Using NLI to detect social biases in toxic data

Setup

To install all required dependencies run:

�pip install -r requirements.txt

To obtain fine-tuned bart-large model and the required tokenizer download them from:

https://drive.google.com/drive/folders/1JxGfoBGCEiFpzlP12u2hwoa17WOzVGQx?usp=sharing

And put then in the fine-tuningfolder

---

Files

Data folder

Includes all the data that was used for training, testing and fine-tuning the models

Results folder

Includes all the data that was outputed from the models for analysis and results.

These can be used to run the code without having to re-run the models each time.

---

Models

Contains the three different models:

• Bart-large
• Deberta-v3
• Deberta base

Helpers

Contains helper functions used by models

---

Initial testing of models and using rationales as hypotheses

testing-models.ipynb

Includes the initial testing of the three models:

• Bart-large
• Deberta-v3
• Deberta base

This includes rationales being used as hypotheses and experimentation with different hypotheses, and investigating neutral cases.

---

Generate results

generate-results.ipynb

Generates csv outputs of the results for the bart-large and deberta-v3 models using custom curated hypotheses.

Also runs fine-tuning of bart-large model using religion data that was manually annotated

---

Anaylsis

misc.ipynb

Contains misc anaylsis like, frequency distribution of lengths of hypotheses in NLI training data.

analysis.ipynb

Contains all f1 scores, AUC scores, graphs and accuracy breadowns for the test and training results