How people tell their stories reveals mental health more than word choice alone

Researchers have long tried to understand mental health through the language people use—counting depressive words, anxious phrases, or trauma-related terms. But a new study suggests this approach misses something crucial: the way people organize their stories.

A team working with 830 Chinese therapeutic texts has introduced a three-level framework that moves beyond simple word-counting. Rather than treating language as a bag of disconnected terms, they show that how narratives are structured—at the level of plot, coherence, and discourse organization—carries far more clinical signal than vocabulary alone.

The three levels of narrative analysis#

The framework breaks down computational approaches into three tiers:

• Micro-level: Traditional lexical features (word counts, sentiment dictionaries)
• Meso-level: Semantic embeddings that capture meaning through neural representations
• Macro-level: Large language model evaluation of narrative structure itself—including story grammar, rhetorical coherence, and how propositions fit together

Why structure beats vocabulary#

The key finding: macro-level narrative evaluation substantially outperforms both lexical and embedding-based methods for predicting depression, anxiety, and trauma. The researchers grounded their analysis in formal discourse theory, examining Labov's story grammar (how events are sequenced), Rhetorical Structure Theory (RST) coherence (how ideas connect), and propositional composition (the logical relationships between claims).

These structural features reveal something that word-counting cannot: how a person's psychological state manifests through the organization of experience itself. Two people might use identical vocabulary but construct their narratives very differently—and those structural differences predict mental health outcomes more reliably.

Practical implications#

Semantic embeddings—widely adopted in recent years—add minimal independent value on their own, though they do contribute incremental improvements when combined with other levels of analysis. This challenges the field's recent emphasis on neural embeddings as a replacement for simpler linguistic features.

By grounding computational methods in discourse processing theory rather than treating language as statistical patterns, the framework opens new directions for intervention design and longitudinal research.