Gluon-nlp: Want to add bi-lstm CRF model

[fierceX]

Recently engaged in NLP work, has been concerned about mxnet, want to add bi-lstm CRF model, TF already has a CRF module, Hope Mxnet and gluon as soon as possible. Thank you

[szha]

Indeed, CRF is useful in many decoding tasks.

Contributions are welcome!

[zhiheng-huang]

+1. LSTM/BLSTM + CRF/segmental CRF would be very useful.

[lixin4ever]

@szha I will handle this issue.

[szha]

Great, thanks @lixin4ever. Feel free to connect with @sxjscience and he can help out if you have some questions :)

[szha]

https://github.com/kenjewu/CRF provides a good starting point. I'm currently looking into hybridizing CRF model.

[lixin4ever]

Sorry for the delay. I can continue to do this after I finish my work (maybe at the end of December), if it still remains unresolved at that time.

[hazelnutsgz]

fastNLP (https://github.com/fastnlp/fastNLP/), a NLP project based on Pytorch may of great help. FYI

CRF

https://github.com/fastnlp/fastNLP/blob/master/fastNLP/modules/decoder/CRF.py
https://fastnlp.readthedocs.io/en/latest/fastNLP.modules.decoder.html?highlight=CRF

BiLSTM + CRF

https://github.com/fastnlp/fastNLP/blob/master/fastNLP/models/sequence_modeling.py

[bratao]

@szha @lixin4ever Any update about the CRF support? It is a must have for many NLP applications

[eric-haibin-lin]

@kenjewu has a PR on this: https://github.com/dmlc/gluon-nlp/pull/466 but it's not merged yet. @kenjewu would you have some bandwidth to address the remaining review comments?

[eric-haibin-lin]

@bratao you might also be interested in https://github.com/dmlc/gluon-nlp/pull/612

[kenjewu]

Ok, I will continue to track the PR.