wyu-du / reinforce-paraphrase-generation

This repository contains the data and code for the paper "An Empirical Comparison on Imitation Learning and Reinforcement Learning for Paraphrase Generation" (EMNLP2019).

A Unified Reinforcement Learning Framework for Pointer Generator Model

This repository contains the data and code for the paper "An Empirical Comparison on Imitation Learning and Reinforcement Learning for Paraphrase Generation".

Useage

Training

1. Model Setting: modify the path where the model will be saved.

vim config.py
log_root = os.path.join(root_dir, "Reinforce-Paraphrase-Generation/log_twitter")

2. Pre-train: train the standard pointer-generator model with supervised learning from scratch.

python train.py

3. Fine-tune: modify the training mode and the path where the fine-tuned model will be saved.

vim config.py
log_root = os.path.join(root_dir, "Reinforce-Paraphrase-Generation/log_rl")
mode = "RL"

Fine tune the pointer-generator model with REINFORCE algorithm.

python train.py -m ../log_twitter/best_model/model_best_XXXXX

Decoding & Evaluation

1. Decoding: first, specify the model path.

vim config.py
log_root = os.path.join(root_dir, "Reinforce-Paraphrase-Generation/log_twitter")

Second, apply beam search to generate sentences on test set:

python decode.py ../log_twitter/best_model/model_best_XXXXX

2. Evaluation:

   • The average BLEU score will show up automatically in the terminal after finishing decoding.

   • If you want to get the ROUGE scores, you should first intall pyrouge, here is the guidance. Then, you can uncomment the code snippet specified in utils.py and decode.py. Finally, run decode.py to get the ROUGE scores.

reinforce-paraphrase-generation's Issues

Question about the vocab size ?

I think 5000 is too small. Do I need to increase vocab size ?

Reinforce-Paraphrase-Generation/src/config.py

Line 29 in 7c788fd

vocab_size= 5000

How to preprocess the data?

Hi,
How to preprocess a user's dataset, to get the chunked data?
Thanks,
Xiao

Is it okay to use teacher forcing in evaluation?

Hi Wu!

I'm studying conditional text generation using your code. Thanks for publishing such a great code.

while reading eval.py, I wondered whether it is allowed to use teacher forcing in evaluation mode, since evaluation process, I think, was supposed to be the same as the model's decoding process. (including beam search)

On the other hand, I suspect it is okay because we use teacher forcing in the training mode. That is, we aim to evaluate its performance in the same environment as the training mode.

Which one do you think is the convention? Thanks in advance!

Dependencies for the code.

Hi,
Could you specify what dependencies could be used for the code? or where I can find them? e.g. The version number. Thanks very much.

Xiao

GPU利用率低

首先，感谢作者大大的杰出贡献。

但是运行代码的过程中，GPU占用率总是在10%+到30%之间跳动，即使将batch size设置为128也不管用。请问这可能是什么的原因？谢谢！