At first sight, it seems correct to me.
For the prediction, we coded a small predictor ourselves and didn’t inherit it from AllenNLP class, even if it basically do the same thing. You can have a look at the code in the
demo folder (hmtlPredictor.py).
There is still a lack of understanding of the settings in which multi-task learning has a significant effect. However there are many evidences that training related tasks in a MTL settings can help.
You can probably start with a non-hierarchical multi-task setting in which all the tasks…
For now, it is not on the roadmap yet, but if the demand is strong, we can probably plan something.
At the moment, you can easily modify the code to use the released weights.