This is the final project report for Umich EECS 595 Natural Language Processing, where we implemented 2 innovative approaches for improving an existing question answering system based on BERT or BERT-related pre-trained models.
The codes rely on python v3.7 and pytorch.
The codes are based on Transformers v2.3.0.
You can install the dependencies by cd to the root directory and run conda install -r requirements.
Here is how to train the model.
For this part, Yin Yuan fine-tune an albert-based model on his own computer, which is too large to be uploaded. You can either fine-tune the model by scripts in the sample code as well, or download a fine-tuned version from https://huggingface.co/.
To fine-tune the model, please run the run_squad.sh in sample_code directory.
After you get the fine-tuned model, you can open /src/Binary_Answerability_Verifier.ipynb in jupyter notebook and run all blocks in sequential order. The accuracy value would occur in the end of this notebook.
For this part, Cheng Qian create a 2 stage reading structure. Here, we uploaded the sentences embeddings using the model deepset/roberta-base-squad2 from https://huggingface.co/. You can run the 2 stage QA system and get a prediction file pred_with2stage.json by running python 2stageQA.py --model_name deepset/roberta-base-squad2 --reduced_range based on the model with reduced context range. Using --whole_range instead of --reduced_range can let stage 2 do prediction without narrowing the context range. You can also change the model name with other models in https://huggingface.co/, but this may cost some time to run. You can then evaluate the prediction by running python evaluate.py dev-v2.0.json pred_with2stage.json.