BiLSTM-CRF NERLearner - Skipping batch of size=1

[gohjiayi]

Hi there, I've been trying to better understand the BiLSTM-CRF NER model, more specifically the NERLearner class in bilstm_crf_ner/model/ner_learner.py.

To run the NER model, I have ran the generate_data and build_data scripts, and subsequently moved on to running the train and test scripts. However, I noticed that when training (and also running test.py), the line 'Skipping batch of size=1' has been logged many times due to the following snippet of code (in both train and test functions).

https://github.com/smitkiri/ehr-relation-extraction/blob/master/bilstm_crf_ner/model/ner_learner.py#L220-L222

if inputs['word_ids'].shape[0] == 1:
    self.logger.info('Skipping batch of size=1')
    continue

All items within my training set and evaluation set will be caught by this if-statement and not move on to the other half of the code. I have tried removing this chunk for evaluation and the model could produce some prediction output - but not to great accuracy as I suspect that it might be affecting the model performance when training.

UPDATE: I realised this was due to the batch size = 1 set, which is not suited for this model. My 2 questions below still remains!

Can I check what is this code for, and will removing it for training and evaluation be okay?

Another question, can I ask how did you derive the results as seen in the BiLSTM-CRF README file? Is there a specific script that you have executed to achieve that?

[smitkiri]

Hey! I haven't worked with the BiLSTM-CRF NER code much, but I'll try to figure this out over the weekend. From a quick glance, I think it might have something to do with how CRF works, it might require sequence length >1 to compute a score? I may be wrong here, but let me know if this makes sense. I got there from this line computing the loss.