I’d like some help figuring out a better strategy for a survey bot that takes very long open-ended responses. My current strategy isn’t working because it has (very) long training times and often triggers the fallback in undesirable ways.
Right now, I have a survey bot that asks about user experiences in a series of questions. The responses are expected to be quite long: they can be paragraphs in length, so very few will be similar.
At the moment, I have a series of actions that are triggered in a Rules sequence. A “inform” intent triggers the next action in the sequence. I’ve tried to train my bot with a few of these long answers, but it isn’t always great at recognizing an 'inform" intent from a long, rambling answer. I can add these in interactive mode, but the long answers in the NLU cause the training to be very long.
One reason I prefer a sequence of questions and answers in a series of actions rather than a form is that I connect these answers mid-conversation to an API that stores the response in an application.
I’ve had difficulty “forcing” the bot to go down a sequence of actions when it has difficulty recognizing an ‘inform’ intent.
It’s doubly difficult because the rotation of these sequences is difficult to predict, so if a user ends up answering a set of questions in an unusual order confuses the bot that relies strictly on known stories.
I’m happy to post more details about the bot (stories, rules, actions config) to get better feedback.