text_style_transfer

• This code implemens paper " Style Transfer in Text: Exploration and Evaluation " (https://arxiv.org/abs/1711.06861)
• The model code is based on https://github.com/nyu-dl/dl4mt-tutorial
• The classifier in Evaluation is based on keras: https://github.com/fchollet/keras/blob/master/examples/imdb_bidirectional_lstm.py
• The data is available at https://github.com/fuzhenxin/textstyletransferdata
• If there is any problem, please contact Zhenxin Fu (fuzhenxin95@gmail.com)
• This work is bsed on Theano
• Python requirement: theano numpy matplotlib

Model

Model is in model

model
    |style_transfer 
    |    |session_multi_decoder
    |    |    |train.sh
    |    |    |test.sh
    |    |    |com.sh
    |    |    |.......
    |    |
    |    |session_auto_encoder
    |    |    |similar to session_multi_decoder
    |    | 
    |    |session_style
    |         |similar to session_multi_decoder
    |data

Preprocess

cd model/style_transfer/data
python get_dict.py # generate vocabulary

Train and Test

$ cd model/style_transfer/session_multi_decoder
$ ./train.sh   # train model
$ ./test.sh    # test model
$ ./com.sh     # show results in compare.txt

Evaluation

Evaluation tool is in eval

Preprocess

• put glove embedding in eval/word_emb
• run bash run1.sh to copy results from model dir to current dir
• test1 test2 test3 for different mode (autoencoder, style embedding. multi decoder)

Transfer Strength (Classifier)

$ python classifier data        # process data of classifier
$ python classifier train       # train classifier
$ python classifier test test1  # test classifier
                                # test1 is the test result dir
                                # results in test1/embedding/style0_classification.txt ...

Content reservation

$cd eval
$python emb_test.py test1   # test1 is the test result dir
                           # results in test1/embedding/style0_semantics.txt ...
                           

Finally, run python eval.py to show results collection.

Example:

dir_name model_type      transfer_strength content_reservation mixture
================================================================================
test1     embedding8 		0.267 		0.943880306299 	0.208126303212
test1 	embedding4 		0.485 		0.915346000157 	0.317023657029
test1 	embedding 		0.593 		0.896598659955 	0.356930373024
.................

Acknowledgment

Thanks for Fangfang Zhang and Yixin Zhang for helping compose data.