DefMoN: A reproducible framework for theory-grounded synthetic data generation in affective AI

Modern NLP systems excel when labels are abundant but struggle with high-inference constructs that are costly to annotate and risky to synthesize without constraints. We present the Defensive Motivational Node framework, henceforth DefMoN (formerly DMN), which operationalizes Vaillant’s hierarchy of ego defenses and Plutchik’s psychoevolutionary emotions into a controllable generative process for text. We release DMN-Syn v1.0 — a quadri-lingual (EN/KO/FR/KA) corpus of 300 theory-constrained utterances — together with a complete, versioned research compendium (data, code, seeds, QC manifests, and evaluation scripts) archived on Zenodo (Kim, 2025). The full package is permanently available at https://doi.org/10.5281/zenodo.17101927 . On the modeling side, we treat defense recognition as 10-way sentence classification and fine-tune a multilingual Transformer (XLM-R) only on DMN-Syn v1. In-domain performance is high (EN macro- , MCC ; KO macro- ), and zero-shot transfer is strong (EN KO macro- ). When evaluated on a small, anonymized real-world benchmark, the model reaches Macro with zero real training data, then rises to 0.76 with only supervised examples per class. Human annotators on that same benchmark agree with each other at , . This shows that DefMoN is not a turnkey classifier, but a theory-grounded primer that enables data-efficient adaptation toward human-level ambiguity theory-grounded primer that enables data-efficient alignment to schema-coded human benchmark without large-scale annotation. We additionally quantify reliability —reporting ECE/MCE and coverage–performance curves for selective prediction—and show robustness under group-aware splits (template/scenario disjoint) and cue ablations, establishing structural coherence in the absence of large-scale human trials. Beyond raw scores, we foreground auditability . Each instance in DMN-Syn v1 is generated with fixed seeds, grouped splits, and guardrails that prevent label leakage and construct drift; validators, manifests, and code are released for byte-exact replication. Results support theory-constrained synthesis as a practical middle path between costly expert labeling and unconstrained LLM generation, particularly for low-resource and cross-lingual settings. Significance. By using psychological theory as an explicit generative constraint rather than a post hoc interpretation, DefMoN reframes synthetic data work as the operationalization of constructs for machine learning. The framework (i) standardizes guardrails to minimize bias amplification and drift, (ii) supplies small but theory-dense corpora that train reliable classifiers with uncertainty awareness, and (iii) provides auditable artifacts (seeds, manifests, validators) enabling reproducibility and extension to new defenses, languages, and dialogue-level settings. Terminology and branding. In prior releases and repositories we used the acronym DMN to denote Defensive Motivational Node . To avoid confusion with the neuroscience“Default Mode Network,” this article adopts the acronym DefMoN for the overall framework. Legacy dataset and repository names (e.g., DMN-Syn v1 ) are preserved for continuity and for reproducibility of prior work. Throughout the paper, we therefore use: DefMoN for the overall framework and method, DMN-Syn v1 for the released dataset and its artifacts, and “DMN node” to refer to an individual ( Defense , Emotion , Scenario ) tuple within that dataset. This split is intentional.