The Closed World Assumption (CWA) is a long-standing controversy in
logic and AI. Are we trying to model a world whose ontology (and other
aspects) are relatively fixed, or do we want to leave room for an
open, ambiguous, changing world, even if that means we can reliably infer less?
This controversy is reappearing in modern machine learning. We show
how many received notions in machine learning, like "labels" or
"classes" for data or "topics" in natural langauge processing,
implicitly rely on closed-world assumptions. We show how this
assumption limits the applicability of some popular techniques.
We introduce *lensing*, a mixed-initiative technique for representing
perspective or "point of view" in machine learning. A closed-world
machine learning technique is first applied to a corpus. Then human users
are asked for open-ended criticism of the output of the machine
learning. The results from this process are fed back into another
iteration of the mixed-initiative loop. By doing so, we can get the
inferential power of closed-world techniques, while assuring that they
are a better fit for the open-ended world in which people actually live.