57 Defense Mechanisms: How We Built a Working Taxonomy
From psychoanalytic theory to computational ontology — the story of creating a classification that works on real speech
Why 57?
The DSM and ICD describe defense mechanisms as a list of 20–30 items. For clinical observation, that's sufficient. For speech analysis — it isn't. When someone says "well, that's normal, everyone does it" — is that NORMALIZATION, MINIMIZATION, or RATIONALIZATION? In a clinical context, the difference is critical.
The Principle of Construction
We didn't start from theoretical classifications, but from speech data. The process:
- Collected a corpus of 3,168 annotated speech fragments
- Each fragment labeled by F-code (functional group) and specific defense
- Clustered by speech markers and usage context
- Tested for discriminability — can the model distinguish two similar classes?
The result: 57 classes, 7 functional categories (CONTROL, AVOIDANCE, DISTORTION, CONCEALMENT, COMPENSATION, EXTERNALIZATION, FRAGMENTATION).
Problem Zones
GASLIGHTING — not a single defense, but a pattern of two components: fact denial + invalidation of the interlocutor's reality. We decomposed it into two markers, and accuracy jumped from 65% to 94%.
FALSE_REPAIR — the most insidious class. "Let's just forget about it" looks like reconciliation, but contains a hidden trap (a condition that makes "forgetting" impossible). Without a trap marker — it's simply AVOIDANCE.
INVERSE_IDEALIZATION — the 57th class, added last. Self-elevation at the expense of devaluing another. It masqueraded as CONTROL_ILLUSION + BINDING for a long time.
Current Metrics
RU pipeline: 94.4% accuracy, 98.6% top-3. EN pipeline: 94.6% accuracy. Both frozen.
The key insight: a taxonomy is not a list, it's a graph with edges reading "similar to / differs because." Without those edges, a classifier inevitably confuses neighboring classes.