Multi-task Fine-tuning with BERT

The goal of this approach is to reduce the number of additional parameters required compared to task-specific models and also realize better transfer learning due to the model architecture as well as training methods.