I see the biggest opportunity/challenge around inferring conceptual metaphors and narrative patterns in a text. There was an earlier DARPA BAA around that sort of thing. Narrative/metaphor is somewhere between affect understanding (sentiment analysis) and true semantic understanding. It is low-order semantic approximation that goes after the "most significant bits" in a semantic sense. For example, I'd like to feed a program stories in classic Aristotlean 3-act form and have it return act boundaries and a tragedy/comedy classification, but not require a gazillion examples to curve-fit a classifier. I'd like to feed a screenplay to a program that automatically segments it in terms of Campbellian monomyth terms (correctly identifying archetypes for example).

It's what I'd be working on if I hadn't entirely lost my appetite for traditional research funding/delivery models. I feel like putting a gun to my head at the very thought of writing SBIR/STTR proposals to get money or papers/patent apps as outputs.

Outside of its Grand Challenges, DARPA seems to have lost its way. They announce programs that _seem_ to aspire to ambitious, disruptive, ARPANET level results, but when you dig in, you find that they have something so narrowly specified in mind that it's clear just from the CFP that nothing even remotely interesting will come out of it.

That said, a few modestly interesting recent things like Siri did originate in DARPA research (CALO).

Philoshically for this domain, I don't like the idea of the sort of Big Data attacks that have proved useful for Siri, Watson etc. for deeper meaning, I think these will fail. At least if used alone.