How to handle the semantics of the TPTP THF0 language?
Christoph Benzmueller
In an ongoing project (funded by the EU IIF
grant THFTPTP 219982) a TPTP infrastructure for typed higher order logic is
being developed. As a starting point the THF0 language format has been
defined. THF0 covers Church's simple type theory which forms the core logic of
choice of many higher order automated theorem provers and proof assistants
including the systems TPS, LEO-II, Isabelle/HOL, PVS, HOL4, and OMEGA.
Currently a library of proof problems encoded in THF0 is built-up which will be
used in the future for the organization of higher order CASC events. There are
clearly many further applications of the new higher order TPTP
infrastructure. In particular, it is intended to provide a standard interface
for connecting higher order proof assistants with automated higher order theorem
provers.
The project distinguishes between three important aspects of THF0:
- Syntax
- Type checking
- Semantics
Syntax
The syntax of THF0 is defined as a conservative extension of the existing TPTP
language formats FOF and CNF for first order logic. An advantage of this approach is that
existing TPTP technology can be easily reused or adapted to also support the new
THFO language. A disadvantage, however, is that the BNF for THF0 appears overly
complicated and even messy. For this note that Church simple type
theory has a simple, single inductive structure at term/formula level
while first order logic has
a more complicated, double inductive structure which explicitly distinguishes
between the term level and the formula level. Church's simple type theory
instead only defines typed terms and all terms of a certain type (type Boolean)
are making up the formulas.
Type checking
The TPTP infrastructure does not provide type checking support for THF0. As a
consequence, the THF0 syntax definition allows many meaningless terms, for example, ~ = &. Additional type checking is therefore needed to distinguish between
well-typed, meaningful THF0 expressions and ill-typed, meaningless THF0
expressions. Such a type checking support for THF0 is provided via Twelf. Meaningful THF0
expressions are thus identified with the help of both the TPTP THF0 parser and
the Twelf type checker for THF0.
Syntax and type checking of THF0 is thus settled and more details can be found
in [1].
Semantics
An open issue in the development of the TPTP THF0 language is concerned with its
semantics. The problem is that there is no obvious, canonical notion of
semantics that we want to pick and stipulate for THF0. In order to illustrate the problems we briefly discuss some candidates:
- Standard semantics: Since there can be no complete theorem prover for
standard semantics this appears not a good option; there would be valid THF0
formulas which no theorem prover could ever prove.
- Henkin semantics: Henkin
semantics appears a strong candidate. It is a suitable generalization of
Standard semantics since it (theoretically) allows for complete
automation. Henkin semantics is fully extensional, that is, the functional
and Boolean extensionality principles are valid. An important question, however,
is whether we want to consider Henkin semantics with or without the axiom of
choice respectively Henkin semantics with or without description. Thus, even if
we would agree on Henkin semantics as an appropriate target semantics for THF0
there are still further options to consider.
- Weaker, non-extensional notions of semantics: Fully extensional Henkin
semantics qualifies as a suitable target semantics for mathematics
applications. For applications in computational linguistics and AI, for example,
it may be the wrong choice though. This is related to extensionality issues as
in these applications full extensionality is often not wanted. As a consequence,
when fixing Henkin semantics as the only and ultimate target semantics for THF0
then some problem statements would be considered as valid in the THF0 context which
would be considered as counter-satisfiable in these other contexts.
This would clearly limit the application range of the THF0 library.
More details on extensionality issues of Church's simple type theory and illustrating examples are given in [2].
- The above notions of semantics are classical and they assume
that the domains associated with types are non-empty for all types. Giving up these assumptions leads to further options:
- Possibly empty types: A Kripke semantics for equational reasoning in the simply typed lambda-calculus (Kripke Lambda Models) has been introduced by Mitchell and Moggi [4]; this semantics allows for empty types.
- Intuitionism: Lipton and Nieva extend the work by Mitchell and Moggi and provide a semantics for full impredicative higher-order intuitionistic logic [3].
The above discussion shows that there are good reasons for not fixing one
ultimate target semantics for THF0 in the first place. Instead, our proposal is
to annotate the THF0 problem files with semantical information. For this we want
to introduce an ontology of interesting notions of semantics for THF0 and let
the user annotate the problem files with the most suitable target semantics for
his/her application. Similarly, the provers shall provide information to which
target semantics they are applicable, that is, for which target semantics they
are performing sound proof search.
At the TPTP party this proposal shall be sketched and further discussed.
References
-
[1] C. Benzmüller, F. Rabe, and G. Sutcliffe, The Core
TPTP Language for Classical Higher-Order Logic. Fourth
International Joint Conference on Automated Reasoning (IJCAR'08),
LNAI, Sydney, Australia, 2008. © Springer.
-
[2] C. Benzmüller, C. Brown, and M. Kohlhase,
Higher Order Semantics and Extensionality. Journal of Symbolic Logic. (2004)
69(4):1027-1088. © JSTOR.
- [3] J. Lipton, S. Nieva, Higher-Order Logic Programming Languages with Constraints: A Semantics , 8th International Conference on Typed Lambda Calculi and Applications (TLCA'08), Paris, France, 2007. LNCS 4583, © Springer.
- [4] J. C. Mitchell, E. Moggi, Kripke-Style Models for Typed lambda Calculus, Ann. Pure Appl. Logic 51(1-2): 99-124 (1991).