rajat1433 / qna_matching

Closed Domain Question and Answer Matching

Question and Answer Matching

Given a pre-defined set of questions and its responses just like a set of FAQ's we try to predict an appropriate response for a user defined query.

We frame the problem as the Multilabel Classification Problem capitalizing on the power of the state of the art models, RoBerta in this project. Also we use a novel training approach to mitigate the lack of the training data across the closed domain.

Getting Started

Install the dependencies in the requirements.txt:

To run the demo :

python3 run-demo.py

then open the link on the browser: http://0.0.0.0:1995/

this can be modified in the run-demo.py file.

To train the model for the QnA Matching (Roberta),

python3 train_roberta.py

To use a pretrained model and get the accuracy score,

python3 roberta.py The already trained model can be downloaded at here

Authors/Contact

Rajat Mittal- Jan 2020 - May 2020

Acknowledgments

Andrew Koh Jin Jie and Prof. Chng Eng Siong Nanyang Technological University For the mentorship and assitance.