TensorFlow with a NER-Tagger

Question

I was wondering if there is any possibility to use Named-Entity-Recognition with a self trained model in tensorflow.

There is a word2vec implementation, but I could not find the 'classic' POS or NER tagger.

Thanks for your help!

Answer 1

You can adapt the Sequence-to-Sequence model for NER tagging. Your training text is the source vocabulary/sequences to the encoder:

Yesterday afternoon , Mike Smith drove to New York .

your BIO / BILOU NER tags are your target vocabulary/sequences to the decoder for NER tagging:

O O O B_PER I_PER O O B_LOC I_LOC O

or instead use POS tags to the decoder for POS tagging:

NN NN ,  NNP NNP VBD TO NNP NNP . 

[IMHO using a deep learning approach usually eliminates the need POS tagging as an intermediate step, unless you specifically need those features as an output for something.]

You would probably want to switch off the word embeddings for the decoder.

This well-known paper applies sequence-to-sequence models to syntactic parsing which has some similarities to the POS and/or NER tasks: Grammar as a Foreign Language