Dancing with Deer: A Constructional Perspective on MWEs in the Era of LLMs

Claire Bonial, Julia Bonn, Harish Tayyar Madabushi · 2025

Read Dancing with Deer: A Constructional Perspective on MWEs in the Era of LLMs by Claire Bonial, Julia Bonn, Harish Tayyar Madabushi (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-15977.txt.

Introduction

Claire Bonial, Julia Bonn, and Harish Tayyar Madabushi argued for a construction grammar view of multiword expressions in two thousand twenty-five. Such expressions, like idioms, pair a form with a meaning that is not built word by word. Construction grammar was designed to handle them with the same machinery as ordinary language.

The chapter traces that history and shows two case studies: one for English expressions in PropBank, and one for complex word units in Arapaho, a highly synthetic language. It then compares how a person and a large language model learn a novel expression from one example. Both generalized from a single use, but only people could reason across two such expressions.

Catalog record

CREATORClaire Bonial, Julia Bonn, Harish Tayyar Madabushi
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE97 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2508-15977

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Dancing with Deer: A Constructional Perspective on MWEs in the Era of LLMs Claire Bonial, Julia Bonn, Harish Tayyar Madabushi Abstract In this chapter, we argue for the benefits of understanding multiword expressions from the perspective of usage-based, construction grammar approaches. We begin with a historical overview of how construction grammar was developed in order to account for idiomatic expressions using the same grammatical machinery as the non-idiomatic structures of language. We cover a comprehensive description of constructions, which are pairings of meaning with form of any size (morpheme, word, phrase), as well as how constructional approaches treat the acquisition and generalization of constructions. We describe a successful case study leveraging constructional templates for representing multiword expressions in English PropBank. Because constructions can be at any level or unit of form, we then illustrate the benefit of a constructional representation of multi-meaningful morphosyntactic unit constructions in Arapaho, a highly polysynthetic and agglutinating language. We include a second case study leveraging constructional templates for representing these multi-morphemic expressions in Uniform Meaning Representation. Finally, we demonstrate the similarities and differences between a usage-based explanation of a speaker learning a novel multiword expression, such as “dancing with deer,” and that of a large language model. We present experiments showing that both models and speakers can generalize the meaning of novel multiword expressions based on a single exposure of usage. However, only speakers can reason over the combination of two such expressions, as this requires comparison of the novel forms to a speaker’s lifetime of stored constructional exemplars, which are rich with cross-modal details. \noautomath 1 Introduction At its inception, Construction Grammar (CxG) approaches have stood apart from other theoretical approaches to language because of their explanatory power when it comes to multiword expressions (MWEs). CxG is able to provide the machinery for elegantly modeling MWEs for two primary reasons. First: in CxG, the basic unit of language is the construction (Cxn), which is a pairing of meaning with a form, where the forms can vary across every level of language – thus there are morphological forms, lexical forms, and phrasal forms. As a result, CxG provides a unified theory for accounting for meaning in multi-morphemic structures in precisely the same fashion as MWEs or transitive phrasal structures. Second: as a usage-based theory, CxG research demonstrates how frequency plays a role in facilitating extension but also preventing overgeneralization of flexible, partially productive Cxns. Thus, CxG can account for the seemingly idiosyncratic flexibility and productivity of MWEs, which has historically presented a challenge to the simplest Natural Language Processing (NLP) techniques of treating MWEs like “words with spaces” ( Sag et al. 2002 ) . To elaborate upon these strengths in a CxG account of MWEs, this chapter begins by detailing the tenets of CxG approaches, with special focus on the notion of a Cxn and the role of frequency in the storage and generalization of Cxns (Section 2 ). We then turn to a description of how constructional templates have been used in modeling English MWEs in the PropBank project ( Palmer et al. 2005 ) , enabling theoretical advances in our understanding of the characteristics of MWEs, as well as computational advances in their detection and semantic interpretation (Section 3 ). In this way, we show the benefit of having one unified treatment of both lexical and phrasal structures. Next, we turn to a description of multi-morphemic expressions in Arapaho, including efforts to develop Uniform Meaning Representation ( Bonn et al. 2023 ) for Arapaho. This requires eschewing traditional notions of word-hood in favor of Cxns, whose forms do not require any clear delineation between morphemes and words (as there are no morphological, input/output rules in CxG) (Section 4 ). Thus, we show the benefit of one theoretical approach to account for meaning in multi-morphemic and lexical structures. Finally, we describe the parallels between a speaker’s usage-based acquisition of grammar and the way that Large Language Models (LLMs) process text in order to acquire an impressive amount of linguistic competency (Section 5 ). We also present two exploratory experiments on both novel and nascent English MWEs that demonstrate that although frequency effects in a speaker’s grammar mimic priors in an LLM’s pretraining, this does not result in fully functional linguistic competence (Section 5.2.1 , 5.2.2 ). This, too, is expected under CxG: whereas an LLM’s “usage” is all through the processing of text, a person’s usage is enriched by cross-modal association, resulting in vastly different meaning features paired up with a given form. We close with recommendations for future work in developing constructional resources that will help to bridge the gap between human linguistic competency and that of our NLP systems (Section 6 ). 2 Accounting for MWEs in Construction Grammar In this section, we begin with the importance of idiomatic expressions, one subset of MWEs, in the historical development of CxG. We then turn to a detailed description of what Cxns are, as well as what features of Cxns are posited to be stored in the mental lexicon, or ‘constructicon’. We then turn to a discussion of the importance of frequency of use and the factors that both facilitate and constrain the extension or generalization of Cxn types to novel instantiations. Such extension leads to Cxns of varying level of abstractness, or ‘schematicity,’ discussed next. We close this section with a description of constructional templates, facilitating the representation of both fixed and flexible constructional slots. 2.1 CxG: a Theory for “Peripheral” Language Phenomena Croft and Cruse write, “It is not an exaggeration to say that Cxn grammar grew out of a concern to find a place for idiomatic expressions in the speaker’s knowledge of a language” ( Croft & Cruse 2004 , 205) . A place for idiomatic expressions had not been satisfactorily found in the Generative linguistic tradition, e.g. Chomsky 1957 , in which idiomatic expressions had been relegated to the “peripheral” language phenomena, which seemingly neither cooperated with nor merited an elegant rule-based explanation of properties. We define idioms as a class of MWEs that exhibit characteristics of lexical idiosyncrasy and syntactic creativity. Idioms are lexically idiosyncratic in that they are tied to certain words with different degrees of fixedness. They exhibit syntactic creativity in that their behavior may not be accounted for by existing syntactic rules. These characteristics are especially problematic when viewed from a Generative perspective, which posits the autonomy of syntax and semantics – semantic features are stored in the lexicon, while syntax dictates the combinatoric potential of lexical items. The Generative solution to the problem of idiomatic expressions was largely to posit that their non-conventional meaning requires speakers to store them as frozen, inflexible chunks in their mental lexicon. However, the majority of idioms are not frozen, showing some syntactic flexibility, such as the ability to be marked for the past tense or passivized: “The beans that Nathan spilled…” Others are more fixed, resisting passivization: *‘‘The bucket that Grandpa kicked...’’ 1 1 1 Usages thought to be less felicitous to native speakers are preceded by ‘*’. If idioms are to be stored in the lexicon, this would reasonably lead one to conclude that some syntactic information is also stored in the lexicon, with respect to what

Read the complete text