I implemented custom NER with bellow trained data first time and it gives me good prediction with Name and PrdName. I mentioned code bellow.

if __name__ == '__main__':
            ('My Name is Rajesh', {'entities': [(11, 17, 'Name')]}),
            ('My Name is Bakul', {'entities': [(11, 16, 'Name')]}),
            ('My Name is Pritam', {'entities': [(11, 17, 'Name')]}),
            ('My Name is Rakesh', {'entities': [(11, 17, 'Name')]}),
            ('My Name is Jayeeta', {'entities': [(11, 18, 'Name')]}),
            ('this is the price of bag', {'entities': [(21, 24, 'PrdName')]}),
            ('what is the price of ball?', {'entities': [(21, 25, 'PrdName')]}),
            ('what is the price of jegging?', {'entities': [(21, 28, 'PrdName')]}),
            ('what is the price of t-shirt?', {'entities': [(21, 28, 'PrdName')]}),

iterations = 20
    model = 'live_ner_model'
    nlp = spacy.load(model)  # load existing spacy model
    model = None
    nlp = spacy.blank('en')  # create blank Language class
    print("Created blank 'en' model")

if 'ner' not in nlp.pipe_names:
    ner = nlp.create_pipe('ner')
    print("Create NER")
    ner = nlp.get_pipe('ner')
    print("Exhisting NER")

# Add new entity labels to entity recognizer
for _, annotations in TRAIN_DATA:
    for ent in annotations.get('entities'):

# get names of other pipes to disable them during training
other_pipes = [pipe for pipe in nlp.pipe_names if pipe != 'ner']
with nlp.disable_pipes(*other_pipes):  # only train NER
    optimizer = nlp.begin_training()
    for itn in range(iterations):
        print("Statring iteration " + str(itn))
        losses = {}
        for text, annotations in TRAIN_DATA:
                [text],  # batch of texts
                [annotations],  # batch of annotations
                drop=0.2,  # dropout - make it harder to memorise data
                sgd=optimizer,  # callable to update weights

# Save model
output_dir = 'live_ner_model'
if output_dir is not None:
    output_dir = Path(output_dir)
    if not output_dir.exists():
    nlp.meta['name'] = model  # rename model
    print("Saved model to", output_dir)

# Test the saved model
output_dir = 'live_ner_model'
print("Loading from", output_dir)

nlp2 = spacy.load('live_ner_model')
test_text = """
   what is the price of cup. My Name is Rahim
doc2 = nlp2(test_text)
for ent in doc2.ents:
    print(ent.label_, ent.text)

But when I am trying to trained with some new data which has entity with only PrdName or any other new entity excluding Name in existing model. Then Name entity prediction goes wrong. I think this issue arises as I updated trained data excluding Name entity.

So is there any way we can improve training by not affecting existing training. Can someone share the idea? If possible please share a sample code.

Environment: Anaconda, spacy=v2.0.1, python=3.7


1 Answer 1


The model depends entirely on the training data: if you train with some data which has only PrdName as label, the model knows only this label and can predict only this label. You need to provide as much training data as possible, containing all the possible labels.

For the record, NER are usually trained with thousands of sentences in order to account for the diversity of the cases where a NE can appear.

  • $\begingroup$ Thanks for share your thought. I want to improve and correct an existing model by giving some more data. So it may not be old entity data. When I am providing more training data then old entity predicted wrongly which correctly predicted before. Not sure why same content which was predicted correctly but after update model it shows wrong prediction. Above example just for demo $\endgroup$
    – Rajesh das
    Jan 2, 2020 at 14:11
  • $\begingroup$ I got a reference in prodi.gy/docs/named-entity-recognition . We need to use ner.correct but I have no idea how to use it. $\endgroup$
    – Rajesh das
    Jan 2, 2020 at 14:27
  • $\begingroup$ @Rajeshdas Oh ok, I didn't know this kind of method. The standard way to achieve this would be to obtain the original training data for the existing model, concatenate it with the new data and re-train a model with that. $\endgroup$
    – Erwan
    Jan 2, 2020 at 15:42
  • $\begingroup$ @Rajeshdas The page you linked to is from Prodigy, not spaCy itself, and is about manual annotation of NER data. What you actually need to do is mix in the original training data, or create some with pseudo-rehearsal: explosion.ai/blog/pseudo-rehearsal-catastrophic-forgetting. $\endgroup$ Feb 1, 2020 at 19:47

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.