The underlying goal is to find ways to represent and store knowledge. This is
complicated by the fact that, as we have already seen, knowledge occurs in both
declarative and procedural forms. Thus, representing knowledge is not merely
the representation of facts, but instead encompasses a much broader spectrum.
Whether a single scheme for representing all forms of knowledge will ultimately
be found is therefore questionable.
The problem, however, is not just to represent and store knowledge. The
knowledge must also be readily accessible, and achieving this accessibility is a
challenge. Semantic nets, as introduced in Section 11.2, are often used as a means
of knowledge representation and storage, but extracting information from them
can be problematic. For example, the significance of the statement “Mary hit
John” depends on the relative ages of Mary and John. (Are the ages 2 and 30 or
vice versa?) This information would be stored in the complete semantic net suggested
by Figure 11.3, but extracting such information during contextual analysis
could require a significant amount of searching through the net.
Yet another problem dealing with accessing knowledge is identifying knowledge
that is implicitly, instead of explicitly, related to the task at hand. Rather
than answering the question “Did Arthur win the race?” with a blunt “No,” we
want a system that might answer with “No, he came down with the flu and was
not able to compete.” In the next section we will explore the concept of associative
memory, which is one area of research that is attempting to solve this
related information problem. However, the task is not merely to retrieve related
information. We need systems that can distinguish between related information
and relevant information. For example, an answer such as “No, he was born in
January and his sister’s name is Lisa” would not be considered a worthy
response to the previous question, even though the information reported is in
some way related.
Another approach to developing better knowledge extraction systems has
been to insert various forms of reasoning into the extraction process, resulting in
what is called meta-reasoning—meaning reasoning about reasoning. An example,
originally used in the context of database searches, is to apply the closed-world
assumption, which is the assumption that a statement is false unless it can be
explicitly derived from the information available. For example, it is the closedworld
assumption that allows a database to conclude that Nicole Smith does not
subscribe to a particular magazine even though the database does not contain any
information at all about Nicole. The process is to observe that Nicole Smith is not
on the subscription list and then apply the closed-world assumption to conclude
that Nicole Smith does not subscribe.
On the surface the closed-world assumption appears trivial, but it has consequences
that demonstrate how apparently innocent meta-reasoning techniques
can have subtle, undesirable effects. Suppose, for example, that the only knowledge
we have is the single statement
Mickey is a mouse OR Donald is a duck.
From this statement alone we cannot conclude that Mickey is in fact a mouse.
Thus the closed-world assumption forces us to conclude that the statement
Mickey is a mouse.