The goal of this project is two folds, first is to create synthetic QA corpora for an unlabeled dataset for training custom QA models. Second goal is to develop general purpose question generation models for use cases like FAQ generation, test creation etc.
Specifically, 3 types of models are developed using the T5 model
You can play with these models on HuggingFace model hub using the inference API.
All training details can be found in this wandb project
All of these models are trained using the unified text-to-text approach.
The multi-task model can do QA, QG and answer span extraction and it's performance is almost similar or better than it's single task counterpart. This is very useful for deployment, because usually question generation systems need three models,
then we can compare the two answers to see if the generated question is correct or not. Having 3 models for single problem is a lot of complexity so a multi-task model is definitely useful here.
Currently these model can generate only factoid question as they are trained using SQuAD dataset. So the next step is to find right dataset for generating non-factoid question and questions whose answers are not explicit part of original text.
Check the repo for more details.
Don't forget to tag @patil-suraj in your comment, otherwise they may not be notified.