Training Chatbots to Be Polite: A Creative Approach

Have you ever thought about teaching a robot to be polite? It's not as easy as it sounds, especially without specific examples to guide them. That's what a team of researchers set out to do. They developed three innovative models to create chatbots that can be either polite or impolite. One model, called Fusion, merges the chatbot's core with a separate polite language model. Another, named Label-Fine-Tuning (LFT), uses a politeness score to teach the bot better responses.The third model, Polite-RL, rewards the bot for its good manners. They also tested two simpler models that select polite responses from a pre-existing database. Humans were asked to evaluate these conversations. While Fusion and the simpler models could produce polite responses, they often missed the mark in terms of relevance. LFT and Polite-RL, on the other hand, managed to be both polite and on-topic.