Large corpora and large language models: a replicable method for automating grammatical annotation

Linguistics Vanguard, Journal Year: 2025, Volume and Issue: unknown

Published: April 9, 2025

Abstract Much linguistic research relies on annotated datasets of features extracted from text corpora, but the rapid quantitative growth these corpora has created practical difficulties for linguists to manually clean and annotate large data samples. In this paper, we present a method that leverages language models assisting linguist in grammatical annotation through prompt engineering, training, evaluation. We apply methodological pipeline case study formal variation English evaluative verb construction “ consider X (as) (to be) Y”, based model Claude 3.5 Sonnet Davies’s NOW Sketch Engine’s EnTenTen21 corpora. Overall, reach accuracy over 90 % our held-out test samples with only small amount training data, validating very quantities tokens future. discuss generalizability results wider range studies constructions change, underlining value AI copilots as tools future research, notwithstanding some important caveats.

Language: Английский

---

Constructing understanding: on the constructional information encoded in large language models

Language Resources and Evaluation, Journal Year: 2024, Volume and Issue: unknown

Published: Dec. 20, 2024

Abstract We review research related to both Construction Grammar (CxG) and Natural Language Processing showing that recent advances in probing Large Models (LLMs) for certain types of linguistic knowledge align with the tenets CxG. However, our survey leads us hypothesize LLM constructional information may be limited constructions within lower levels postulated taxonomical “constructicons” enumerating a particular language’s constructions. Specifically, studies show at taxonomy, which are more substantive fixed elements corresponding frequently used words construction, type accessible LLMs. In contrast, general, abstract schematic slots can filled by variety different not included test this hypothesis on collection 10 distinct constructions, each is exhibited 50 or corpus instances. Our experimental results strongly support lead conclude that, order LLMs generalize point where purely recognized regardless frequency instantiating (as psycholinguistic experimentation has shown people can), additional semantic resources needed make explicit role slot. To ensure transparency reproducibility, we publicly release data, including prompts model.

Language: Английский

---

The development of the theory of construction grammar in 1985–2024

OOO Zhurnal Voprosy Istorii, Journal Year: 2024, Volume and Issue: unknown, P. 152 - 159

Published: Nov. 4, 2024

This article provides a detailed overview of the history development theory construction grammar, which began to take shape in late 1980s and reached its peak 21st century. The key stages also are analyzed. review aims ensure further laying groundwork for scientific progress.

Language: Английский

---