Course: Modeltheoretic Semantics and the Syntax-Semantics Interface
Eddy Ruys and Yoad Winter
Topics: types, domains, characteristic functions, function application, the category to type matchingIn the first lecture we have talked rather sloppily about models and various sets, functions and truth-values they contain. In this lecture we will go somewhat deeper into the structure of the model: what objects are there in the model, how can we describe them, and does this give us more insight into the relations between syntax and semantics?
As we have said, natural language expressions are taken to denote objects in the model. However, different expressions denote different kinds of things. For instance, we let Tina denote a member of the set of entities E, whereas tall denotes a subset of E and a sentence like Tina is tall denotes a truth value. We would like to have a systematic way to describe these distinctions. For this reason we divide the model into different domains, where each domain has a type: a label describing its internal structure. The three expressions mentioned above denote members of domains with different types.
The simplest types and the simplest domains are the primitive ones. There are two primitive types, each with its corresponding domain. One primitive type is called e (for entities) and its corresponding domain is the non-empty set of entities E, which we from now on denote by (the e Domain). The other primitive type is called t (for truth-values) and its corresponding domain is the set . The proper name Tina takes its denotation from the domain. In short, we say that proper names are of type e. A sentence like Tina is tall denotes a truth value in .
All the non-primitive types are recursively defined on top of these two primitive types. To see how, let us start with an example. We have said that predicates like tall denote subsets of . For instance, consider a model where the set of entities is . Let us indicate that only Tina and John are tall in this model. We can do that by letting tall refer to the set , a subset of . Alternatively, we can say the same thing in a slightly different way: let tall, instead of denoting the set T, denote a corresponding function from entities to truth-values, i.e. from to . For each member of the set of entities, assigns it true if it is in T and false if it is not in T. Thus, we have:
We obviously want predicates like tall to be able to denote any subset of , or, as we now say, any function from to . Thus, we want the domain from which tall takes its denotation to be the set of all functions from to . In set theory this set of functions is denoted by . This domain will be of type (et), for which we will often use the notation et as an abbreviation, erasing parentheses whenever this does not cause confusion. Here are the eight members in for the tripleton domain we are dealing with:
To summarize, a proper name like Tina denotes an element of , a sentence denotes an element of and a predicate like tall denotes a function from the first domain to the latter, i.e.\ an element of . Of course, when we consider more expressions, we are likely to need more types than these ones. How do we define them?
What we obviously want is a general strategy for defining types.
The tall example has illustrated the way to do that: we defined
a new type, et, and its corresponding domain on the basis of the given types
e and t and their domains.
Generally, if we are given
the types and and
their corresponding domains and , we can define
the new type of a domain that consists of
all the functions
from to .
In this way we can go on and use the new type and domain we get
to create more new types and domains.
For instance, we can use et and e to create the type (et)e and its
domain
of all functions from to .
There are infinitely many types we can define using this inductive
procedure.
This is precisely what we do now,
using the following two definitions.
These two simple definitions summarize all we have discussed above.
A model
is the infinite collection
of all typed domains. Note that the primitive domain is given, so
in order to determine a model, all we need to do
is to determine the other primitive domain: . After we have done this, all
the compound domains are inductively fixed by definitions 1 and 2.
Working with functions in this way as the building blocks of compound domains gives us an easy way to describe what is going on compositionally in analyzing a sentence. Consider a maximally simple sentence like Tina smiles. The sentence denotes a truth value simply because the subject denotes an entity of type e and the predicate denotes a function of type et: from e entities to truth-values. Once this function applies to the subject denotation we get a truth value. This operation of function application is crucial: it is the main operation that ``glues'' denotations together into new denotations. Let us exemplify the notation we use in more detail for the simple sentence Tina smiles:
Note how easy it is to know from two types if function application is allowed between objects in the corresponding domains. When function application is allowed also the outcome type is obvious. For instance:
The postulation of function application also facilitates the typing (=the determination of the type) of different expressions. For instance, consider the auxiliary verb is in the sentence Tina is tall. We have decided to let Tina get the type e and tall get the type et. The constituent is tall has to denote a function that applies to and derives a truth-value. Hence, it must be, like tall, of type et. To get this, the denotation of is will be a function of type (et)(et): when it applies to an et object like it will give us an et object. In fact, is will denote the identity function on predicates: the function that for each et predicate assigns the same predicate itself. Thus, the denotation of is tall is the same as the denotation of tall.
These typing considerations are summarized in figure 1. Note that now it is natural to use a more realistic structure than the one that was tentatively used in the first lecture for expository purposes.
Figure 1: compositionality and typing
What we have done above is a kind of puzzle solving. We were given parts of the puzzle: the types for the denotations of the words Tina and tall and the type for the whole sentence. We had the structure of the sentence and sought types for the word is and the category PRED that would match the other types by allowing function application to work compositionally. Because we knew that the denotation of PRED would have to apply to an e denotation to produce a t denotation, we concluded that this category should be of type et. To generate this type, the denotation has to apply to the et type denotation of tall and give another et function. Thus, it has to be of type (et)(et).
We should stress the relations between typing and sentence structure. For instance, the (et)(et) typing of is is motivated by the structure in figure 1, but not necessarily by different structures that might have been postulated. Suppose tentatively that we gave the sentence the implausible structure of figure 2, in which Tina and is are assumed to form a constituent. In this structure the typing of is could not have remained (et)(et): functions of this type apply to et type objects and not to e type entities. A possibility would be to change the typing of is to ee: a function from entities to entities. This would allow the denotation to apply to and produce an e type denotation for the category FUNNY, to which the predicate could apply and produce a truth value.
This simple example shows that in assigning a type to a particular expression, structural considerations play a major role: the syntactic configurations in which an expression appears often affect the semantic type we choose to assign to it.
So far, types have just put some order in things that were already done in the previous lecture. But the idea to let expressions denote typed objects is very helpful when we want to define denotations of new expressions. For instance, what does the transitive verb see denote? Consider a sentence like Tina saw Mary. Compositionally, the constituent saw Mary is going to have the type et: like intransitive predicates (e.g. tall) its denotation applies to entities to produce a truth value. Since we have decided to let Mary be of type e there is no question as for what type the verb see has to get: the type of functions from e objects to et objects. In short: e(et). Similarly, a ditransitive predicate like show, as in Tina showed Mary John denotes objects of type e(e(et)): functions from entities to functions from entities to functions from entities to truth values. It is surely convenient that we have so many types at our disposal!
Note that in the discussion above we have made two implicit assumptions:
Figure 3: possibilities with function application
Such a strict organization of the grammar is challenging, and therefore theoretically desired: it leaves us with only a small amount of possibilities to develop syntactic theories with suitable typings of lexical items. At the present stage of semantic theory, however, it is debatable whether such a strict configuration can succeed as a linguistic program. These restrictions may be too strict for natural language, and attempts are being made to relax them in principled ways. Nevertheless; since these two constraints are simple and methodologically convenient, we will adopt them throughout the first part of this course as an illustration of the power of modeltheoretic techniques.
A note on characteristic functions: We have shown above that we can freely switch between a set denotation of a predicate like tall and the corresponding characteristic function of type et. The situation is in fact more general. Consider the transitive predicate see. Intuitively, such a predicate can be viewed as a two-place relation between entities: a subset of the cartesian product . The relation will specify which entities see which entities. For instance, in case , we can describe a situation like the one given in (1) by letting see denote the relation S in (2).
Note that is itself a characteristic function. In our case, the one characterizing the set of entities that saw x. The successive application is sometimes written in short . This switch in the intuitive order of the arguments is confusing upon first encounter. However, it is necessary if we want to be able to speak of the denotation of constituents like saw Mary: it is much more natural then to think of Mary as the first argument of the verb's denotation and not as in the intuitive relation S. After some exercising you will surely get used to this and realize that the method is completely innocuous. Moreover, the procedure we have used for moving back and forth between the intuitively convenient sets/relations and their corresponding characteristic functions is easily extended to many other useful types.
Exercises