Motivation

The npl language may be viewed as a proof of concept for the solution of a problem that has been present in logics since Gottlob Frege uncovered it. Here I will try to introduce this problem and my proposed solution.

Note that this “motivation” page is mainly for logicians. If you just want to use the language because it seems useful, you can (must!) leave this page and head on to the reference or the tutorial. And if you are a logician, please bear in mind that I am a self taught “logician”, so you should not expect the orthodox academic terminology of any established line of thought.

Historical introduction

Frege

Predicate logic was first developed by Gottlob Frege. He developed it in the belief that he was unraveling the foundations of the natural logic behind scientific theories (or behind any unambiguous text expressed in a natural language) (see this). He pursued a mathematical formalism capable of expressing any informal scientific theory. To that end, he developed predicate logic and, at the same time and on top of it, a particular formal theory which we may call “theory of concepts”. In this theory he had individuals (or values) and he had classes (or value-ranges), connected through class predicates (belongs, subclass). He also had an unlimited amount of other predicates (what he called concepts) (see here). With this, he very nearly had all the elements he needed to have a formal system with the same expressive power as the natural languages. With his class predicates he was able to represent the class relations in the natural languages, generally expressed in them through copular verbs; and with his concept/predicates he was able to represent the rest of relations expressed in the natural languages through non-copular verbs.

However, there was a problem with this scheme. His theory was second order, so he had 2 types of variables: those that range over individuals, and those that range over concepts. This, in itself, was not satisfactory. In the natural languages you do not have 2 types of variables, but just one, that can range over anything. I think that this can be easily seen with an example. Let us examine the sentence: “if Mary wants something, she gets it”. I will take that this sentence is equivalent to “for all X, if mary wants X, mary gets X”. So, we have constructs in the natural languages that behave very much like variables. And “to go to the cinema” can fall under X, so predicates can be in the range of variables. And also “an apple” can fall under the same X, which would correspond to an individual; so variables are first order, there is only one order of variables.

To bridge this gap, Frege had his Basic Law V, that basically established a correspondence between classes and concepts. (again, see here). This allowed him to unify the use of his 2 orders of variables. But along came Bertrand Russell, showing that this axiom leads to contradictions and must be dropped. This was a heavy blow for Frege, who felt that his project had failed, and never quite recovered from it.

His failure was however a very productive one, and predicate logic became a hugely successful technology. After Russell’s antinomies, predicate logic developed in 2 separate but interconnected ways. On one hand, logicists, following the lead of Zermelo, developed a first order set theory with which they provided a foundation for most mathematics. On the other hand, Russell, and the logical positivists, pursued the development of higher order logics with the original aim of Frege, of developing a formal system with the expressive power of the natural languages. They did not succeed in this effort.

First order logic

The problem encountered by Frege can also be seen in first order logic. To show this, we will imagine a first order set theory, with 3 predicates: “equals”, “belongs to”, and “is a subset of”. In principle, this theory will have 3 axioms, of equality, extensionality, and definition of subsets. This axioms provide the predicates with the basic form of class/set relations, and can thus be put in correspondence with the natural usage of the copular verbs. We then have a formal system where we can express any scientific taxonomy, and reason about it. Obviously this is not enough to express the logic of any scientific theory; we need other predicates apart from copular ones, to represent other relations apart from that of belonging to classes.

First order logic allows us to use more predicates, of course. We can define new predicates and give them whatever form we like through additional axioms. Nevertheless, there is still one problem to overcome if we aspire to a mechanization of the scientific use of the natural language. And that is the ability to have variables that range over predicates, the ability to express predicates through constraints, classes of predicates, etc. In first order predicate logic, you cannot have variables ranging over predicates. In the natural languages, I hope I have shown that you can do the equivalent.

The natural way out of this problem is an axiom schema of unrestricted comprehension, that (like Frege’s Basic Law V) establishes a correspondence between predicates and sets (thus classes). But if we add UC to the system, Russell’s paradox apply, that same paradox that toppled Frege’s edifice. As Paul Bernays stated in the introduction to his “Axiomatic set theory”:

What really is excluded by the antinomies is only that interpretation (easily suggested at first) of set theory (...) whose domain of individuals contains for every predicate B an assigned individual p such that:

(x)(x € p <-> B(x)).

—Paul Bernays

So with first order logic we come to exactly the same point where Frege foundered.

Modern manifestations of the problem

This problem I have described can be seen in many modern logic programming systems, where any attempt to use our natural logic as design model for software development is futile.

A paradigmatic example is the OWL language of the semantic web. This language had two flavours: DL, and full (well, and lite). It is an ontology language that can be processed by reasoners to extract consecuences. It has basic class predicates, and it has UC: you can have anonymous classes defined by predicates. But, in the DL flavour, you cannot treat classes as individuals: you cannot have variables ranging over them. In the full flavour, you can, but, as they say, “It is unlikely that any reasoning software will be able to support every feature of OWL Full” (see here). And it is the full flavour that would provide full (natural) expresivity.

A possible solution

My proposition is to use the predicates of set theory to express the natural copular verbs, just as Frege (or OWL) did, but then, as we shall see below, instead of representing the rest of the natural verbs as formal predicates, we represent them through individuals of the theory. We limit our (first order) theory to only have the basic predicates of set theory in their barest form. With their barest form, I mean defined by just equality, extensionality and a definition of subset, and perhaps some boundary axioms to define a universal and an empty “set”. In contrast, Paul Bernays, in the text quoted above, after dismissing UC, goes on to provide a number of additional axioms, like the axiom of choice or the axiom of infinity; what he called constructive axioms, that gave further form to the set predicates (and quite estranged them from the natural copulas). But his aim was different from ours. He wanted a formalism to base on it the whole of mathematics, so he needed a few axioms to produce a finished theory that would contain all mathematical structures. We do not need a closed theory, nor an initial complex structure as model for our theory. All we need is a simple and empty starting theory, that allows us to extend it with ad hoc new individuals and axioms to model each particular informal theory.

The theory NPL

We call this theory NPL. To sketch it, we will only use implication -> and conjunction & as logical connectives, and the only production rule will be modus ponens. Variables are denoted by x1, x2... and are always universally quantified in their outernmost scope (sentence); and individuals are denoted by any sequence of lower case letters. The predicates are isa, equivalent to “belongs to”, and are, equivalent to “is a subset of” (for this quick sketch of the theory, we do not need equality). We use these predicates in an infix form, and we have that:

x1 isa x2 & x2 are x3 -> x1 isa x3

x1 are x2 & x2 are x3 -> x1 are x3

Now to the representation of natural verbs other than copulas. For simplicity, we will only consider natural verbs that represent binary relations, so a natural sentence with such a verb would have the form of a triplet subject-verb-object. To represent this relation, we use a ternary operator f (from fact). So, a non-copular sentence, in our system, would have the form f(s, v, o) (where s, v, and o are just individuals of the theory). Since f is an operator, this sentence stands for just another individual of the theory, and has no truth value. We will call this sort of individuals “facts”. To attach truth value to facts, we use the set predicates, to put them in relation with another individual of the theory, fact. So a complete non-copular sentence, in this theory, would have the form (with prefix operators and infix predicates):

f(s, v, o) isa fact

Since we only have 2 (or 3, with equality) formal predicates, we do not need UC at all, and yet we can have variables that range over the equivalents of our natural verbs (and also over whole “facts”). The point is that we can model the forms of natural logic with very few predicate and operator symbols, and that any new term we may want to introduce, when modelling any kind of natural discourse, will be quantifiable by first order variables. Those symbols that can not be quantified, like are or isa or f, are so few that do not merit to be so.

We can be even more fine-grained. If we call “predication” to a pair verb-object, we may want to have variables that range over them. To do this, we can define a new operator p, that produces predication individuals, so that now the f operator takes 2 operands, the subject and a predication, to have something like:

p(v, o) isa predication

f(s, p(v, o)) isa fact

And, to show a little more of what can be obtained from such a system, note that facts and predications are individuals of the theory, so we can use them where we have used s or o, to build as complex a sentence as we may want (I think it wouldn’t make much sense to use them in place of v).

An example derived theory

An example developed on top of this theory might be (using a primitive universal set word):

person isa word

man are person

john isa man

woman are person

yoko isa woman

verb isa word

loves isa verb

x1 isa person &
x2 isa verb &
x3 isa person &
f(x1, x2, x3) isa fact
->
f(x3, x2, x1) isa fact

Now, john loves yoko will imply that yoko loves john.

There is a semantics for this theory here.


comments
blog comments powered by Disqus