Custom NER for new entities - Model differences and NER details

[teresa-arima]

I am interested in using a Custom NER to differentiate two types of Person names, and I am not interested in the entities the trained NER has, only in the new ones I will train it for. I am considering a model with the tok2ver and NER components. I have some probably quite naÏve questions I cannot find the answer for:

1. What are the differences between employing a spacy blank “es” model or the large model, (es_core_news_lg in my case)?
2. Would there be any differences using the transformer model apart from the obvious performance gains?
3. When I train a custom NER, am I training the model from scratch with only the training set I provide or am I building upon the model’s knowledge? If so, what information learnt from the original corpus would be preserved?
4. In this line, should I also train the tok2ver component?

Thanks!

[svlandeg]

Hi @teresa-arima. If you're not interested in any of the NER labels from a pretrained model like the ner model from es_core_news_lg, then I would advice starting from a new ner model all together. But as with most things in ML, the proof is often in the pudding: you could try both approaches and see what works best for your use-case in terms of accuracy/F-score on an independent dev set.

If you want to avoid training from scratch and would like to use some pretrained tok2vec layer, you can source it from an existing pipeline and then use a Tok2VecListener within your (new) ner component so that it will have pretrained embeddings to start from. You can also have a look at our pretrain functionality.

If you're relatively new at this all, I advice reading through the training documentation in detail first, and in particular get more familiar with the config system which allows you to flexibly define and experiment with various different setups for pipelines.