neural-network-tagger

A General Purpose Tagger for POS Tagging, NER Tagging, and Chunking.

Syntaxnet

1. Prepare WSJ Data for Part-OF-Speech Tagging a. convert to conll format https://gist.github.com/khlmnn/3cc07407a002bb1773cd b. map XPOSTAG to UPOSTAG before training using convert.py

2. Install Syntaxnet https://github.com/tensorflow/models/tree/a9133ae914b44602c5f26afbbd7dd794ff9c6637/syntaxnet

3. Train and test the model using taggerTrain.sh, taggerTest.sh and tagger.pbtxt

FeedForward Model

cd PATH_TO_TAGGER/src/feedforward_model

Training: python tagger_trainer.py

Evaluating: python tagger_eval.py

BiLSTM-CRF Model

Reference: https://github.com/guillaumegenthial/sequence_tagging

Mention2Vec with FF/BiLSTMs

python build_data.py python main_ff.py/main.py

Experiments

Model	POS	NER	Chunk
Feedforword (word)	95.89	-	-
Feedforword (history and spelling features)	97.31	-	-
Bi-LSTM (word)	95.88	78.66	-
Bi-LSTM (Character Embedding)	97.08	78.87	-
Bi-LSTM-CRF (Character Embedding)	97.34	-	-

Decoding Time (words/sec, with GPU enabled)

Model	POS	NER	Chunk
Feedforward Model(word feature only)	11000/s	-	-
Feedforword Model (history spelling feature)	~8000/s	-	-
Bi-LSTM-CRF (word feature only)	~2000/s	-	-
Bi-LSTM-CRF (Character Embedding)	~1500/s	-	-