Title: Light or Full Verb? A Minimal-Pair Dataset for Probing Phraseological Competence in Language Models

Abstract:

Common English verbs like 'have' and 'make' serve dual roles: they can act as collocates within light-verb constructions or function as full lexical predicates, exemplified by the contrast between "make a decision" and "make a cake." It remains uncertain whether current language models adequately capture this linguistic distinction. To address this, we present a large-scale, controlled dataset comprising minimally varied English sentence pairs where identical contexts feature the same verb used both as a light verb and as a full verb. Our probing experiments demonstrate that language models can distinguish between these two uses even within minimal contexts, revealing distinct behavioral patterns based on object types. We make the dataset, generation code, and associated materials available as a reusable resource. This framework is designed to be extensible, allowing for integration into broader contexts, the inclusion of additional verbs, and application across other languages.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC