yuzhimanhua / Multi-BioNER

Cross-type Biomedical Named Entity Recognition with Deep Multi-task Learning (Bioinformatics'19)
https://arxiv.org/abs/1801.09851
Apache License 2.0
131 stars 28 forks source link

The co-train Parameter #14

Closed hantingge closed 4 years ago

hantingge commented 4 years ago

Hi,

What does the co_train parameter do in train_wc.py?

In the pre-trained model you provided, co_train is false.

yuzhimanhua commented 4 years ago

Hello,

co_train=True means the LSTM-CRF model is co-trained with a language model. Please refer to the following paper:

Liu et al., Empower Sequence Labeling with Task-Aware Neural Language Model. AAAI'18.

In our model, we do not incorporate this co-training strategy.

hantingge commented 4 years ago

Thank you for the reply.

So as long as I input multiple datasets, the model is trained as a MTM (Multi-task) by default. If I input only one dataset, it is (equivalent to) a Single-Task Model (LSTM-CRF). Is that correct?

yuzhimanhua commented 4 years ago

Yes, you are right.