We adopt assumptions consistent with Gricean maxims of conversation, thereby mapping the problem into a simple Bayesian framework that allows us model the behavior of the speaker. This framework has been shown to give us the advantage of being able to systematically incorporate evidence from various knowledge that a listener may acquire during conversation. In particular, knowledge about speaker's beliefs and knowledge about focus are shown to be easily incorporated.
Now the whole point of search (as opposed to just picking whichever child looks best to an evaluation function) is to insulate oneself from errors in the evaluation function. When one searches below a node, one gains more information and one's opinion of the value of that node may change. Such "opinion changes" are inherently probabilistic. They occur because one's information or computational abilities are unable to distinguish different states, e.g. a node with a given set of features might have different values. In this paper we adopt a probabilistic model of opinion changes, de-1This is a super-abbreviated discussion of [Baum and Smith, 1993] written by EBB for this conference.
This class derives from link grammar, a context-free formalism for the description of natural language. We describe an algorithm for determining maximum-likelihood estimates of the parameters of these models. The language models which we present differ from previous models based on stochastic context-free grammars in that they are highly lexical. In particular, they include the familiar n-gram models as a natural subclass. The motivation for considering this class is to estimate the contribution which grammar can make to reducing the relative entropy of natural language.
Recently, there have been a number of efforts to cast natural language understanding in a probabilistic framework. The argument that there is a probabilistic element in natural language interpretation is essentially the following: Most inferences or interpretation decisions cannot be made based on logical criteria alone. Rather, evidence of varying strengths needs to be combined, and probability theory offers a principled basis for doing so. For example, consider the following story1: (1) John got his suitcase. He went to the airport. The interpretation that John is a terrorist intent on blowing up an airplane, say, is not favored by human readers, and is probably not even considered by them, although structurally it is identical to *This work was supported by Nation Science Foundation grant IRI-9123336.
We consider the problem of enabling an autonomous We would like to build a robot which can, for a reasonable class of environments, fulfill tasks of the form "go mobile robot to get from one specified point in its environment to another. The solution to this problem to z and get /," or perhaps "take z to z." We would requires the robot to build the representations of its like to be able to specify x, Z/, and z in the same form environment to enable it to determine its current location, its destination, and how it can get from one We realize that we will not soon be able to specify to our robot that we would to another human being.