Why do discourse patterns predict anxiety better than single words?

The primary clinical manifestation of anxiety is cognitive distortion — illogical reasoning about life events. The key insight is that these distortions are expressed at the discourse level, not the lexical level. Single words or words in context are insufficient to detect them.

Consider the catastrophizing statement: "I'm sick. Now I'm going to miss my classes and fail them all." To recognize that "fail them all" catastrophizes "I'm sick" requires understanding causal explanation across statements — this is discourse-level semantics, not lexical features.

Four discourse relations are relevant to anxiety detection:

• Causal explanations — statements of why an event happened (strongest predictor, related to overgeneralization)
• Counterfactuals — imagining alternatives to actual events
• Dissonance — behavior or belief contradicting a prior belief
• Consonance — the opposite of dissonance

All four discourse dimensions correlate with anxiety scores, but causal explanations show the highest difference between high and low anxiety groups. The mechanism: anxious individuals overgeneralize through causal reasoning — "You know life is going to be permanently complicated when your in-laws start turning their backs on you like a domino effect."

A dual lexico-discourse model combining both representation levels outperforms either alone, suggesting lexical and discourse features capture complementary information about cognitive state.

The Diagnosis of Thought (DoT) prompting framework operationalizes this insight for therapeutic chatbots. DoT uses a structured three-stage process to detect cognitive distortions: (1) subjectivity assessment — identifying whether a statement contains subjective elements, (2) contrastive reasoning — comparing the statement against an objective baseline to identify distortion, and (3) schema analysis — classifying the distortion into one of 10 cognitive distortion types from CBT (catastrophizing, overgeneralization, mind-reading, emotional reasoning, should statements, labeling, personalization, black-and-white thinking, mental filtering, fortune-telling). Since Can structured prompting improve cognitive distortion detection?, DoT provides evidence that discourse-level cognitive patterns can be detected computationally — but only through structured multi-stage reasoning, not through end-to-end classification. The three-stage decomposition mirrors the discourse-level analysis this note advocates: detecting distortions requires understanding the reasoning structure between statements, not just classifying individual statements. The limitations of word-counting approaches are concrete: LIWC-style methods cannot handle negation ("not bad"), sarcasm, or context-dependent polysemy, and manually defined dictionaries omit synonyms. Transformer-based models resolve these by leveraging proximal cues, but even they default to lexical-level features unless explicitly designed for discourse-level reasoning.

The implication for therapeutic AI is direct: since Why does ChatGPT fail at implicit discourse relations?, current LLMs may struggle precisely where anxiety detection matters most — at the implicit discourse relations that reveal cognitive distortions. A chatbot that detects sentiment words ("sad," "worried") but misses discourse-level causal reasoning patterns will miss the cognitive structure of anxiety.

This also connects to the observation that since Do LLM therapists respond to emotions like low-quality human therapists?, therapeutic chatbots are not just failing at emotional attunement — they may be failing at the discourse-level comprehension needed to even detect what kind of cognitive distortion is occurring.

Source: Psychology Users; enriched from Psychology Therapy Practice