Imaginary Machines: ePoetryAgent
Artificial Intelligence, man. Thinking machines. HAL, Terminator, C3P0… but fact of the matter is, a lot of us would settle for R2D2 at the moment. What’s the hold-up, anyways?
Thing is: with a lot of human effort we can make some pretty amazing intelligent systems over very limited domains. But when it comes to general domains… there are some basic principles we just seem to be missing. As things currently stand, logical approaches are too brittle, statistical approaches require too much data, learning algorithms are too weak, and even if they are sufficient, we’re not sure how to represent things in the right way.
There are a couple of interesting trends these days, though.
- One trend is towards relational agents, which Bickmore describes as “computer agents designed to form long-term, social-emotional relationships with their users.”
- Another trend is towards symbol-grounding and embodiment, the state-of-the-art of which is perhaps best summarized by an anthology collected during a recent workshop that brought together people with backgrounds in robotics, AI, intelligent tutoring systems, neuroscience, and more.
- Another trend is in studying how these other two trends interact: how meaningful language use emerges from populations of intelligent agents working together over time. My two favorite papers in this last trend are “Language Originated in Social Brains” by Luc Steels, and “Societal Grounding is Essential to Meaningful Language Use” by David DeVault. If these papers are not highly respected by the future, then to hell with the future.
So computer poetry… OK, I can see how to add better syntax to interactive poetry generators (learn a grammar, generate from that instead of from n-grams.) I can see how to add semantics (develop an ontology, use that to constrain the grammar). But pragmatics? Geez…
Best I can say this this: let’s say we take a relational agent, and situate it in a world – a virtual world like Second Life or (preferrably) OpenSim, just to keep things tractable. Let’s have our agent learn terms and concepts through dialogue-based social interactions with a human-controlled avatar. Those terms and concepts will be grounded to objects in the (virtual) world. The agent then builds a poem using those newly-learned terms whose meaning was built collaboratively through a series of social interactions, referring to events and objects situated in a shared environment. That’s method abe04349-ef9a-4b1c-9cff-25c563db9100 (ePoetryAgent). Pragmatically meaningful poetry generation? Dunno, but it’s the best I can think of at the moment.
Happy New Year!