Which level of abstraction, and when? Resource-rational granularity in inflection class learning

Abstract Morphological knowledge can be characterised at multiple levels of abstraction: the same facts about the lexicon can be stated as a single schema per inflectional macro-class, as intermediate sub-schemas over lexeme sets that share a given property, or as stored individual forms. This is usually cast as a choice of cognitive architecture: pure prototype, pure exemplar, or intermediate/multi-level systems. We reframe the question of the correct level of abstraction to model morphological knowledge as not one of architecture, but one where optimal level of abstraction is itself the variable of interest, to be derived from properties of the data. Under a resource-rational view (Anderson, 1991), humans build as much representational structure as the evidence justifies and no more, so the optimal granularity is inferred from the input, shifts as evidence accumulates, and may differ across regions of the lexicon. We treat the question of how we partition and represent lexical knowledge as a categorisation problem, and model it as Bayesian nonparametric inference in which granularity is a free parameter. We model trial-by-trial performance in an artificial-language experiment on plural allomorphy with exceptions with models that make different assumptions about granularity. On the basis of the results, we argue that granularity is thus productively thought of not a property of the morphological architecture to be discovered but a quantity inferred from evidence. The method speaks directly to theories of morphology which have an architecture where all levels of abstraction coexist, enabling them to make predictions about which grain a given speaker brings to bear on a particular linguistic task for a particular word.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-13
DOI
https://doi.org/10.21203/rs.3.rs-11362813/v1
Primary Topic
Categorization, perception, and language
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Which level of abstraction, and when? Resource-rational granularity in inflection class learning

Maria Copot, Jayden Macklin-Cordes
2 links
Research Square
Categorization, perception, and language
preprint

Which level of abstraction, and when? Resource-rational granularity in inflection class learning

Maria Copot, Jayden Macklin-Cordes
preprint en

Abstract

Abstract Morphological knowledge can be characterised at multiple levels of abstraction: the same facts about the lexicon can be stated as a single schema per inflectional macro-class, as intermediate sub-schemas over lexeme sets that share a given property, or as stored individual forms. This is usually cast as a choice of cognitive architecture: pure prototype, pure exemplar, or intermediate/multi-level systems. We reframe the question of the correct level of abstraction to model morphological knowledge as not one of architecture, but one where optimal level of abstraction is itself the variable of interest, to be derived from properties of the data. Under a resource-rational view (Anderson, 1991), humans build as much representational structure as the evidence justifies and no more, so the optimal granularity is inferred from the input, shifts as evidence accumulates, and may differ across regions of the lexicon. We treat the question of how we partition and represent lexical knowledge as a categorisation problem, and model it as Bayesian nonparametric inference in which granularity is a free parameter. We model trial-by-trial performance in an artificial-language experiment on plural allomorphy with exceptions with models that make different assumptions about granularity. On the basis of the results, we argue that granularity is thus productively thought of not a property of the morphological architecture to be discovered but a quantity inferred from evidence. The method speaks directly to theories of morphology which have an architecture where all levels of abstraction coexist, enabling them to make predictions about which grain a given speaker brings to bear on a particular linguistic task for a particular word.

Research Square
University of Zurich (CH), University of Surrey (GB)
Categorization, perception, and language
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon