Spatio-Temporal Graph Neural Networks for Infant Language Acquisition Prediction

Andrew Roxburgh, Floriana Grasso, Terry R. Payne · 2025

Read Spatio-Temporal Graph Neural Networks for Infant Language Acquisition Prediction by Andrew Roxburgh, Floriana Grasso, Terry R. Payne (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-2503-14341.txt.

Introduction

Andrew Roxburgh, Floriana Grasso, and Terry Payne predict which words a young child will learn next, in a paper published in two thousand twenty-five. Knowing the likely next words could help boost early language acquisition. Earlier predictors used either vocabulary change over time or relationships between words, but not both together.

Their model is a spatio-temporal graph convolutional network. Words form a graph whose edges capture different linguistic relationships. The network also tracks how each child's vocabulary state evolves. This combination aims to mirror the real complexity of infant learning. The authors evaluate how each relationship type affects prediction quality. Their observations shed light on which connections matter most during early word learning.

Catalog record

CREATORAndrew Roxburgh, Floriana Grasso, Terry R. Payne
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE38 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2503-14341

Browse the full library · About the Tower of Babel

Full text

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

Spatio-Temporal Graph Neural Networks for Infant Language Acquisition Prediction Andrew Roxburgh, Floriana Grasso, Terry R. Payne Abstract Predicting the words that a child is going to learn next can be useful for boosting language acquisition, and such predictions have been shown to be possible with both neural network techniques (looking at changes in the vocabulary state over time) and graph model (looking at data pertaining to the relationships between words). However, these models do not fully capture the complexity of the language learning process of an infant when used in isolation. In this paper, we examine how a model of language acquisition for infants and young children can be constructed and adapted for use in a Spatio-Temporal Graph Convolutional Network (STGCN) , taking into account the different types of linguistic relationships that occur during child language learning. We introduce a novel approach for predicting child vocabulary acquisition, and evaluate the efficacy of such a model with respect to the different types of linguistic relationships that occur during language acquisition, resulting in insightful observations on model calibration and norm selection. An evaulation of this model found that the mean accuracy of models for predicting new words when using sensorimotor relationships (0.733) and semantic relationships (0.729) were found to be superior to that observed with a 2-layer Feedforward neural network. Furthermore, the high recall for some relationships suggested that some relationships (e.g. visual) were superior in identifying a larger proportion of relevant words that a child should subsequently learn than others (such as auditory). 1 Introduction Developmental Language Disorder (DLD) is a condition whereby a child shows significant difficulties with language development for no clear reason. It is the most common disability in pre-schoolers, affecting somewhere between 5-7% of all UK children [ Lindsay et al.(2002)Lindsay, Dockrell, Mackie, et al. , Scerri et al.(2011)Scerri, Morris, Buckingham, Newbury, Miller, Monaco, Bishop, and Paracchini , Tomblin et al.(1997)Tomblin, Records, Buckwalter, Zhang, Smith, and O’Brien ] . The condition is exacerbated by other factors associated with learning difficulties, or environmental or familial factors. For example, in the UK, between 40% and 56% of children in areas of social disadvantage begin nursery school with such a language delay [ Law et al.(2011)Law, McBean, and Rush , Locke et al.(2002)Locke, Ginsborg, and Peers ] . This has been proven to cause a knock-on effect within child education; for example, DLD has an incidence rate of 50-90% on reading difficulties [ Stothard et al.(1998)Stothard, Snowling, Bishop, et al. ] , which can have a subsequent impact on educational outcomes (in fact reading levels have been shown to be a good predictor of educational outcomes [ Hulme and Snowling(2013) ] ). Furthermore, children show great difficulties catching up with their peers without adequate support [ Bleses et al.(2016)Bleses, Makransky, Dale, et al. , Conti-Ramsden et al.(2001)Conti-Ramsden, Botting, Simkin, et al. , Feinstein and Duckworth(2006) , Van Dulm and Southwood(2016) ] , and thus are more likely to perform worse academically, to show a lower employment rate, and to have poorer future mental health [ Clegg et al.(2005)Clegg, Hollis, Mawhood, et al. , Snowling et al.(2001)Snowling, Adams, Bishop, et al. , Young et al.(2002)Young, Beitchman, Johnson, et al. ] . Language delay is also an indicator of many other developmental issues; at least 3% of all UK children have a language delay linked with hearing impairment, specific learning difficulties such as dyslexia, and general learning needs [ Lindsay et al.(2002)Lindsay, Dockrell, Mackie, et al. ] . Communication difficulties also feature in Autism Spectrum Disorder, which affects 1% of the UK population [ Baron-Cohen et al.(2009)Baron-Cohen, Scott, Allison, et al. ] . Such challenges are not limited to neurologically ‘atypical’ children, as the child’s communication environment and family circumstances (e.g. poverty or limited language exposure) can have an effect on any early years language development, with a negative, cascading effect on school outcomes years later [ Roulstone et al.(2011)Roulstone, Law, Rush, et al. , Walker et al.(1994)Walker, Greenwood, Hart, et al. ] . With such a public health issue of this scale, identifying affected children early is extremely important. One of the most commonly-used methods for identifying such children is through the use of standardised tools such as the MacArthur-Bates Communicative Development Inventory (CDI) [ Fenson et al.(2007)Fenson, Marchman, Thal, Dale, Reznick, and Bates ] , a paper instrument originally developed in US English but now adapted into many languages [ Alcock et al.(2016)Alcock, Rowland, Meints, Christopher, Just, Brelsford, and Summers ] . It consists of a series of questions and checklists designed to assess how children understand and produce words and gestures, as well as their acquisition of grammar. It is typically accompanied by a Family Questionnaire, which gathers information about the family members and individuals who generally spend time with the child during a typical week. This data can be used to generate a spatial model of language acquisition (where the spatial component corresponds to words with specific linguistic relationships), which can then be used to identify children who may suffer from language delay. Figure 1: General operation of a Spatial Graph Neural Network (GNN). By automating the process of acquiring this data, through mobile applications such as Babytalk 1 1 1 http://www.lucid.ac.uk/ [ Roxburgh et al.(2017)Roxburgh, Grasso, and Payne ] , a model of a child’s language acquisition can be built over time, and used to predict those words that the child should learn in the future [ Beckage et al.(2015)Beckage, Mozer, and Colunga , Beckage et al.(2020)Beckage, Mozer, and Colunga ] . This knowledge can be used as a potential intervention for children with DLD, by suggesting words that a child’s carer should emphasise when teaching language [ Johnson and Jusczyk(2001) ] ; for example, using the Babytalk application. This could also be used to boost the language acquisition of neurologically atypical children, and could have further research applications. One way to accomplish this is to refer to Age-of-Acquisition norms [ Stadthagen-Gonzalez and Davis(2006) ] that can be used to encourage caregivers to emphasise words that typical children would acquire at a similar age. However, children (as individuals) do not tend to learn the same words at the same rate or age, and so a more tailored approach is required. The accuracy of the model used to predict forthcoming words is therefore an important factor. Graph Neural Networks (GNN) [ Wu et al.(2021)Wu, Pan, Chen, Long, Zhang, and Yu ] extend the concept of conventional feedforward neural networks to capture the relationships and interactions between nodes within a graph, by learning a vector representation of each node that not only depends on its own features but also on the features of its neighbours. These node embeddings encode the structural and feature data of themselves, their neighbouring nodes, and ultimately extend to the overall graph structure. The resulting model can be used for edge prediction, node or graph classification, labelling, and feature prediction; or in the context of language acquisition, it could be used to determine whether the word that the node represents will be ‘known’ by the child. In this paper we examine how a model of language acquisition for infants and young children can be constructed and adapted for use in a Spatio-Temporal Graph Convolutional Network

Read the complete text