sducouedic / bert-for-context-retrieval

Use a BERT model to retrieve the correct context for each question from SQuAD dataset.

Find Best Context With SQuAD

Installation

• In the root directory, create a new python environment and install all the requirements

python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt

• Download and unzip BERT pretrained for passage reranking in /models/. You should now have the following files:
  • /models/BERT_for_passage_ranking/config.json
  • /models/BERT_for_passage_ranking/pytorch_model.bin

The model is taken from link which adapted BERT for passage ranking. The model is converted to PyTorch using Hugging Face's transformers library.

Inference

• Go to src/ folder and run main.py with the desired question. The algorithm will print his best guess for the context:

cd src
python3 main.py --question "In what country is Normandy located?"

Future work

I didn't have the time to finish all I planned, here is what is missing:

• Fasten the computation by using batches and doing inference using torch tensor (see src/eval.py)
• Get feedback on the actual rank of the correct context
• Implement MRR evaluation metric on dataset
• Fine-tuning the model
• Add new arguments to the parser to handle test dataset