Re-Implementation of EMNLP 2018 paper Paraphrase Generation with Deep Reinforcement Learning
Still under construction.
- Created parse_data.py to process Quora dataset into tokens and .bin files (tf.Example) files for generator to read
- To train evaluator:
- Fix train.py to take in the data we want - no need to read_corpus, we just need to have pairs of (sent1, sent2, label) --> where label is 0 or 1
- Embeddings: Use the ones from the paper, there should be a function to load them
- Fix train.py to take in the data we want - no need to read_corpus, we just need to have pairs of (sent1, sent2, label) --> where label is 0 or 1
- To pre-train generator()
- Fix batcher.py
- modify text_generator() to extract question 1 and question 2 from tf.Example file IFF they have a positive label
- modify fill_example_queue() --> Possibly just name change of (article, abstract)
- Fix batcher.py
- Try generating one sentence with initial generator
- Write the whole RL architecture:
- Find a good monte carlo simulator library for this task
- Function to compute value
- Make function to rescale Q value
- Make function to update gradient
- Steps 1-3 from RbM-SL
- Calculate gradient according to algorithm
- Write method to train evaluator