nreimers
commented
6 years ago Hi @prohit93 you are asking though questions.
I would use the multi-task BiLSTM implementation. Then you train the model on coarse grain labels first.
After that, you take the base model and train a new classifier for your fine grained labels. The old classifier (softmax or CRF) for the coarse grained tags are no longer needed.
You can also try to run this as a multi task setup and provide data for the coarse grained labels along with the data for your fine grained labels.
However, I'm not certain if you will achieve that much improvement compared to directly training on your dataset with your fine grained labels. Most multi-task frameworks do not yield that much improvement and they are most often beaten by a simple pipeline approach:
- First, trained the network on task 1 (e.g. coarse NER tags)
- Add tags from task 1 as features to the input of task 2 and train the network on the sentences + the tags from task 1.
This pipeline approach was in all my experiment often much better than more complicated multi-task setups (or transfer learning setups).
Hello @nreimers , I really enjoyed going through your work. Could we improve your existing model / code to retrain for say 50 fine grained ner classes ?