Please be polite: Towards building a politeness adaptive dialogue system for goal-oriented conversations

Politeness embalms interactions, putting the parties in conversation at ease. Even the most well-intended communication can fall through if there is a manifestation of rudeness. Complementarily, even the most angst-prone situation can be communicated with much less hurt if lathered with politeness. In this paper, we propose a novel task named as Politeness Adaptive Dialogue System (PADS) to incorporate politeness feedback of agents’ actions according to user’s mood and demands. As there is no politeness labeled conversational data available, we annotate the recently released MultiDoGO dataset having six domains with appropriate politeness labels. The proposed end-to-end dialogue system comprises of a transformer-based politeness classifier that interacts with a reinforced learning framework with four different politeness-oriented reward algorithms, forcing the agent to adapt to polite actions upon encountering the user’s dissatisfaction in the dialogue. We train and evaluate PADS by building six different user simulators, each corresponding to a domain. Quantitative and qualitative analysis show that our proposed politeness based reward algorithms improve the task success rate and reduce the dialogue length with state-of-the-art results.