Question Answering on SQuAD 2.0

• 

  Poster Demo

Inspiration

A big challenge for machines right now is reading and understanding text. How can we make them understand the meaning of a question and subsequently find an appropriate answer? This is the challenge that I tried to tackle in this project using the SQuAD 2.0 dataset.

What it does

It takes a question asked over a paragraph of text from a Wikipedia article. The output is the predicted answer segment in the text paragraph. Sometimes, the questions will not have a possible answer on the text segment provided.

How I built it

I used a pre-trained BERT for question answering fine-tuned over the SQuAD dataset. Then I screened for impossible answers to improve prediction results.

Challenges I ran into

The model took a lot of computational power to run, so I introduced a sampling rate to reduce the amount of data.

Accomplishments that I'm proud of

Right now, I have been able to improve the accuracy from using the basic BERT model by screening for possible answers. I have learned a lot throughout this project and know that I will definitely be able to put this knowledge to use in the future.

What's next for Question Answering on SQuAD 2.0

Ultimately, the goal is to keep improving the accuracy of the predictions. In the future, the model could be fined tuned further.

Built With

• bert
• python