Multi-Hop Question Answering

Implemented a question and answering model for multi-hop questions that requires logical inference or aggregation of information from various parts of the information text (like referring multiple wikis to answer a question). Please check out the report for results.

Instructions

• To view answer prediction results on test data with trained weights or baseline model, use the Colab notebooks from the resources below.

• To train a fresh model with the following configurations, use utils/model_trainer.py

  • Configurations
    Tokenizer Max Length = 1024
    Epochs = 2
    Learning Rate = 0.00005
    Architecture Name = allenai/longformer-base-4096
    Save Name for Weights = neew_weights\
  • Command
    python model_trainer.py 1024 8 2 0.00005 allenai/longformer-base-4096/ new_weights\

• The data pre-processor script and the data splitter script can be found in utils/model_trainer

Resources

Data set

https://drive.google.com/file/d/1pFJ0NAvMSn7C-vI-hzeSsCG7ppJGa8EV/view?usp=sharing

Tokenizer

https://drive.google.com/file/d/1Ra6HNBnP8bGutLi7076Vk0kizznjOL-X/view?usp=sharing

Model

https://drive.google.com/file/d/1DMWe7bLI0FZ6Qd5tUyIYBCWMy6uLrHy5/view?usp=sharing

Interactive Notebooks (NYU Account Only)

Baseline Model Inference Notebook

https://colab.research.google.com/drive/1Yn7ARYrp3JGKNeBrTnuL2XMMvVkpQwto?authuser=1

Trained Model Inference Notebook

https://colab.research.google.com/drive/10B71qnh9oAkeWJ7x71dTOABJh92KxHGs?authuser=1