Moving to continuous classifications of bilingualism through machine learning trained on language production

Q: What problem does the study address in bilingualism research?

It addresses how strict categorical labels for bilingualism may overlap and fail to support consistent comparisons across studies, since individuals can change language-related profiles over time.

Q: How were machine learning models used in this research?

Support vector classifiers were trained on two datasets of Italian speakers’ linguistic productions to predict classes: monolingual, attriters, and heritage.

Q: Which linguistic features and class relationships were most informative?

Clitic clusters were identified as the most discriminative features. The confusion matrix showed attriters were often classified as heritage speakers, indicating they may lie between the two categories.

Aug 5, 2026

Upload by Aurora

Recent conceptualisations of bilingualism move away from strict categorisations toward continuous approaches. This study combines empirical psycholinguistics data with machine learning classification modelling. Support vector classifiers were trained on two datasets of linguistic productions by Italian speakers to predict classes (monolingual, attriters, heritage). All classes were identified above chance, with monolinguals best and attriters most confusable. Confusion patterns indicate attriters often align with heritage speakers, suggesting a middle position. Clitic clusters proved especially informative, supporting bilingualism as a continuum of linguistic behaviours.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages34

Reading Time1 h 26 min

AccessFree to download

Abstract

Introduction

Background and limitations of categorical bilingualism

Research question and implications for terminology and methodology

1 Short title: Moving to continuous classifications of bilingualism through machine 2 learning 3 4 Long title: Moving to continuous classifications of bilingualism through machine 5 learning trained on language production 6 7 Authors: Coco, M. I. 1,2*, Smith, G.3*, Spelorzi, R.4 & Garraffa, M.3 8 1Department of Psychology,“Sapienza” University of Rome, Rome, Italy 9 2 I.R.C. S. S Fondazione Santa Lucia, Rome, Italy 10 3 School of Health Sciences, University of East Anglia, Norwich, UK 11 4Department of Linguistics and English Language, University of Edinburgh, Edinburgh, UK 12 13 * Denotes equal contribution 14 15 Competing interests: the authors declare none. 16 17 Addresses for correspondence: 18 19 Dr Moreno I. Coco 20 Dipartimento di Psicologia 21 Sapienza, Universita’ di Roma 22 Via dei Marsi, 78 23 00185, Roma, 24 Italy 25 [email: ](email: moreno.coco@uniroma1.it)[moreno.coco@uniroma1.it](email: moreno.coco@uniroma1.it)[ ](email: moreno.coco@uniroma1.it)26 27 Dr Giuditta Smith 28 School of Health Sciences 29 University of East Anglia 30 Norwich Research Park 31 NR4 7TJ 32 Norwich 33 UK 34 Email: [giuditta.smith@uea.ac.uk](giuditta.smith@uea.ac.uk) 35 36 37 Abstract 38 39 Recent conceptualisations of bilingualism are moving away from strict categorisations, 40 towards continuous approaches. This study supports this trend by combining empirical 41 psycholinguistics data with machine learning classification modelling. We trained support 42 vector classifiers on two datasets of linguistic productions coded for type of production of 43 Italian speakers to predict their class (i.e., “monolingual”, “attriters”, and “heritage”) . All 44 classes were predicted above chance (> 33%), even if the classifier’s performance substantially 45 varies, with monolinguals identified better (f-score > 70%) and attriters being the most 46 confusable (f-score < 50%) . The confusion matrices qualify that attriters are identified as 47 heritage speakers nearly as often as they could be correctly classified, suggesting this class to 48 sit in the middle. Clitic clusters were found to be the most identifying features for 49 discrimination. Overall, this study supports a conceptualisation of bilingualism as a continuum 50 of linguistic behaviours rather than sets of a-priori established classes. 51 Keywords: bilingualism, heritage speakers, attrition, support vector machine, classification 52 53 1. Introduction 54 55 In a globalized and highly integrated world, the boundaries of languages have become fluid 56 and seemingly continuous. Speakers are more likely to move across countries, transfer their 57 homeland language to their offspring and acquire other languages, with bilingual proficiency 58 reaching native-like language abilities well after childhood (Steinhauer, 2014; Roncaglia- 59 Denissen & Kotz, 2016; Hartshorne, Tenenbaum, & Pinker, 2018; Köpke, 2021; Gallo et al. 60 2021). However, bilingualism is known to substantially vary among individuals, as it is 61 shaped by intra and extralinguistic factors such as amount of exposure, social status, and 62 education (Gullifer et al. 2018; Haranto & Yang, 2016; Rodina et al. 2020; Bialystok, 2016; 63 Gullifer & Titone, 2020) . Consequently, research in bilingualism has progressively 64 abandoned strict categorical approaches in favour of more nuanced ones. The increased 65 complexity of a “winner-take-all” definition of bilingualism has created a plethora of labels 66 to classify speakers (see Surrain & Luk, 2017 for a systematic review), sometimes leading to 67 the same speakers being labelled differently according to whether the classification is based 68 on language dominance, learning history, age, etc., which renders it impractical to perform 69 consistent comparisons across different studies. Moreover, strict classifications disregard that 70 individuals can also change

FAQs

What problem does the study address in bilingualism research?

It addresses how strict categorical labels for bilingualism may overlap and fail to support consistent comparisons across studies, since individuals can change language-related profiles over time.

How were machine learning models used in this research?

Support vector classifiers were trained on two datasets of Italian speakers’ linguistic productions to predict classes: monolingual, attriters, and heritage.

Which linguistic features and class relationships were most informative?

Clitic clusters were identified as the most discriminative features. The confusion matrix showed attriters were often classified as heritage speakers, indicating they may lie between the two categories.

Aug 5, 2026

Upload by Aurora