Graph- and surface-level sentence chunking

Abstract

The computing cost of many NLP tasks increases faster than linearly with the length of the representation of a sentence. For parsing the representation is tokens, while for operations on syntax and semantics it will be more complex. In this paper we propose a new task of $\textit{sentence chunking}$: splitting sentence representations into coherent substructures. Its aim is to make further processing of long sentences more tractable. We investigate this idea experimentally using the Dependency Minimal Recursion Semantics (DMRS) representation.