nggsam/paragen

Neural Paraphrase Generation based on OpenNMT-py

1. Quora Question Pair Dataset

1.1. preprocessing:

python preprocess.py -train_src para_data/quora/train.q1 -train_tgt para_data/quora/train.q2 -valid_src para_data/quora/val.q1 -valid_tgt para_data/quora/val.q2 -save_data data/para_quora -dynamic_dict -share_vocab

1.2. training:

python train.py -data data/para_quora -save_model models/quora-model -copy_attn -global_attention mlp -word_vec_size 128 -rnn_size 256 -layers 1 -encoder_type brnn -epochs 16 -seed 777 -batch_size 256 -max_grad_norm 2 -share_embeddings -gpuid 1

1.3. paraphrasing:

python translate.py -model models/quora-model_acc_xxx_ppl_yyy_e2.pt -src para_data/quora/test.q1 -output para_gen.txt -beam_size 10 -replace_unk -verbose

2. WikiAnswers Dataset

2.1. preprocessing:

python preprocess.py -train_src para_data/wikianswers/train.q1 -train_tgt para_data/wikianswers/train.q2 -valid_src para_data/wikianswers/val.q1 -valid_tgt para_data/wikianswers/val.q2 -save_data data/para_wikianswers -dynamic_dict -share_vocab

2.2. training:

python train.py -data data/para_wikianswers -save_model models/wikianswers-model -copy_attn -global_attention mlp -word_vec_size 128 -rnn_size 256 -layers 1 -encoder_type brnn -epochs 16 -seed 777 -batch_size 32 -max_grad_norm 2 -share_embeddings -gpuid 1

2.3. paraphrasing:

python translate.py -model models/wikianswers-model_acc_xxx_ppl_yyy_e2.pt -src para_data/quora/test.q1 -output para_gen.txt -beam_size 10 -replace_unk -verbose

3. Paraphrasing the SQuAD questions training dataset

python squad_process/preprocess_squad.py
split -l 15000 squad_process/squad_train.qtns.input -d -a 1 squad_process/squad_train.qtns.input.split_ 
python translate.py -model models/quora-model_acc_66.37_ppl_5.94_e14.pt -src squad_process/squad_train.qtns.input.split_0 -output squad_process/squad_train.qtns.para.split_0  -beam_size 10 -replace_unk -share_vocab -verbose -batch_size 1 -gpu 0
python translate.py -model models/quora-model_acc_66.37_ppl_5.94_e14.pt -src squad_process/squad_train.qtns.input.split_1 -output squad_process/squad_train.qtns.para.split_1  -beam_size 10 -replace_unk -share_vocab -verbose -batch_size 1 -gpu 1
python translate.py -model models/quora-model_acc_66.37_ppl_5.94_e14.pt -src squad_process/squad_train.qtns.input.split_2 -output squad_process/squad_train.qtns.para.split_2  -beam_size 10 -replace_unk -share_vocab -verbose -batch_size 1 -gpu 2
python translate.py -model models/quora-model_acc_66.37_ppl_5.94_e14.pt -src squad_process/squad_train.qtns.input.split_3 -output squad_process/squad_train.qtns.para.split_3  -beam_size 10 -replace_unk -share_vocab -verbose -batch_size 1 -gpu 3
python translate.py -model models/quora-model_acc_66.37_ppl_5.94_e14.pt -src squad_process/squad_train.qtns.input.split_4 -output squad_process/squad_train.qtns.para.split_4  -beam_size 10 -replace_unk -share_vocab -verbose -batch_size 1 -gpu 4
python translate.py -model models/quora-model_acc_66.37_ppl_5.94_e14.pt -src squad_process/squad_train.qtns.input.split_5 -output squad_process/squad_train.qtns.para.split_5  -beam_size 10 -replace_unk -share_vocab -verbose -batch_size 1 -gpu 5
cat  squad_process/squad_train.qtns.para.split_0 squad_process/squad_train.qtns.para.split_1 squad_process/squad_train.qtns.para.split_2 squad_process/squad_train.qtns.para.split_3 squad_process/squad_train.qtns.para.split_4 squad_process/squad_train.qtns.para.split_5 > squad_process/squad_train.qtns.para


cd squad_process
python insert_paraphrased_questions.py

4. Paraphrase Context via Bidirectional Translation

4.1. Preprocessing Wikipedia Parallel Corpora

mkdir nmt_data
cd nmt_data
wget http://opus.nlpl.eu/download.php?f=Wikipedia/de-en.txt.zip
unzip de-en.txt.zip
cd wiki_de-en

python preprocess.py en
python preprocess.py de

This parallel corpora has 2,459,662 sentence pairs. We used fist half (1,229,381) for en-de and last half for de-en, each with 1,229,381 -> 1,220,000 for training, 9,381 for validation.

4.2. Training Translators

4.2.1 Preprocessing

python preprocess.py -train_src nmt_data/wiki_de-en/en_de_train.en -train_tgt nmt_data/wiki_de-en/en_de_train.de -valid_src nmt_data/wiki_de-en/en_de_val.en -valid_tgt nmt_data/wiki_de-en/en_de_val.de -save_data data/wiki_en_de -dynamic_dict -share_vocab
python preprocess.py -train_src nmt_data/wiki_de-en/de_en_train.de -train_tgt nmt_data/wiki_de-en/de_en_train.en -valid_src nmt_data/wiki_de-en/de_en_val.de -valid_tgt nmt_data/wiki_de-en/de_en_val.en -save_data data/wiki_de_en -dynamic_dict -share_vocab

4.2.2 Training

python train.py -data data/wiki_en_de -save_model models/wiki_en_de -copy_attn -global_attention mlp -word_vec_size 256 -rnn_size 512 -layers 2 -encoder_type brnn -epochs 16 -seed 42 -batch_size 256 -max_grad_norm 2 -gpuid 6
python train.py -data data/wiki_de_en -save_model models/wiki_de_en -copy_attn -global_attention mlp -word_vec_size 256 -rnn_size 512 -layers 2 -encoder_type brnn -epochs 16 -seed 42 -batch_size 256 -max_grad_norm 2 -gpuid 7