Language or Culture? Diagnosing Zero-Shot Transfer of Performance-Anxiety Detectors

Many people feel strained by keeping up a polished version of themselves online: the exam results they report, the career they project, the family they reassure. Detecting this kind of distress in social media text is almost entirely an English-language effort. The standard way to reach other languages is zero-shot cross-lingual transfer: train a multilingual model on English labels, then apply it unchanged to German, Russian or Hindi. This works for tasks like question answering. Anxiety is different, because what people are anxious about, and how they say it, changes with culture, not just with vocabulary. This note proposes a study that separates the two. We train an English anxiety detector and measure how much accuracy it loses on each of three targets: English written by Indian users (a culture shift with no language shift), German (a language shift with roughly similar Western norms), and Russian (both shifts at once).

Our hypothesis is specific. The English training data mostly shows distress about failing one's own ideals, written in self-focused first-person "I-talk". Much non-Western distress is about failing obligations to others, especially family and academic duty. Psychology already names this split: Higgins' distinction between the gap from the "ideal self" and the gap from the "ought self". We will label a small English sample for which kind of gap a post expresses. We will then train the detector so that ought-self distress is not drowned out, still using no target-language labels. The study succeeds if this closes a measurable share of the culture gap and a cheaper surface fix, reweighting by pronoun use, does not.

Core Claim

For performance-related anxiety, most of the zero-shot transfer gap is cultural, not linguistic. That cultural gap is concentrated in one identifiable sub-construct, ought-self discrepancy, which is under-represented in Western English training data. If this is true, collecting data in every language is not the main remedy. The remedy is to rebalance the kinds of distress the source data covers, which costs a few thousand English annotations rather than a new corpus per language. If it is false, and the culture-only gap is small next to the language gap, then the field's usual fixes (better multilingual alignment, more target-language data) are the right ones. Either answer is useful, and the 2×2-style design is what makes it possible to tell them apart.

What Prior Work Establishes

Work on digital performance anxiety motivates the construct but offers no detector. Self-Discrepancy: The Discrepancy Between Digital Identity and Real Identity on Social Media and Its Psychological Impact maps retouching and selective posting onto Higgins' actual/ideal/ought framework. A two-wave panel study, The Double-Edged Influence of Self-Expansion in the Metaverse, shows that seeing one's virtual self as superior lowers self-esteem over time. The Evolution of Occupational Identity in Twitter Biographies shows at scale that online self-descriptions drift toward prestige, but only for US English users. We take Higgins' split from this literature as our annotation scheme. These papers rely on surveys or English-only corpora. None builds a classifier, and none asks whether their constructs hold outside WEIRD (Western, educated, industrialized, rich, democratic) samples.

Work on cross-lingual transfer supplies the method and also warns about its limits. Crosslingual Generalization through Multitask Finetuning shows that English-only instruction tuning improves zero-shot non-English performance: BLOOMZ-P3 gains 50% on non-English XNLI over base BLOOM. Its mT0 models are our prompting baseline. MLQA documents a persistent gap even on a task whose meaning is the same in every language, with Hindi reaching only 48.8 F1. XPersona shows zero-shot transfer collapsing for persona-consistent dialogue generation. Its "persona" is a list of stated facts, not a psychological self, so it does not tell us whether distress transfers.

The strongest constraint on our design comes from Cross-Cultural Differences in Mental Health Expressions on Social Media. Using English Reddit, it finds that Indian users express distress through second- and third-person reference, family and academic duty, and advice-seeking, while Western users use I-talk and clinical vocabulary. That paper predicts cultural cohort; it never trains a condition detector or tests transfer. Its corpus is what lets us isolate culture from language. Two monolingual studies constrain what counts as success. SMHD-GER builds a German Reddit corpus by adapting English self-disclosure patterns and finds hand-built psycholinguistic features beating German BERT. Exploring Large Language Models for Detecting Mental Disorders finds anxiety detection in Russian stuck at 55–68% F1-macro even with in-language training. Neither tests transfer from English.

Cross-Lingual Mental Health Ontologies for Indian Languages proposes a knowledge graph of Hindi, Kannada and Marathi idioms of distress. It reports no empirical results. We treat it as a possible source of probe items, not as a benchmark.

We considered and rejected one tempting bridge: grounding the three-dimensional affect model of Bayesian Affect Control Theory of Self in multilingual embeddings. That model measures inauthenticity as a divergence between situational and core identity. It has only been run in synthetic simulations, and no labeled text exists against which to validate an embedding-to-affect mapping. Building it would add an unfalsifiable layer before the basic transfer question is answered.

Research Program

The source domain is English posts from Western users. We take the MH-RoW cohort of the Rai et al. corpus (about 300,000 posts from 4,530 users in total, split across cohorts). Anxiety-positive users are those active in anxiety-focused subreddits. Controls are drawn from general subreddits with matched activity. We will check that subreddit-derived labels behave like the self-disclosure labels used in SMHD-GER, and we will report where they disagree.

There are three targets:

  • MH-India, the Indian-user English cohort from the same corpus, built by the same procedure. This is the culture-only shift.
  • The anxiety-disorder and control cohorts of SMHD-GER. Its labeling procedure was itself adapted from English, which keeps label definitions comparable. This is the language-dominant shift.
  • The Russian anxiety datasets of Kuzmin et al. These carry both shifts.

If the idiom inventory behind the Indian-language ontology is obtainable, we add a small Hindi probe set built from it. We will describe that set as a probe, not a benchmark.

The central annotation is a facet label for performance-related distress. Licensed psychologists will label about 2,000 English source posts and about 500 posts from each target. Each post gets one of three labels:

  • Ideal-self discrepancy: distress about falling short of one's own aspirations or projected image.
  • Ought-self discrepancy: distress about failing duties or others' expectations.
  • Neither.

Target-side facet labels are used only for analysis, never for training. The first analysis tests the mechanism directly. We predict that ought-facet posts are rare among English source positives, and that the baseline detector's false negatives on MH-India are concentrated in ought-facet posts well beyond their base rate.

The intervention is deliberately small. A multilingual encoder (XLM-R large) is trained with two outputs: the anxiety label, and an auxiliary head that predicts the facet. The anxiety loss is reweighted so that ought-facet and ideal-facet positives contribute equally in total. Facet labels exist only for the annotated subset. A facet tagger trained on that subset propagates weights to the rest of the source data, and we report the tagger's own accuracy. No target-language or target-culture examples enter training or model selection. Hyperparameters are chosen on held-out English-Western data only, because tuning on target data would quietly break the zero-shot claim.

We measure four quantities for every method:

  • F1-macro at the user level, on each target.
  • The culture gap: in-domain F1 minus MH-India F1.
  • The language gap: in-domain F1 minus SMHD-GER F1.
  • Recall on ought-facet positives in each target.

We will describe the German and Russian comparisons carefully. They also differ in platform, time period and labeling details, so the "language gap" is really "language plus residual differences". The MH-India comparison is the cleanest, because it holds platform and collection procedure fixed.

Baselines

The naive reference points are a majority-class predictor and a detector trained and evaluated on English-Western data only. The in-domain English score sets the ceiling from which both gaps are measured. Any method that reports target scores without this reference cannot say whether it closed a gap.

The standard zero-shot baseline is the same XLM-R model fine-tuned on the same English source data with plain cross-entropy loss. It uses identical data, number of training steps and hyperparameter search budget, and it is selected on the same English validation split. This is the comparison that isolates our contribution, since the only difference is the facet head and the reweighting.

The strongest instruction-following baseline is zero-shot prompting of mT0-xxl from Crosslingual Generalization through Multitask Finetuning. It uses English prompts and one prompt template selected on English validation data. It is the strongest competitor because instruction tuning is exactly what the file shows transferring best to unseen languages. We also run a zero-shot open LLM prompt, in the spirit of the models evaluated by Kuzmin et al., under the same prompt-selection budget.

The strongest reported figure for anxiety detection in any target we use is the monolingual Russian result from Exploring Large Language Models for Detecting Mental Disorders. Its best models, trained or prompted in Russian, reach "between 55% and 68%" F1-macro for anxiety, and its noisy social-media dataset DSM tops out at 63.69%. For SMHD-GER, no reported result found on Lacuna for exact anxiety F1. We will therefore train an in-language GBERT and a psycholinguistic-feature model on SMHD-GER as a supervised ceiling. We cannot claim a number from the paper.

We count two outcomes as beating the baselines. The first is closing at least half of the MH-India culture gap relative to plain zero-shot XLM-R, with confidence intervals from user-level bootstrap. The second is a zero-shot Russian anxiety F1-macro that falls within the 55–68% range reported for in-language models.

Three ablations isolate the key ingredient:

  • Random facets: reweighting with facet labels shuffled, to rule out generic reweighting effects.
  • Facet head without reweighting: to separate representation learning from loss balancing.
  • Pronoun reweighting: weights derived from the ratio of first-person to other-person pronouns, which requires no annotation.

The pronoun ablation is the one we most need to beat.

Why not just…?

Why not just translate target posts into English and apply the English model? Translation removes the language gap by construction, but it cannot remove the culture gap. An Indian user writing in English already needs no translation, and our hypothesis predicts the detector still misses their ought-facet distress. The file points the same way. Zero-shot Reading Comprehension by Cross-lingual Transfer Learning with Multi-lingual Language Representation Model reports that routing through translation degrades comprehension relative to multilingual representations. The Indian-language ontology work argues that literal translation strips somatic idioms of distress. We will run translate-test on German and Russian anyway. If it matches our method there, we will report that honestly: the language half of the gap would then be solved by translation, and our contribution would reduce to the cultural half.

Why not just reweight by pronoun use, since Rai et al. already found the I-talk difference? This is the most serious threat. If pronoun ratio is a good proxy for ideal versus ought distress, the psychologist annotation is wasted effort. It may not be a good proxy. Ought-self distress can be written in the first person ("I can't face my parents after these marks"), and Rai et al. document a correlation, not an equivalence. The pronoun-reweighting ablation settles this directly. If it closes as much of the MH-India gap as facet reweighting does, we will adopt it as the recommended method and report the facet construct as explanatory rather than necessary.

Why not just collect a few labeled examples in each target and skip zero-shot altogether? If 64 target examples close the gap, arguing about a zero-shot method is pointless for high-resource targets. We therefore fine-tune the baseline with 16, 64 and 256 target-labeled users and plot F1 against labeling cost. The claim we can defend is narrower: facet rebalancing is a one-time English cost that should help every culturally similar target at once, including the many Indian languages for which Kandala et al. note that native data collection is infeasible. If a few dozen target examples beat our method everywhere, the proposal's practical value is limited to languages where even that is unavailable. We will say so.

Why not just prompt an instruction-tuned LLM? In-language prompting already caps out at 55–68% F1-macro for Russian anxiety. Psychologists rated LLM explanations at only 2.84/5 in the same study, which leaves little room to trust these models' judgments of subtler, duty-framed distress. The mT0 baseline tests this directly, and its ought-facet recall is the specific number to watch.

Decisive Experiments

The first experiment decides whether the premise holds. Train the plain baseline and compare the culture gap on MH-India with the language gap on SMHD-GER. If the MH-India gap is within bootstrap noise of zero, the cultural hypothesis is wrong for this construct, and we stop and report that. The second experiment tests the mechanism: are baseline false negatives on MH-India over-represented among ought-facet posts, relative to the facet distribution annotators find among true positives? The third experiment tests the fix. Facet-balanced training should raise ought-facet recall and overall F1 on MH-India. It must beat random-facet and pronoun reweighting, and it must not lower in-domain English F1 by more than about one point. The fourth experiment tests generality. Does the same English-trained model also improve German and Russian F1, even though it never saw those languages? That would show the cultural fix and multilingual transfer stack together. If it helps only MH-India, the facet imbalance is specific to Indian cultural framing rather than to non-Western distress in general.

All data are public research corpora, analyzed at the user level. No model is deployed or used to flag individuals, and annotators work under clinical supervision.