17.5.6 Pitfalls

As noted in Section 17.4, there are a number of common mistakes in representing the meaning of natural language utterances, that arise from confusing, or equating, elements from real languages with elements in FOPC. Consider the following example, which on the surface looks like a candidate for a standard implication rule:

(17.82) If you're interested in baseball, the Rockies are playing tonight.

A straightforward translation of this sentence into FOPC might look something like this:

$$ \begin{aligned}Have Interest In(Hearer,Baseball)\\\Rightarrow Playing(Rockies,Tonight)\end{aligned} $$

This representation is flawed for a large number of reasons. The most obvious ones arise from the semantics of FOPC implications. In the event that the hearer is not interested in baseball, this formula becomes meaningless. Specifically, we can not draw any conclusion about the consequent clause when the antecedent is false. But of course this is a ridiculous conclusion, we know that the Rockies game will go forward regardless of whether or not the hearer happens to like baseball. Exercise 17.10 asks you to come up with a more reasonable FOPC translation of this example.

Now consider the following example:

(17.83) One more beer and I'll fall off this stool.

Again, a simple-minded translation of this sentence might consist of a conjunction of two clauses: one representing a drinking event and one representing a falling event. In this case, the surface use of the word and obscures the fact that this sentence instead has an implication underlying it. The lesson of both of these examples is that English words like and, or and if are only tenuously related to the elements of FOPC with the same names.

Along the same lines, it is important to remember the complete lack of significance of the names we make use of in representing FOPC formulas. Consider the following constant:

InexpensiveVegetarianIndianFoodOnTuesdays

Despite its impressive morphology, this term, by itself, has no more meaning than a constant like X99 would have. See McDermott (1976) for a discourse on the inherent dangers of such naming schemes.

As we noted at the beginning of this chapter,

Basics and current applications/versions OWL as a kind of description logic as applied to the Semantic Web. See the Brachman thing.

basically we're going to arrange the concepts in a domain into a hierarchy. Then we're going to relate the elements in the hierarchy via type slot-filler relations (value/resritictions). Then there's inheritance.

unary and binary predicates only? I.e. what you get in a network.

17.6 RELATED REPRESENTATIONAL APPROACHES

Over the years, a fair number of representational schemes have been invented to capture the meaning of linguistic utterances for use in natural language processing systems. Other than First Order Logic, the most widely used schemes have been semantic networks and frames, which are also known as slot-filler representations. The KL-ONE (Brachman and Schmolze, 1985), and KRL (Bobrow and Winograd, 1977) systems were influential efforts to represent knowledge for use in natural language processing systems.

In semantic networks, objects are represented as nodes in a graph, with relations between objects being represented by named links. In frame-based systems, objects are represented as feature-structures similar to those discussed in Ch. 16, which can, of course, also be naturally represented as graphs. In this approach features are called slots and the values, or fillers, of these slots can either be atomic

values or other embedded frames. The following diagram illustrates how example (17.76) might be captured in a frame-based approach.

I believe Mary ate British food.

$$ \begin{aligned}&\begin{bmatrix} \\{{{\mathrm{BELIEVING}}}} \\{{{\mathrm{BELIEVER}}}}&{{{\mathrm{SPEAKER}}}} \\{{{\mathrm{BELIEVED}}}}&{{{\begin{bmatrix} \mathrm{EATING}}}} \\{{{\mathrm{EATER}}}}&{{{\mathrm{MARY}}}} \\{{{\mathrm{EATEN}}}}&{{{\mathrm{BRITISH FOOD}}}} \\\end{bmatrix}\\ &\end{bmatrix}\\ \end{aligned} $$

It is now widely accepted that meanings represented in these approaches can in principle be translated into equivalent statements in FOL with relative ease. The difficulty is that in many of these approaches the semantics of a statement is defined procedurally. That is, the meaning arises from whatever the system that interprets it does with it.