Towards Construct-Valid Computational Analysis of Human Selfhood

Computational approaches are increasingly used to identify Self-related constructs in text, but predictive success alone does not establish that a model validly captures the intended aspects of human selfhood. We argue that computational analysis of the Self requires alignment between the theoretical constructs being studied, the linguistic evidence through which they are operationalised, and the outputs produced by computational models. We develop this argument through a structured representation of selfhood in which distinguishable Self-aspects may interact, share cross-cutting dimensions, and form higher-level configurations. We further consider how these constructs become observable in different forms of text and how this should inform annotation and model selection. On this basis, we outline a computational framework in which modelling approaches are evaluated not only for predictive performance, interpretability, grounding, and efficiency, but also for construct fidelity and evidential traceability. Rather than proposing a single optimal model of the Self, we argue for computational methods whose assumptions and outputs remain aligned with the structure of the phenomenon they aim to measure.