gabriellin / drlparaphrase

Paraphrase Generation with Deep Reinforcement Learning

Re-Implementation of EMNLP 2018 paper Paraphrase Generation with Deep Reinforcement Learning

Still under construction.

---

Updates:

• Created parse_data.py to process Quora dataset into tokens and .bin files (tf.Example) files for generator to read

TODO:

RbM-SL

1. 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
2. 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)
3. Try generating one sentence with initial generator
4. 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

RbM-iRL

1. Steps 1-3 from RbM-SL
2. Calculate gradient according to algorithm
3. Write method to train evaluator