Or, consider a narrative example. In the story Men Are Different, the reader is presented with a first-person narrative. It is acceptable to assume that the narrator is a human. This is reinforced by the narrator informing the reader that he is an archaeologist. On the other hand, the fact that the narrator is studying the ``lost'' civilization of Mankind leads the reader to suspect that their understanding of narrator-as-human is flawed; problem reformulation is needed to see if any other explanation ``fits'' the available data. In this case, the narrator turns out to be a robot, so another explanation is actually better than the original one.
For a reasoner, there are two ways in which this problem reformulation can occur. The first mechanism depends on the fact that I am studying reading. There is a communicative agreement between the author of a text and the reader--that is, the author is expecting the reader to be able to comprehend what is being presented in the text. The author may intentionally delay the presentation of all the relevant facts (consider a mystery story, for example), but if the reader has the prerequisite knowledge, they should be able to eventually comprehend the text. This allows a conservative approach to handling problem reformulation. If a text can be assumed to always present enough information to allow a ``proper'' understanding to occur (if one is possible), a reasoner can simply delay problem reformulation until it is explicitly indicated. This indication will occur when the current understanding fails. In the phaser example, seeing someone point it at a television, immediately followed with the television switching to a new channel, is an explicit cue that something is flawed--the current understanding cannot explain this behavior. In Men Are Different, for instance, the narrator continually talks about lost Mankind and eventually describes itself as a robot. These explicit text pointers indicate to the reader that the initial narrator-as-human understanding is flawed; problem reformulation must occur. The problem reformulation algorithm is shown in Figure 28.
The original concept is given to the task, along with the specific understanding failure which occurred. This understanding failure will have resulted from a prediction, abduction, or explanation provided by the original concept failing to be true in the current context of the story. For example, if the reader is understanding a lycanthrope in a story as the ``standard'' magical creature, then the prediction made at the time of a transformation will be for a wolf-like creature to result. If, as in the story Lycanthrope, a car is the result, this represents an understanding failure. These two pieces of information, the original concept and the specific failure, allow memory retrieval to be called to return a set of possible candidates to act as the reformulated concept. These are then checked to ensure that they maintain consistency with the other uses of the original concept. As soon as a concept is discovered which fills both roles (i.e., explains the immediate failure and explains the prior uses of the concept), it is returned as the reformulated concept to be understood.
Unfortunately, in the ``real'' world, it will not be possible to guarantee that problem reformulation will always be triggered explicitly. In a story, however, the reader assumes that needed information will be provided to them at some point by the author. The author provides this information as their part of the communicative agreement which exists between the author and the reader. Also, a story is a very dense representation with respect to the concepts which are presented--there is not enough space in most stories to present irrelevant information. As Anton Chekov stated: ``If there's a gun on the wall in act one, scene one, you must fire the gun by act three, scene two. If you fire a gun in act three, scene two, you must see the gun on the wall in act one, scene one.'' Thus, as a result of the domain in which I have placed the creative understanding process, I am able to greatly reduce the effort which is required for this important task of the process. In other words, THE READING PROCESS CONSTRAINS THE PROBLEM OF IDENTIFYING UNDERSTANDING FAILURES, THEREBY ALLOWING PROBLEM REFORMULATION TO WORK.
However, it is possible to suggest a more general-purpose technique which is the second mechanism by which problem reformulation can occur. This other mechanism to consider is spontaneous retrieval ([#!memory:ram-francis-1996!#]). As the memory system is working in parallel with understanding, previous indexes and cues still active may trigger a new memory which can then interrupt the process and ``force'' a problem reformulation to occur. To illustrate this, consider the story Zoo. A fabulous, intergalactic zoo visits the Earth each year. The reader eventually learns that the ``creatures'' in the zoo are intelligent beings which view themselves as going on a great adventure to many foreign worlds; they are in the cages in order to protect themselves from the dangerous natives. One way to understand the novel use of zoo is to perform an analogy between the standard idea and the novel one. Another possibility, however, relies on problem reformulation via spontaneous retrieval. As the story is being read, memory retrieval is continuing based on the original specification. If the memory system is given enough time, it will return the concept of safari in addition to the original concept of zoo (assuming that these concepts exist in the memory system; it is also possible that the safari concept will never be returned). If this retrieval occurs, the reasoning supertask has been informed that the concept of zoo is not necessarily the best concept any longer. Problem reformulation is triggered as a result.
There are other approaches to problem reformulation which differs from my technique. Most of these, however, exist in problem solving or design domains and are not directly applicable to my work with understanding. For example, work on this topic has been done by Jones (creat:jones). His thesis is that the reformulation should be from an actual problem to a more abstract description; in particular, a primary source of culturally shared abstract knowledge exists in the form of proverbs which reasoners can make use of in solving novel problems. In order to accomplish this transformation, Jones formulates a set of viewing schemas which control how redescription of a problem is performed and to what extreme it is taken. Since Jones' work is in problem solving, he does not have the advantage of the communicative agreement to assist in the problem reformulation process. My task of problem reformulation is more simple than his because I can rely on that aspect of my domain.
