A semantic
``Trampe'' for HPSG: Easy ways from bottom to top and back
in Trondheim
Frank Richter and Manfred Sailer
Invited talk given at
HPSG 01
in Trondheim,
Norway, on August 3rd, 2001.
Abstract
HPSG offers linguists a choice between a fairly wide range of semantic
representation languages. Some of these representation languages
endorse ideas of underspecified semantic systems and are usually
influenced by computational considerations, while others are inspired
by fully specified systems.
In this talk we want to clarify some important theoretical issues
pertaining to the architecture and potential range of semantic systems
in HPSG. For our investigation, we start with an easily accessible,
type-theoretical language, Ty2, which is similar to Richard Montague's
Intensional Logic. We use techniques offered by the HPSG formalism
outlined in Pollard & Sag 94 in order to build different combinatorial
semantic systems for that language in HPSG grammars and to explore
various ways of underspecification. The techniques that we apply will
allow us to distinguish at least four possible semantic systems
available in HPSG: systems with a truly underspecified denotation,
systems that use an indirect representation, systems with
discontinuous representations, and systems that only employ classical,
``fully specified'' representations with a traditional combinatorial
system such as lambda abstraction and functional application. With
respect to this small taxonomy of semantic systems, we can classify
the current semantic representation languages of the HPSG literature.
Having characterized a range of interesting options that are
technically available, we then turn to questions of the empirical
adequacy of the resulting alternative architectures of
grammar. Drawing on well-known empirical phenomena in various
languages, we will show that different semantic systems force the
linguist to express certain empirical generalizations in different
modules of the grammar. Depending on the module in which the
generalizations are expressed, different natural classes of data are
predicted. To the extent to which these predictions can be tested, the
choice between the semantic systems becomes accessible to empirical
tests. Besides arguments of computational feasibility and largely
aesthetical differences in elegance, we thus obtain an additional
dimension for choosing a suitable semantic system for HPSG grammars of
natural languages.
There is one more important issue concerning semantic systems that we
will pay attention to throughout the talk. HPSGians adopt two kinds of
underlying mathematical foundations. While some follow the
non-unification, constraint-based framework introduced by Pollard &
Sag 94, others adhere to the unification-based formalism of Pollard
& Sag 87. Concrete formalisms that currently represent these two
directions are RSRL for Pollard & Sag 94 and the unification formalism
underlying the LKB system for Pollard & Sag 87, which is inspired by
the unification formalism developed and implemented by Carpenter and
Penn. When we formalize the four types of semantic systems
distinguished above, we do so in RSRL. We will thus have to keep an
eye on where particular properties of the RSRL formalism that have no
direct counterpart in a unification-based formalism might come into
play at each step and tie the choice of the semantic framework to a
particular formalism for HPSG. Moreover, the different ontological
status of the structures in the denotation of a grammar and the
different function of the mathematically corresponding structures in
an information-based approach leads to plausible alternatives for the
interpretation of the structures that are used as semantic
representations.
Then handout of the talk is available in two formats:
Frank Richter