17.1.3 Canonical Form

The notion that single sentences can be assigned multiple meanings leads to the related phenomenon of distinct inputs that should be assigned the same meaning representation. Consider the following alternative ways of expressing example (17.1):

(17.4) Does Maharani have vegetarian dishes?

(17.5) Do they have vegetarian food at Maharani?

(17.6) Are vegetarian dishes served at Maharani?

(17.7) Does Maharani serve vegetarian fare?

Given that these alternatives use different words and have widely varying syntactic analyses, it would not be unreasonable to expect them to have substantially different meaning representations. Such a situation would, however, have undesirable consequences for our matching approach to determining the truth of our representations. If the system's knowledge base contains only a single representation of the fact in question, then the representations underlying all but one of our alternatives will fail to produce a match. We could, of course, store all possible alternative representations of the same fact in the knowledge base, but this would lead to an enormous number of problems related to keeping such a knowledge base consistent.

The way out of this dilemma is motivated by the fact that since the answers given for each of these alternatives should be the same in all situations, we might say that they all mean the same thing, at least for the purposes of giving restaurant recommendations. In other words, at least in this domain, we can legitimately consider assigning the same meaning representation to the propositions underlying

each of these requests. Taking such an approach would guarantee that our matching scheme for answering Yes-No questions will still work.

The notion that inputs that mean the same thing should have the same meaning representation is known as the doctrine of canonical form. This approach greatly simplifies various reasoning tasks since systems need only deal with a single meaning representation for a potentially wide range of expressions.

Canonical form does, of course, complicate the task of semantic analysis. To see this, note that the alternatives given above use completely different words and syntax to refer to vegetarian fare and to what restaurants do with it. More specifically, to assign the same representation to all of these requests our system will have to conclude that vegetarian fare, vegetarian dishes and vegetarian food refer to the same thing in this context, that the use here of having and serving are similarly equivalent, and that the different syntactic parses underlying these requests are all compatible with the same meaning representation.

Being able to assign the same representation to such diverse inputs is a tall order. Fortunately there are some systematic meaning relationships among word senses and among grammatical constructions that can be exploited to make this task tractable. Consider the issue of the meanings of the words food, dish and fare in these examples. A little introspection, or a glance at a dictionary, reveals that these words have a fair number of distinct uses. Fortunately, it also reveals that there is at least one sense that is shared among them all. If a system has the ability to choose that shared sense, then an identical meaning representation can be assigned to the phrases containing these words.

In general, we say that these words all have various word senses and that some of the senses are synonymous with one another. The process of choosing the right sense in context is called word sense disambiguation, or word sense tagging by analogy to part-of-speech tagging. The topics of synonymy, sense tagging, and a host of other topics related to word meanings will be covered in Chs. 17 and 18. Suffice it to say here that the fact that inputs may use different words does not preclude the assignment of identical meanings to them.

Just as there are systematic relationships among the meanings of different words, there are similar relationships related to the role that syntactic analyses play in assigning meanings to sentences. Specifically, alternative syntactic analyses often have meanings that are, if not identical, at least systematically related to one another. Consider the following pair of examples:

(17.8) Maharani serves vegetarian dishes.

(17.9) Vegetarian dishes are served by Maharani.

Despite the different placement of the arguments to serve in these examples, we can still assign Maharani and vegetarian dishes to the same roles in both of these

examples because of our knowledge of the relationship between active and passive sentence constructions. In particular, we can use knowledge of where grammatical subjects and direct objects appear in these constructions to assign Maharani, to the role of the server, and vegetarian dishes to the role of thing being served in both of these examples, despite the fact that they appear in different surface locations. The precise role of the grammar in the construction of meaning representations will be covered in Ch. 18.