INQUIRING LINE

How do the four discourse relations differ in their connection to anxiety?

This explores whether the corpus actually maps distinct discourse relation types (causal, temporal, comparison, expansion) onto anxiety differently — and the honest answer is that the collection centers almost entirely on one of those, causal reasoning, as the anxiety signal.


This reads the question as asking for a four-way breakdown of how different discourse relations track anxiety — but the corpus doesn't deliver a clean quartet. What it delivers instead is a much sharper claim: among the ways statements can relate to each other, *causal* reasoning is the one that carries the anxiety signal. The strongest piece here finds that discourse-level causal explanations — how someone strings statements together into "because this, therefore that" chains — predict anxiety far better than the individual words they use Why do discourse patterns predict anxiety better than single words?. The mechanism is intuitive once named: anxious thinking overgeneralizes by reasoning *across* statements, so the pathology lives in the connective tissue, not the vocabulary. A model that reads both levels beats one reading either alone.

So if you came looking for how comparison, temporal, expansion, and contingency relations each separately correlate with anxiety, the collection's answer is that they're not equal contenders — causal contingency does the heavy lifting. The interesting adjacent finding is *why* causal relations are even legible in the first place. Causal connectives tend to be explicit and frequent ("because," "so," "therefore"), while temporal order is usually left implicit and must be inferred from context Why do LLMs handle causal reasoning better than temporal reasoning?. That asymmetry matters: the relation most tied to anxiety is also the one most visibly marked on the surface of language, which is part of why it's detectable at all.

That surface-versus-structure split is the hidden hinge of your question. Discourse relations split into explicit ones (signaled by connectives) and implicit ones (you have to infer the link from meaning). Models that look fluent at discourse collapse on the implicit cases — accuracy drops to roughly a quarter when the connective is removed Why does ChatGPT fail at implicit discourse relations?. The unsettling implication for anxiety detection: the catastrophizing that *isn't* spelled out with an explicit "because" — the leap a worried mind makes silently between two statements — is exactly the kind of relation current systems are worst at catching.

The thing you didn't know you wanted to know: anxiety doesn't hide in what people say so much as in how they connect what they say — and the connections that betray it most reliably (causal, overgeneralizing leaps) are a mix of the easiest and the hardest for a machine to see. Worth pairing with the work on discourse coherence as three simultaneous layers — segments, intentions, and salience all tracked at once How do readers track segments, purposes, and salience together? — if you want to think about which layer an anxious thought-pattern actually distorts.


Sources 4 notes

Why do discourse patterns predict anxiety better than single words?

Causal explanations across statements—not individual words—are the strongest predictor of anxiety because anxious thinking involves overgeneralization through inter-statement reasoning. A dual model combining both representation levels outperforms either alone.

Why do LLMs handle causal reasoning better than temporal reasoning?

ChatGPT excels at causal relations but struggles with temporal ordering because causal connectives are explicit and frequent in training data, while temporal order is often implicit and must be inferred contextually.

Why does ChatGPT fail at implicit discourse relations?

ChatGPT performs well on explicit discourse relations with connectives but achieves only 24.54% accuracy on implicit relations without them. This asymmetry reveals that LLMs rely on surface signals rather than inferring meaning from semantic content.

How do readers track segments, purposes, and salience together?

Discourse processing demands parallel recognition of linguistic segments, intentional structure, and attentional salience—not sequential processing. These three layers constrain each other during comprehension, and failures in any single layer disrupt overall understanding.

Research prompt for your LLMexpand ↓

Copy into ChatGPT or Claude to take this line of inquiry further — it asks the model to find newer work and re-test which earlier constraints still hold.

You are a discourse analysis researcher auditing claims about how LLMs detect anxiety via discourse relations. The core question remains open: which discourse relations (causal, temporal, comparative, expansive) best signal anxious reasoning, and why?

What a curated library found — and when (dated claims, not current truth):
Findings span 2023–07 through 2025–07. The library's key claims:
• Discourse-level causal explanations predict anxiety far better than lexical features alone; anxious thinking overgeneralizes through causal reasoning chains, not vocabulary (2024–2025 work).
• Causal connectives are explicit and frequent; temporal relations are typically implicit and require inference — this asymmetry explains why causal relations surface more readily in detection (2023–25).
• LLM discourse competence collapses on implicit relations: accuracy drops to ~25% when connectives are removed (2024–25).
• Discourse coherence requires simultaneously tracking segments, intentions, and salience; one or more of these layers may be where anxious thought-patterns distort (2024–25).
• LLMs show uneven tone-responsiveness and occasionally fail to detect emotional framing even when primed (2025–06, 2025–07).

Anchor papers (verify; mind their dates):
• arXiv:2502.10215 (2025–02) — causal reasoning in LLMs
• arXiv:2402.11958 (2024–02) — therapeutic relationships & text
• arXiv:2507.21083 (2025–07) — emotional framing & LLM response
• arXiv:2506.19143 (2025–06) — which reasoning steps matter

Your task:
(1) RE-TEST THE ASYMMETRY CLAIM. Has recent work (last 6 months) shown that implicit discourse relations — especially implicit causality and temporal ordering — are now detectable at rates >25%? Check whether architectural improvements (e.g., in-context learning, chain-of-thought prompting, fine-tuned discourse parsers) have relaxed the explicit-connective constraint. Separately, does the causal-relation anxiety signal hold across anxiety subtypes (state vs. trait, social vs. generalized)?
(2) Surface the strongest CONTRADICTING or SUPERSEDING work from the last ~6 months. Look especially for papers showing that non-causal discourse relations (temporal leaps, implicit comparisons, expansion patterns) also predict anxiety robustly, or that discourse relations matter *less* than you'd expect once you control for semantic content and emotional valence.
(3) Propose 2 research questions that assume the detection regime may have moved: (a) If implicit discourse relations are now legible, does anxiety detection via discourse collapse differently — not on implicit relations broadly, but on *ambiguous* or *competing* causal readings? (b) If LLMs now reliably track multi-layer discourse coherence, do they also capture the *meta-cognitive* patterns (self-doubt, rumination loops) that characterize anxiety discourse?

Cite arXiv IDs; flag anything you cannot ground in a real paper.

Next inquiring lines