Context Recognition in Language Cornprehension
Eugene Charniak Brown University
People can read a line or two of a story and immediately know "what is going on." So in reading
As the boy walked down the aisle he took a can of tuna fish from the shelf and put it in his basket.
we have no trouble deciding that we are witnessing a person shopping at a supermarket. When I talk of "context recognition" this is what I mean. The question I wish to raise is: How do we do this?
As we see in the course of this chapter, there are many aspects to this problem, but we concentrate on the most obvious ones, coming up with a "supermarket" hypothesis in a story that never mentions any such thing. Superficially, this ability requires that we combine "low-level" clues (tuna fish, shelf, basket, aisle) to form a "high-level" hypothesis (supermarket). We might then ask if there are any limits to this ability, or are we able to take any set of clues, and find a consistent hypothesis that integrates them, assuming such a hypothesis exists. At a glance this seems plausible, because if the reader tries, he or she could probably make up dozens, if not hundreds, of examples, all of which imply supermarket shopping, yet use new combinations of different clues.
But are people really capable of all this? In this chapter I argue that such an assumption will have very far-reaching consequences, many of which make me uncomfortable. Instead I propose an alternate model of this ability, which makes much more modest processing assumptions, and which, in particular, assumes