Repository for dialogue models which enhance response diversity or coherence, coded in Pytorch.
Contains shared code for multiple concurrent projects. First, this repository contains code for the processing of Reddit comments for use in training a dialogue model. Second, this repository contains code related to training a baseline sequence to sequence model (run_seq2seq.py). Finally, run_mismatch.py trains a classifier which diagnoses response coherence with respect to a dialogue history. With the --mismatch_path flag, the sequence-to-sequence can train using learned coherence features from the coherence classifier.