Psycholinguistic Lexicons as a Free Concept Bottleneck Against Shortcut Learning in Depression Detection

Text-based depression detectors often score well for the wrong reasons. On the most widely used clinical interview corpus, a model that reads only the virtual interviewer's questions, and never the patient's answers, predicts depression better than models that read the patient. Fine-tuned language models also serve women measurably worse than men. Computer vision has a partial remedy: force the model to first predict human-meaningful clinical signs, then predict the diagnosis only from those signs. This "concept bottleneck" blocks many shortcuts. It is expensive, though, because someone has to label the signs on every training example.

Psychology already has a large, expert-built set of such signs for text: lexicons of affect, subjectivity, self-focus, cognitive processing and bodily complaints. We propose to use those lexicons as zero-annotation concept labels. A multilingual text encoder is trained to predict lexicon-derived concept scores for each patient turn. A small, sparse classifier then predicts depression from those predicted concepts alone. We then test whether this routing reduces three failure modes: reliance on interviewer prompts, gender performance gaps, and collapse when moving to other languages and interview styles. The test has to hold against the obvious cheaper option of simply counting lexicon words.

Core Claim

Routing a learned encoder through lexicon-defined concepts retains most of an end-to-end model's accuracy while doing three things. It resists the interviewer-prompt shortcut even when prompts are left in the input. It narrows the gender gap without using gender labels in training. And it transfers across languages better than raw lexicon counts can, because the encoder learns to recognise a concept's paraphrases, slang and translations that a fixed word list scores as zero.

The proposal also carries a caveat that the file's framing tends to skip. Lexicons are not demographically neutral. Understanding gender and age differences in language use found that gender correlates with dictionary categories, for example female Weibo users with "biological processes" at $r = 0.245$. A lexicon bottleneck could therefore reintroduce gender shortcuts through communal or bodily vocabulary. The program audits each concept for exactly this and reports it as a first-class result, whichever way it comes out.

Relation to Prior Work

Two papers supply the method. Concept Bottleneck Models predict human-defined concepts and then the label. That paper showed bottlenecks lose little accuracy (CUB error 19.9% vs 17.5% end-to-end) and resist a background shift far better (48–49% vs 62.7% error on CUB+Places). Fairer AI in ophthalmology via implicit fairness learning forced disease predictions through 67 annotated retinal lesions and narrowed sex and age gaps without demographic labels. Both depend on manual concept annotation and both are image-only. We replace annotation with lexicon scoring on text, which turns their cost into a free supervision signal. We also inherit the known concept-completeness and leakage risks, which we test directly.

Three papers motivate the proposal by documenting the failure. DAIC-WOZ: On the Validity of Using the Therapist's prompts showed that a graph model trained on interviewer prompts alone reaches F1 0.88. A Longformer rises from 0.64 on participant text to 0.80 when prompts are added. Speech and Text Foundation Models for Depression Detection found female F1 trailing male by 0.05–0.1 in English clinical interviews. Principles from Clinical Research for NLP Model Generalization showed that deep models can track surface heuristics, with surrogate agreement reaching $\kappa = 0.85$ out of distribution. We reuse its surrogate audit as a shortcut metric.

Mitigating Interviewer Bias in Multimodal Depression Detection is the closest competing mitigation. It applies adversarial gradient reversal, a layer that pushes the encoder to hide information a side classifier tries to recover. Its side labels are interviewer question functions, and those labels come from GPT-4o. It removes one named shortcut, whereas our bottleneck needs no shortcut labels at all.

Several papers show that lexicons plus topics carry real signal but never used lexicons as a constraint on a deep model. Using Topic Modeling to Improve Prediction of Neuroticism and Depression, Triaging content severity in online mental health forums and Head versus heart all concatenate lexicon and topic features. In the first, LIWC alone gave $r = .288$ for depression and adding LDA topics gave $r = .416$. This is a constraint on our design: pure lexicons are incomplete. We therefore test a narrow open-vocabulary side channel as a variable rather than assuming it away.

Research Program

The concept inventory is built from openly available lexicons: NRC emotion, MPQA subjectivity and DepecheMood, as used by Cohan et al. We add small construct lists for first-person self-focus, cognitive-process words, and somatic or vegetative complaints, following Resnik et al. and Liu and Ungar. This yields roughly 40–60 concepts. For each participant turn, the concept target is the normalised lexicon score. In the bottleneck vocabulary, $x$ is a turn, $c(x)$ its concept vector and $y$ the PHQ-8 depression label.

A multilingual encoder, XLM-RoBERTa (a pretrained transformer covering many languages), is trained to regress $c(x)$ from English turns. To make it learn concepts rather than memorise word lists, we mask the lexicon words themselves in a fraction of training turns. The encoder must then infer "negative affect" or "self-focus" from context. Turn-level predicted concepts are pooled over the interview, and depression is predicted by a sparse logistic head.

Training is sequential, following Koh et al.: concept model first, then the head on predicted concepts. This limits leakage, meaning label information smuggled through the soft concept values. As a leakage control, we also train the head on quantised concepts.

A side-channel variant adds $m$ LDA topic proportions or free latent dimensions to the head. We sweep $m$ from 0 to a large value to trace how accuracy and shortcut reliance trade off as the bottleneck opens.

The audit arm tests each concept for population dependence. For each concept we estimate its association with gender among non-depressed participants, and whether its association with depression differs by gender. When concepts carry demographic style, we report which ones and test whether dropping or group-standardising them helps. Group-standardising requires group labels at deployment, so this is flagged as a separate regime.

Baselines

The naive reference is a majority-class predictor. Alongside it we run a participant-text Longformer, for which DAIC-WOZ: On the Validity of Using the Therapist's prompts reports depression-class F1 0.64 on DAIC-WOZ. That number is the accuracy level our model must roughly retain: within 0.03 F1 counts as "retained". We rerun this model ourselves with our splits.

The end-to-end competitor is XLM-RoBERTa fine-tuned directly on depression labels. It uses the same encoder, the same data and splits, and the same hyperparameter-search budget of equal trials and GPU hours as the bottleneck. This is the strongest fair comparison, because it differs from our model only in the missing bottleneck. For fully multimodal numbers, the best text-plus-speech system in Speech and Text Foundation Models reports F1 0.75 on DAIC+. We cite this as context only, since it uses audio.

The strongest existing debiasing method is the adversarial interviewer-debiasing approach of Zhang et al., reimplemented on the same text encoder. We also run a gender-adversarial variant, which uses gender labels that our method does not. No reported F1 for this method on DAIC-WOZ was found on Lacuna, so we rely on our reimplementation and report its tuning budget. Beating it means one of two things: a smaller gender gap and prompt-shortcut gain at equal F1, or equal fairness without any shortcut or demographic labels.

The key ablation is raw lexicon counts fed to the same sparse head, with no learned encoder. This isolates whether the learned concept predictor adds anything beyond word counting. A second ablation adds the lexicon features alongside the end-to-end encoder by concatenation, which is the standard hybrid of Cohan et al. and Resnik et al. It tests whether the constraint, rather than the features, is what matters.

Why not just…?

Why not just use lexicon counts directly? A count-based model is already a concept bottleneck with a fixed concept extractor, and it may be equally shortcut-resistant. This is the most serious threat. Its expected weakness is lexical blindness: an English list scores Chinese or Italian turns, and unlisted slang, at zero. That weakness should show up mainly in the cross-lingual tests. If count models match ours in-language and across languages on both accuracy and fairness, the honest conclusion is that the learned encoder is unnecessary. We would report that plainly as a useful negative result.

Why not just drop the interviewer's turns? For DAIC-WOZ this removes the prompt shortcut trivially, and all our main comparisons use participant-only text for that reason. However, speaker-separated transcripts are not always available, and dropping prompts does nothing for the gender gap. The gap reported by Gómez-Zaragozá et al. arises in participant-centred models. We keep the prompts-included condition only as a stress test of whether the bottleneck resists a known shortcut it was never told about.

Why not just fine-tune end-to-end and fix fairness with post-hoc thresholds? Exploring Text Specific and Blackbox Fairness Algorithms found that equalised-odds post-processing lowered AUC. Towards Gender Fairness for Mental Health Prediction found that standard mitigations failed on complex behavioural data. Thresholding also needs group labels at deployment and cannot help cross-lingual transfer. We run it on the end-to-end model so the accuracy-for-fairness exchange rate can be compared directly.

Proposed Experiments

All experiments use DAIC-WOZ and E-DAIC with PHQ-8 labels and gender metadata for training and in-domain testing. Each model is trained on English only and evaluated zero-shot on three other corpora. EATD is a Chinese interview corpus with more generic prompts. The Androids Corpus is an Italian interview corpus. DEPTALK is Spanish casual conversation. Together they span cross-language and cross-elicitation shift.

We report depression-class F1 and the male–female F1 gap. We also report equal opportunity, the ratio of true-positive rates across genders, because Towards Gender Fairness for Mental Health Prediction showed fairness metrics disagree. The shortcut score has two parts. The first is Elangovan et al.'s surrogate $\kappa$: a Naive Bayes model on turn count, response length and prompt identity, whose agreement with each model's predictions is measured. The second is the F1 gain from adding interviewer prompts back into the input, benchmarked against the 0.64→0.80 jump reported by Burdisso et al. Every comparison uses at least five seeds and bootstrap confidence intervals, because these test sets are small.

Four checks decide the outcome. First, with prompts included, the bottleneck's F1 should rise far less than the end-to-end model's, and its surrogate $\kappa$ should be lower. Second, the gender gap should shrink below the end-to-end model's without gender labels in training. Matching the gender-adversarial baseline on that gap would be a strong result. Third, on EATD and Androids, the learned bottleneck should beat raw lexicon counts. If it does not, the encoder is not earning its keep. Fourth, sweeping the side-channel width $m$ should show whether shortcut reliance returns as the bottleneck opens. Our prediction is that it does, which would identify how much open-vocabulary capacity is safe.

Across all four, the concept audit reports which lexicon categories carry gender style. It also reports whether removing them trades accuracy for parity, directly testing the untested assumption that closed lexicons are population-invariant.