Work

Computational Modelling of Early Language Acquisition : Towards naturalistic simulations and robust model assessment

Year: 2025
Type: dissertation
Abstract: Early language acquisition is a complex developmental process in which infants rapidly develop linguistic capabilities. During the first year of life, infants demonstrate superficially effortless early proficiency in this learning process despite cognitive and sensory constraints related to the developmental stage in which this process occurs. Studying infant language acquisition is inherently challenging, as infants cannot explicitly communicate their learning progress. This requires researchers to rely on indirect behavioural and neural measurements to assess their linguistic skills. Complementary computational modelling offers an approach to studying the process, enabling the testing of hypotheses about language learning mechanisms, required language input, and other developmental factors that are difficult to investigate experimentally. However, the effectiveness of computational models is affected by how ecologically valid the simulation setup is, including the assumptions and data used to model the learning environment, the decisions made about learning mechanisms (including, for example, naturalness of the processing and the model representations), and the robustness of evaluation methods used to compare the model’s behaviour against infant data. This thesis addresses two key challenges in computational modelling of infant language acquisition: improving and extending the practices on comparing models against empirical infant data, especially when the available infant data is noisy, and increasing the ecological validity of the learning simulations. To tackle the first challenge, this work introduces a new evaluation framework that incorporates robust empirical data on infant language, accounts for multiple linguistic capabilities in parallel, and aligns with the experimental practices used to study infants’ language behaviours. This approach supports creating and assessing increasingly holistic models that better reflect infant learning trajectories and contributes to standardising model evaluation practices. To address the second challenge, the thesis explores how computational models can simulate infant learning under more realistic conditions than in the past studies. The selected self-supervised neural network models rely solely on acoustic speech input without linguistic priors, simulating statistical learning as a core mechanism of language development. Furthermore, this work integrates empirical estimates of daily speech exposure and pioneers the inclusion of prenatal language experience in modelling studies, an aspect rarely considered in previous research. The results of these simulations suggest that statistical learning and other cognitive hypotheses, such as the “less is more” hypothesis, are not necessarily mutually exclusive, and that statistical learning can operate despite limitations in the speech signal available to infants. Overall, the contributions of the thesis are bridging the gap between computational models of infant language learning and real-life language acquisition experiences by advancing more ecologically valid simulations of infant language learning and closer comparisons between models’ and infants’ linguistic behaviours that account for multi-capability learners and noisy behavioural data. The results of this thesis highlight the importance and challenges of developing more realistic simulations of language learning.
Language: English
Field: Psychology
Open Access status: green
OpenAlex ID: W7133019561
Cites: 273
Cited by:
Related to: 10