Here are some unstructured notes I’ve been collecting on conversational frameworks. I’ve been meaning to publish this as a more polished post for the past year, but haven’t gotten around to it quite yet.
I think we’re 5-10 years away from having machines that can pass the Turing test. Moreover, the reason why they haven’t passed them yet is because we haven’t taught machines about human games, status, self-presentation, and a concrete model of how humans develop.
We’re making a lot of progress in structured games where there’s a discrete action set at every state. Examples: AlphaGo, AlphaChess.
In self-play, we can train machines up against themselves. But in human conversation, we need some sort of framework that acts as a “free body diagram” of sorts to explain conversational and human dynamics.
You also need a lot of data. I can’t propose any definitive data set, but I think that these heuristics might help us leverage existing data sets so that we can train machines to better understand human interaction.
This is more an exploration - if you’re doing work in any of these fields please let me know and I’d be happy to provide a link to your work. Also, these are some notes I’ve collected, so it’s deliberately unstructured.
Eric Berne introduces the notion of a game in his seminole works Games People Play. What is a game?
We play games all the time—sexual games, marital games, power games with our bosses, and competitive games with our friends. Detailing status contests like “Martini” (I know a better way), to lethal couples combat like “If It Weren’t For You” and “Uproar,” to flirtation favorites like “The Stocking Game” and “Let’s You and Him Fight,” Dr. Berne exposes the secret ploys and unconscious maneuvers that rule our intimate lives.
I was surprised when I first read this book, because I could pretty accurately describe a lot of interactions with important people in my life by one of the games he outlines. There are details that vary in each game, but the general archetype and structure remains the same.
In order to understand a game, Berne introduces the notion of super-egos. In a reductionist view, you can model an interaction between two entities as a stimulus and response (let’s not get too Skinnerian here):
Berne believed that insight could be better discovered by analyzing patients’ social transactions. Berne mapped interpersonal relationships to three ego-states of the individuals involved: the Parent, Adult, and Child state. He then investigated communications between individuals based on the current state of each. He called these interpersonal interactions transactions and used the label games to refer to certain patterns of transactions which popped up repeatedly in everyday life.
With this framework, we’re able to summarize a lot of one-on-one human interaction with a 2x2 grid:
Me: I’m ok / I’m not okay.
You: You’re ok / you’re not okay.
I’d be curious to see how long it takes machines to understand this dichotomy.
One of my all-time favorites. I had to read it twice before I really understood and appreciated it. I think Johnstone’s framework describing status is relevant here:
If status can’t even be got rid of, then what happens between friends? Many people will maintain that we don’t play status transactions with our friends, and yet every movement, every inflection of the voice implies a status. My answer is that acquaintances become friends when they agree to play status games together.
Suddenly we understood that every inflection and movement implies a status, and that no action is due to chance, or really ‘motiveless’.
Normally we are ‘forbidden’ to see status transactions except when there’s a conflict. In reality status transactions continue all the time. In the park we’ll notice the ducks squabbling, but not how carefully they keep their distances when they are not.
In order to enter a room all you need to know is what status you are playing.
Johnstone talks extensively about the notion of status, specifically how it relates to acting. One larger takeaway, though, is that most of what humans do is some form of improvisational acting, so it’s far more relevant than one might think. I couldn’t find a concrete definition of status, probably because it’s one of those things that you likely just “get” from being human.
I think the notion of a machine discovering status through unsupervised learning is difficult, at least for now. Humans have enough trouble doing it; if they didn’t, everybody would probably be a good actor. But what if there was some data set that detailed conversation and the status conversations that came as a result of it?
I believe this would be a case where “end-to-end” deep learning might be hard, but a labeled dataset could really do wonders and teach machines how to communicate with us.
Is life just theatre?