Mapping relational mechanism clusters in Of Human Bondage: a theory-driven multiscale embedding analysis.

INTRODUCTION

Some psychological patterns appear as events; others become visible only as sequences. Literary fiction is unusually rich in such sequences, because it records how attachment, injury, dependence, and release accumulate across time. This study asks whether the language of one novel-W. Somerset Maugham's Of Human Bondage-is organized around specific psychological mechanisms, using theory-guided text embeddings.

METHODS

We wrote short "anchor" texts for eight relational or existential constructs and two negative-control domains, then measured, in embedding space, how closely each chapter and scene matched each anchor. At chapter level (122 chapters), we tested whether construct signals peak at the same chapters, whether the Mildred Arc differs from the Resolution segment, and whether neighbouring chapters resemble one another. At scene level (165 scenes from Chapters 55-122), we identified which scenes drive chapter peaks. Three embedding models tested robustness; six independent readers rated 20 scenes for external validation.

RESULTS

The construct signals clustered at the same chapters far more than chance permutations allow. In the late novel, four constructs-a narcissistic relational style, coercive control, intermittent reinforcement, and trauma bonding-were higher in the Mildred Arc than in the Resolution segment, while other constructs were not. Scene analysis showed that each chapter peak came from a few distinct local scenes rather than from uniform elevation. The negative controls did not follow the toxic-relational pattern; they peaked instead in reasoning-heavy and travel passages. The pattern held across all three models; the largest model did not give the cleanest separation. Human ratings of key scenes converged with the model on the core constructs (mean Spearman ρ = 0.46, individually significant for the narcissistic-style, coercive-control, trauma-bonding, helplessness, and existential signals) but not on three broader anchors.

DISCUSSION: Without diagnosing fictional characters, the study shows how theory-guided anchors can map where a novel organizes interpersonal mechanisms, combining chapter-level trajectories, scene-level localization, negative-control anchors, cross-model checks, and human-rater validation. The workflow is portable to other long-form literary texts.