Developing a semantic representation for predictive TAG parsing

This talk is intended to present our on-going efforts in constructing a
semantic formalism that preserves the incremental properties of PLTAG
while remaining both flexible and formally robust. PLTAG is a
Tree-Adjoining Grammar variant which supports strict parser
incrementality and near-future syntactic prediction by using modified
and augmented LTAG operations and a lexical induction process that
generates underspecified "prediction trees" alongside LTAG-style
elementary trees.  Of particular note is its unification-based
prediction/verification mechanism that allows for the explicit
incorporation of expectation into the syntactic structure.

Existing approaches to robust semantic parsing often adopt
neo-Davidsonian/minimally recursive approaches due to their flexibility
in representing variable argument structures, adjuncts, and so on.  We
adopt a neo-Davidsonian approach as well, but we have the added burden
of connecting a highly restricted, psycholinguistically-motivated
syntactic parsing algorithm to the process that constructs semantic
representations.

We describe our approach to augmenting the PLTAG lexicon with semantic
information, including how we represent prediction.  We then give a
high-level overview of how a PLTAG parse can result in a neo-Davidsonian
representation.  Finally, we bring up three issues for discussion: (1)
the binding of unbound variables during adjunction, (2)
linguistically-plausible limits on quantifier scope and
underspecification, and (3) potential conflicts between
minimally-recursive approaches and leading-edge work in distributional
semantics.