Operator: Learn-comprehension

Problem space: (Virtual-)Top-space
Operator Overview:
The learn-comprehension (or LC)operator is a (pseudo-)top-space language operator that provides the option of deliberately deciding to learn to comprehend a language input even though there may be other operators available. This is seen as an attractive alternative to the option of creating a deliberate impasse, since causing a deliberate impasse requires the construction of a dummy operator without semantic content. This operator is constructed instead, and it can be processed by the decision procedure with a full understanding of what its selection entails.

Operator Proposal:
Proposed when there is a sentence attribute on the top state and when the u-model-success annotation is not being made by NL comprehension.

Operator Application:
This operator is never actually applied; the impasse created causes the NL comprehension spaces to take over processing and, if successful, produce a proposal for one or more new comprehension operators.

Operator Reconsider:
Because the Create-operator space (which responds to the impasse on learn-comprehension) copies and walls off the language state during learning, the conditions for reconsidering the LC operator lead to the "blippability" of the comprehension process. One opportunity that is always available is the possibility of reconsidering the LC operator when a constructor operator is made acceptable by the Create-operator space ("blipping from below" or, less idiomatically, resolving the impasse).

"Blipping from above" is accomplished by the addition of search control that causes a reconsider whenever there is a proposal for another LC operator for a word later in the sentence than the one that led to the proposal of the current one. Such search control knowledge will generate a reconsider preference for the selected, blippable LC operator. This has the effect that if a new word arrives before the impasse terminated naturally (by producing a constructor operator), the impasse will be terminated prematurely. It is possible that the LC operator that was previously in the slot will be reselected, but the "blipping" will have caused the stack to collapse in any case.

A consequence of designing comprehension to be blippable is that it creates a learning/training trade-off. When words are coming into the phonological buffer in real-time there is rarely enough time to learn from scratch and the system will generally be blipped from above before a constructor can be completed. Since chunking occurs throughout the goal stack, however, a second (or third...) run over the same data will take advantage of the chunks created in lower spaces to build a top-level constructor. Alternatively, data can be run in a stand-alone version of NL-Soar where time can be stretched to allow for the learn-comprehension to self-terminate.

Productions are in file:

Back to the operator hierarchy.

This page written by Jill Fain Lehman (jef@cs.cmu.edu) and Greg Nelson (ghn@cs.cmu.edu)