  year = {2024},
  pages = {},
  title = {An Evaluation of Large Language Models for Geological Named Entity Recognition}
  url={https://www.researchgate.net/publication/383822506_An_Evaluation_of_Large_Language_Models_for_Geological_Named_Entity_Recognition},
}