| # Hugging Face with Bias Data in CoNLL Format | |
| ## Introduction | |
| This README provides guidance on how to use the Hugging Face platform with bias-tagged datasets in the CoNLL format. | |
| Such datasets are essential for studying and mitigating bias in AI models. | |
| This dataset is curated by **Shaina Raza**. | |
| The methods and formatting discussed here are based on the seminal work "Nbias: A natural language processing framework for BIAS identification in text" by Raza et al. (2024) (see citation below). | |
| ## Prerequisites | |
| - Install the Hugging Face `transformers` and `datasets` libraries: | |
| ```bash | |
| pip install transformers datasets | |
| ``` | |
| ## Data Format | |
| Bias data in CoNLL format can be structured similarly to standard CoNLL, but with labels indicating bias instead of named entities: | |
| ``` | |
| The O | |
| book O | |
| written B-BIAS | |
| by I-BIAS | |
| egoist I-BIAS | |
| women I-BIAS | |
| is O | |
| good O | |
| . O | |
| ``` | |
| Here, `B-` prefixes indicate the beginning of a biased term,`I-` indicates inside biased terms, and `O` stands for outside any biased entity. | |
| ## Steps to Use with Hugging Face | |
| 1. **Loading Bias-tagged CoNLL Data with Hugging Face** | |
| - If your bias-tagged dataset in CoNLL format is publicly available on the Hugging Face `datasets` hub, use: | |
| ```python | |
| from datasets import load_dataset | |
| dataset = load_dataset("newsmediabias/BIAS-CONLL") | |
| ``` | |
| - For custom datasets, ensure they are formatted correctly and use a local path to load them. | |
| If the dataset is gated/private, make sure you have run `huggingface-cli login` | |
| 2. **Preprocessing the Data** | |
| - Tokenization: | |
| ```python | |
| from transformers import AutoTokenizer | |
| tokenizer = AutoTokenizer.from_pretrained("YOUR_PREFERRED_MODEL_CHECKPOINT") | |
| tokenized_input = tokenizer(dataset['train']['tokens']) | |
| ``` | |
| 3. **Training a Model on Bias-tagged CoNLL Data** | |
| - Depending on your task, you may fine-tune a model on the bias data using Hugging Face's `Trainer` class or native PyTorch/TensorFlow code. | |
| 4. **Evaluation** | |
| - After training, evaluate the model's ability to recognize and possibly mitigate bias. | |
| - This might involve measuring the model's precision, recall, and F1 score on recognizing bias in text. | |
| 5. **Deployment** | |
| - Once satisfied with the model's performance, deploy it for real-world applications, always being mindful of its limitations and potential implications. | |
| Please cite us if you use it. | |
| **Reference to cite us** | |
| ``` | |
| @article{raza2024nbias, | |
| title={Nbias: A natural language processing framework for BIAS identification in text}, | |
| author={Raza, Shaina and Garg, Muskan and Reji, Deepak John and Bashir, Syed Raza and Ding, Chen}, | |
| journal={Expert Systems with Applications}, | |
| volume={237}, | |
| pages={121542}, | |
| year={2024}, | |
| publisher={Elsevier} | |
| } | |
| ``` |