Confidence and start/stop index in NER

[Grossmend]

Good day!

It is possible return from NER model confidence and (or) start/stop position entities?
By default model return ["x_tokens", "y_pred"]

[yoptar]

Hi @Grossmend,

1. It is possible to return probabilities for NER models based on BERT: you can set the return_probas option to true in the configuration, remove the last tag_vocab component from the pipeline and rename y_pred_ind to y_pred in bert_sequence_tagger’s out field.
2. There is no way at this moment to return start/stop positions as the model works with tokenized text. You can however tokenize your inputs beforehand remembering start and stop positions for every token and then match those with the model outputs.

[Grossmend]

Thanks!