The input is twitter chat that contains questions and their responsed answers. I clean all messy characters. In training state, we forward questions and try to predict answers. The model is followed by

In sampling pharse, i use multinomial distribution to sampling next word.

• Python 2.7
• Tensorflow 1.xx
• I only test it on Ubuntu 16.04

All sequences is padded to be same fixed length by using PAD token, and all questions is reversed according to Sutskever et al., 2014

Q : [ PAD, PAD, PAD, PAD, PAD, PAD, “?”, “you”, “are”, “How” ]
A : [ GO, “I”, “am”, “fine”, “.”, EOS, PAD, PAD, PAD, PAD ]

I use tensorflow so if you haven't installed tensorflow yet, just following the link and install approriate tensorflow for your pc To train model

python seq2seq.py

Here i plot chart of cross-entropy loss. the loss is still going down at the end of training phase. It mean the model is learning somethings.

I show somes answers at iter

Q: if youre gonna fall asleep in class just stay home
A:
Q: this was my favorite play of his jumped off the screen
A:
Q: former president george hw bush said hell vote for hillary clinton according to sources close to bush
A:
Q: sweet he should be here in his staff meeting now
A: