I have a few thousands of sentences with their B/I/O NER tags. I also have access to millions of sentences for which only some words are tagged. That is, for the other words, I don't know whether they are B, I or O. Is there some literature that discusses how to leverage those partially tagged sentences to improve the training of the classifier?
My tagger is based on https://github.com/guillaumegenthial/sequence_tagging. But I'm welcoming ideas that are not restricted to CRF+LSTM.
