The word as a unit of internal predictability

Abstract

A long-standing problem in linguistics is how to define ‘word’. Recent research has focused on the incompatibility of diverse definitions, and the challenge of finding a definition that is cross-linguistically applicable (e.g. Haspelmath 2011; Gijn and Zúñiga 2014; Bickel and Zúñiga 2017; Tallman 2020). In this study I take a different approach, asking whether one structure is more word-like than another based on Shannon’s (1948) concepts of predictability and information. I hypothesise that word constructions tend to be more ‘internally predictable’ than phrase constructions, where internal predictability is the degree to which the entropy of one constructional element is reduced by mutual information with another element. I illustrate the method with case studies of complex verbs in German and Murrinhpatha, comparing verbs with selectionally restricted elements against those built from free elements. I propose that this method identifies an important mathematical property of many word-like structures, though I do not expect that it will solve all the problems of wordhood.

Cite this work

No public copy is stored for this record, so only the metadata is cited.

Stored only when both papers are in the corpus — 24 citations exist upstream.

Papers that cite the same references — the relational fallback while embeddings are still backfilling.