UKPLab / emnlp2017-bilstm-cnn-crf

BiLSTM-CNN-CRF architecture for sequence tagging
Apache License 2.0
824 stars 263 forks source link

Train MTL model for two tasks using one dataset only. #38

Open Dragon615 opened 5 years ago

Dragon615 commented 5 years ago

I would like to train a multi-task learning model for two tasks, pos tagging, and language id. I have one dataset that is annotated for pos tagging and language id. My question is that is it possible to train a MTL model that learns these two tasks jointly using one dataset?

Thanks in advance for your response.

nreimers commented 5 years ago

Hi @Dragon615 one trick to achieve this is to create two copies of the folder with your train/dev/test files, i.e. data/pos_task/ and data/language_id_task/

Then you can configure the model so that it is trained on 'pos_task' for POS tagging and on 'language_id_task' for language identification.