Artificial Intelligence for Automatic Attachment Assessment in School-Age Children: An Approach Based on Language and Paralanguage.

Abstract

Attachment is a psychological construct that provides a framework for understanding how individuals perceive and interpret social interactions, navigate relational dynamics, and experience and regulate their emotional states, particularly under conditions of stress. An attachment style begins to develop within the first few months of life, shaped by a child’s interactions with their primary caregivers. Consistent and nurturing care promotes the development of a secure attachment style, whereas inconsistent or inadequate caregiving often gives rise to insecure attachment patterns. Insecure attachment is linked to a range of challenges, including behavioural issues such as antisocial tendencies; mental health difficulties like anxiety, emotional dysregulation, and body image concerns; and heightened risks of physical health problems, including sleep disturbances.

Early recognition and intervention for insecure attachment increases the likelihood of reshaping maladaptive patterns into secure ones, potentially reducing attachment-related challenges. Automated approaches for attachment recognition offer significant benefits, including consistent delivery of assessments, such as the MCAST, and broader accessibility to a wider population. While there are a few available systems for delivering attachment tests (e.g., CMCAST and SAM), the limited studies focused on developing automated classifiers to analyse the collected data have shown a suboptimal performance. These classifiers often struggle to recognise insecure attachment, achieving a maximum Accuracy of only 62.7%. Furthermore, these studies fail to offer insights into the reasoning behind their classifications, missing an opportunity to advance the understanding of attachment in early to middle childhood. This developmental stage—characterised by significant changes that include the expansion of social circles and the internalisation of emotional representations—has historically received less attention in a field predominantly focused on studying attachment markers in infants and adults.

This thesis focuses on two primary objectives: enhancing the automated classification of attachment styles in children, particularly insecure attachment, and identifying markers associated with these styles. The study employs two modalities—language and paralanguage— along with emotions derived from both modalities. These modalities are utilised within a unimodal and a multimodal framework. Among all classifiers developed using the same dataset, the language-based unimodal approach demonstrated the highest effectiveness, achieving exceptional performance in recognising insecure attachment with an Accuracy of 82.2%, all while relying on relatively simple methodologies. Furthermore, this research identified linguistic, acoustic, and emotional markers of attachment, offering valuable insights into attachment representations in children.