QREFINE-PPO

This is implementation of QREFINE model described in "Generative Question Refinement with Deep Reinforcement Learning in Retrieval-based QA System" (https://arxiv.org/abs/1908.05604)

Datasets:

• Yahoo dataset:

https://ciir.cs.umass.edu/downloads/nfL6/

• Customer Service Userlog (CSU):

../data/Huawei/large_unique0.95_without_alluserlog

Rewards network:

• Wording Reward:

BA_Seq2Seq.py

• Answer Coherency Reward:

QA_similiarity.py

Baselines:

• Seq2Seq and Seq2Seq+ Attention and Seq2Seq + Bidirectional:

BA_Seq2Seq.py

• QREFINE_sen:

Sen_seqRL_training.py

• QREFINE_word:

Word_seqRL_training.py

• QREFINE:

Whole_seqRL_training.py

Question Answer Retrieval experiment: 

Retrieve_qa.py

Evaluation:

Bleu-1-2-3-4: Evaluation_Package/bleu
Meteor: Evaluation_Package/meteor
Rouge: Evaluation_Package/rouge