I am struggling during training an NER model using spacy transformer in GPU. This NER is related to build an accurate CV Parser. The main challenge here is that I can't give a lot of data to this model at once. There are approximately 35 labels to train NER for. I am getting very poor accuracy which is about 40% after giving 2.5K training data and 500 testing data. I trained 3 separate models to extract NER information. I need to improve it further for two aspects.
- Accuracy needs to be improved.
- Response tome should be improved to few seconds which is 45 seconds now. I am using Roberta Base model to do NER training. Please suggest me ideas. I would also like to understand the config.cfg file for spacy ner training. Thank you!!!
I tried simple NER training and transformer based training. I failed to achieve a good accuracy but when I see any video in Youtube or any blog related to NER training, there the accuracy is easily achieved 80 to 90 %.