Combining logical and distributional methods in type-logical grammars

Abstract

We propose a low-level way of combining distributional and logical ideas into a single formal system. This will be an instantiation of a more general system, adding weights to proof rules. These weights will not measure some sort of "confidence the proof is valid", but rather act as a way to prefer some proofs over others, where preference can mean
"easier to process (for humans)" or "more coherent (combining words that make sense together)". The resulting system of weighted theorem proving can be implemented either as a best-first proof search strategy or as a polynomial-time approximation of proof search for NP-complete parsing problems.