Re-thinking Supertags in Linear Context-free Rewriting Systems for Constituency Parsing

Abstract

Recently, a supertagging-based approach for parsing discontinuous constituent trees with linear context-free rewriting systems (LCFRS) was introduced. We reformulate their algorithm for the extraction of supertags from treebanks to be more concise. Moreover, we add some extensions that give us control over the extraction process in terms of supertag granularity and which terminal symbols are associated with supertags. Our additions lead to an increase in parsing quality with LCFRS supertagging in all three compared treebanks. The scores are among the state of the art in discontinuous constituent parsing.

Tasks

Reproductions