Chatbots: conversation as AI planning

I just love those YouTube videos of Internet-connected dolls. They're Wifi-linked to IBM's Watson, or some similar AI-in-the-cloud.

You'd think that the conversation would be scintillating - insofar as talking to a small child might have its moments.

But no, in the clip the adorable child is always shown asking the doll, "How far away is the Moon?" and the doll shows its conversational prowess by reciting Wikipedia.

Surely we can do better.

It's good to recall that we converse in order to achieve objectives. Sometime those objectives are indeed objective - as when we wonder how far away is the Moon.

Most of the time, though, the objectives are more about our more basic human desires. The need for sympathy, approbation, bonding against a third party, humiliation and revenge.

Oh dear, I've been listening to Cersei Lannister again..

Cersei and Margaery

Conversation is composed of speech acts directed towards goals, the goals often being the alteration of emotional states. Real-life chat is therefore very dependent on the accurate reading of emotional cues: facial expressions, tone of voice or other body language. No wonder email and WhatsApp chats go awry so often.

This suggests that a chatbot (a better Eliza) should treat conversation as an AI planning task.*

Naturally I'm not the first to realise that - this 1980 article - but it's the next place to go.

This relatively recent Stanford review - Conversational Agents - is quite illuminating (PDF slides).

---

Call me an old GOFAI-nostalgic but bottom-up number-crunching the stats is subject to diminishing returns as we move to higher cognitive tasks. Which is why I'm less than impressed by the tedious Google Assistant in Allo.

---

* I mean like GPS. Most conversational chatbots - like Google Assistant - cope with the extreme difficulties of unrestricted natural language input by giving a series of response-text buttons for the user to select ("Tell me a joke").

In a formalised conversational model, the chatbot should treat the conversation as a two-person game-tree. The opponent (user) has a choice of possible conversational rejoinders ('moves') which the program has a chance of understanding, and which propel the conversation forward in some direction.

Best to look at the planner conversation 'speech-act' rule-base and list the possible responses (need a way to flag input-variables) so as to guide the user.

Example:

Chatbot>: "What's your brother's name?

[I understand best:
   1. <first-name>
   2. <first-name> <second-name>
   3. <name> but we call him <nickname>
   4. I don't have a brother.
   5. Could we move on to another topic?
You can enter the response number if there's no information to give.]

User>:  Jimmie.