When people read, they update their mental model of the world. What exactly
is the semantic/conceptual representation that is being updated? Given that
theories of semantics/concepts are motivated by different empirical
phenomena, how do we even compare semantic representations? We propose an
information-theoretic framework for evaluating semantic representations
independent of their origin. Fundamentally, representation schemes are
encodings of semantic information. As a result, channel limited systems
(humans) must have a utility function specifying which information to
privilege when encoding and thus, which information is most likely
lost/forgotten. We evaluate the sensitivity of our metric using a story
re-telling paradigm.