POS Tagging for Twitter

TensorFlow implementation of Part-of-Speech Tagging for Twitter with Adversarial Neural Networks.

The code is partially referred to https://github.com/mkroutikov/tf-lstm-char-cnn and https://github.com/shucunt/domain_adaptation.

Requirements

• Python 2.7 or higher

• Numpy

• Tensorflow 1.0

• Gensim

• In addition, anyone who want to run these codes should download the word embedding and the saved model from http://pan.baidu.com/s/1boSlljL. The folder 'word2vec' and the files 'adv_model...' should be placed at 'TPANN/.'.

Usage

1.Reproducing the results of paper:

$ python adv_train.py --choice 0

which outputs dev_results.txt, test_results.txt and the accuracy of the model. Note that the model at the begnning might take a while reloading the parameters.

adv_model.ckpt contains the parameters of the model, which is to be used in the test time.

2.retraining the model

Then, run

$ python adv_train.py --choice 1

which will save the parameters of the final model and output the wrong labels(dev_text.txt and test_text.txt).

The parameters in this model are:

params	meaning	default
rnn_size	size of LSTM internal state	250
kernels	CNN kernel widths	[1,2,3,4,5,6]
kernel_features	number of features in the CNN kernel	[50,50,100,100,200,200]
adv_l	meta-parameter λ in gradient reversal layer (GRL)	0.7
char_embed_size	dimensionality of character embeddings	25
word_embed_size	dimensionality of word embeddings	200
max_word_length	maximum word length	35
param_init	initialize parameters at	0.05
batch_size	number of sequences to train on in parallel	20
max_epochs	number of full passes through the training data	100

You can change these values by passing the arguments explicitly. For example:

$ python adv_train.py --choice 1 --rnn_size 300 --adv_l 1 ...

Reference

Part-of-Speech Tagging for Twitter with Adversarial Neural Networks