Hacker News new | ask | show | jobs
by stephenroller 2485 days ago
We originally set out to see whether we could use conditional generation to sort of adjust the behavior of these neural generative dialogue systems. Much of the original idea was about sort of "planning" our response to improve the quality and diversity of neural generation.

As we studied though, we became more aware of how much nuance there was in each of these methods, especially when we started having humans chat with the models. In the end, the paper ended up being more of a thorough analysis into what behaviors are more preferred in dialogue agents. For example, we found that you should spend about ~60-70% of your turns asking questions to maximize engagement: enough to convey interest in your partner without coming off as super nosy.
2 comments
matchagaucho 2485 days ago
That's amazing. Do you have any hobbies?
link
stephenroller 2485 days ago
Yes, that's a great response for one of the models to have. :)
link
zitterbewegung 2484 days ago
How is the weather?
link
thomk 2485 days ago
How is the context defined here? Certainly when you are first getting to know someone 60% of your dialogue is reasonable for questions. After you know someone, 60% feels high. Speaking from one humans perspective of course, I may be wrong. :)

Most conversations are based on a pretext that individuals understand. In fact, if someone entering a conversation doesn't know that pretext it could awkward for them. That's why most people enter a conversation by saying something neutral like 'hello' and not by saying 'wow is the food horrible here!'. You never know who is listening. You wouldn't greet your boss the same way you greet your mother for instance (in most cases).

The tl/dr question here is 'Do you take any precondition (like relationship) into account before the parties begin chatting?'

link
stephenroller 2485 days ago
We limited the scope of our study to the PersonaChat dataset, which is a standard benchmark in dialogue. In PersonaChat, both participants are presented with Personas and asked to roleplay as that person. (e.g. "I love the beach; Horses are my favorite animal; I am on a diet now").

So we restrict ourselves to meeting someone for the first time and trying to learn about them and share a bit about ourselves.

The personas that were assigned to the bot and the human were both randomized (from a large, fixed set) and unseen in training data.

link
contingencies 2485 days ago
That's really cool. So, do you have any hobbies?
link