The research conducted for this thesis has been guided by the vision of a computer program that could check the correctness of mathematical proofs written in the language found in mathematical textbooks. Given that reliable processing of unrestricted natural language input is out of the reach of current technology, we focused on the attainable goal of using a controlled natural language (a subset of a natural language defined through a formal grammar) as input language to such a program. We have developed a prototype of such a computer program, the Naproche system. This thesis is centered around the novel logical and linguistic theory needed for defining and motivating the controlled natural language and the proof checking algorithm of the Naproche system. This theory provides means for bridging the wide gap between natural and formal mathematical proofs.

We explain how our system makes use of and extends existing linguistic formalisms in order to analyse the peculiarities of the language of mathematics. In this regard, we describe a phenomenon of this language previously not described by other logicians or linguists, the implicit dynamic function introduction, exemplified by constructs of the form "for every x there is an f(x) such that ...". We show how this function introduction can lead to a paradox analogous to Russell's paradox. To tackle this problem, we developed a novel foundational theory of functions called Ackermann-like Function Theory, which is equiconsistent to ZFC (Zermelo-Fraenkel set theory with the Axiom of Choice) and can be used for imposing limitations to implicit dynamic function introduction in order to avoid this paradox.

We give a formal account of implicit dynamic function introduction by extending Dynamic Predicate Logic, a formalism developed by linguists to account for the dynamic nature of natural language quantification, to a novel formalism called Higher-Order Dynamic Predicate Logic, whose semantics is based on Ackermann-like Function Theory. Higher-Order Dynamic Predicate Logic also includes a formal account of the linguistic theory of presuppositions, which we use for clarifying and formally modelling the usage of potentially undefined terms (e.g. 1/x, which is undefined for x=0) and of definite descriptions (e.g. "the even prime number") in the language of mathematics. The semantics of the controlled natural language is defined through a translation from the controlled natural language into an extension of Higher-Order Dynamic Predicate Logic called Proof Text Logic. Proof Text Logic extends Higher-Order Dynamic Predicate Logic in two respects, which make it suitable for representing the content of mathematical texts: It contains features for representing complete texts rather than single assertions, and instead of being based on Ackermann-like Function Theory, it is based on a richer foundational theory called Class-Map-Tuple-Number Theory, which does not only have maps/functions, but also classes/sets, tuples, numbers and Booleans as primitives.

The proof checking algorithm checks the deductive correctness of proof texts written in the controlled natural language of the Naproche system. Since the semantics of the controlled natural language is defined through a translation into the Proof Text Logic formalism, the proof checking algorithm is defined on Proof Text Logic input. The algorithm makes use of automated theorem provers for checking the correctness of single proof steps. In this way, the proof steps in the input text do not need to be as fine-grained as in formal proof calculi, but may contain several reasoning steps at once, just as is usual in natural mathematical texts. The proof checking algorithm has to recognize implicit dynamic function introductions in the input text and has to take care of presuppositions of mathematical statements according to the principles of the formal account of presuppositions mentioned above. We prove two soundness and two completeness theorems for the proof checking algorithm: In each case one theorem compares the algorithm to the semantics of Proof Text Logic and one theorem compares it to the semantics of standard first-order predicate logic.

As a case study for the theory developed in the thesis, we illustrate the working of the Naproche system on a controlled natural language adaptation of the beginning of Edmund Landau's Grundlagen der Analysis.

Table of contents

Complete version (2,5 MB)