Polite Chatbots: A New Approach to Dialogue Generation

Creating chatbots that respond politely is tough, especially when you don't have examples showing how to do it. We've come up with three clever models that can generate polite (or rude) responses without needing perfect training data. Our first model, called Fusion, combines a dialogue model with a language model that's been trained on polite sentences. Another model, Label-fine-tuning (LFT), adds a politeness score to each conversation and adjusts it during tests to get different levels of politeness. Our third model uses reinforcement learning, or Polite-RL, to reward politeness in responses. We also made two simpler models that find polite responses in a database. When we tested these models with humans, we found that Fusion and the database models were polite but sometimes missed the point of the conversation. The LFT and Polite-RL models, however, were both polite and stayed on topic.