Learning Culture-Specific Dialogue Models from Non Culture-Specific Data

We build culture-specific dialogue policies of virtual humans for negotiation and in particular for argumentation and persuasion. In order to do that we use a corpus of non-culture specific dialogues and we build simulated users (SUs), i.e. models that simulate the behavior of real users. Then using these SUs and Reinforcement Learning (RL) we learn negotiation dialogue policies. Furthermore, we use research findings about specific cultures in order to tweak both the SUs and the reward functions used in RL towards a particular culture. We evaluate the learned policies in a simulation setting. Our results are consistent with our SU manipulations and RL reward functions.