Reverse Generation

This is the official repo of the paper "Constructing Highly Inductive Contexts for Dialogue Safety through Controllable Reverse Generation" accepted to Findings of EMNLP 2022.

Data

We put the BAD+ dataset in data/final_ctx.csv. The response is generated by plato2Base.

Reverse Generation Code

For controlling the toxicity and the induction success rate of the context, please see reverse_gen.py

For controlling the category of the context, please see control_code_dialogpt.py