Evaluating a bracketing protocol for multiword terms

Abstract

Multiword terms (MWTs) are frequently used to encapsulate and convey meaning in scientific and technical texts. However, they can also make these texts difficult to understand because the relations between constituents are not transparent. When MWTs have more than two constituents, a dependency analysis (bracketing) is often necessary to facilitate their interpretation. NLP has proposed various models to automatize bracketing operations, but none has been entirely satisfactory. This paper presents a protocol that combines various models and applies it to a set of three-constituent MWTs in order to: (i) sort rules by their disambiguation potential, based on their likelihood of retrieving results from any corpus and their ability to solve bracketing; and (ii) ascertain the influence of corpus size and type in the results obtained.