Inspiration

For our research project involving question generation, we have had a difficult time to evaluate the results based on SQuAD [1]. We envision this project to simplify the process for anyone doing research and development involving BERT and SQuAD. Since it is TF 2.0, the code should be much easier to modify and customize for your needs, whether it be research, development, work, hackathons, and competitions (i.e., Kaggle). We therefore look forward to contributing to accelerating the development of question-answering systems based on BERT and TF 2.0!

[1] https://arxiv.org/abs/1909.05017

What it does

BERT-QA is an (Apache 2.0) open-source software repository that simplifies developing question-answering models based on BERT [2]. It is also available as a Python package via pip install bert_qa [3].

[2] https://github.com/artitw/BERT_QA

[3] https://pypi.org/project/bert-qa/

How we built it

We used pre-trained BERT models, SQuAD, GitHub, Colab, and TensorFlow 2.0!

Challenges we ran into

A lot of time was devoted to testing and making the project reproducible and consumable by the community.

Accomplishments that we're proud of

We have a working pip package [3] and Colab notebook [4] for everyone to try!

[3] https://pypi.org/project/bert-qa/

[4] https://colab.research.google.com/drive/1-tLvxSuI0ik2BaruaY_Ivoh_4eobWzEW

What we learned

TensorFlow 2.0 made it much simpler to develop the project, so we have developed the project to make it much simpler to develop BERT and question-answering related applications.

What's next for BERT QA

We will use this project as a springboard for conducting more research and competitions involving BERT and question-answering systems. There are also a lot of undocumented features and options that we will uncover for everyone.

Built With

Share this project:
×

Updates