Figure 8:

Side-by-side comparison of chatbots with multiple dialog management approaches for the same starting meal description for the goal to “Eat no more than 2 portions of carbs (30g).” The Random chatbot dialog (a) continues for 2 more turns, and is longest of the four dialogs. The Scripted dialog (b) consistently asks the same 3 follow-up questions for every meal, based on the user’s goal. In this example, the Rule-based (c) and RL (d) dialogs include similar follow-up questions, with the exception that the rule-based dialog includes at least one “Search” question (“What else will you put in your cereal?”), whereas the RL chatbot does not. Note that while the dialogs start with the same seed meal description, each dialog is continued by a different crowd worker, so the final description of meals diverge by the end of each dialog.