From Lexicon to Prediction

Suicide prevention requires attention not only to clinically recognized crises but also to communicative signals that may precede or accompany them. Language offers an observable behavioural trace through which distress can be studied outside formal clinical encounters, although linguistic evidence must not be treated as a diagnosis. The Psycho-Forensic Linguistic Surveillance Initiative (PFLSI) has progressed from culturally situated Pakistani suicide notes to corpus-scale English analysis, theoretical integration, and a proposed multilingual Urdu-Punjabi corpus. Series Paper 5 extends that programme by specifying a protocol for transformer-based modelling and multilingual validation of suicide-related discourse in English, Urdu, Punjabi, Roman Urdu, and Urdu-English code-switched communication. Because the proposed Multilingual Suicide Discourse Corpus (M-SDC) has not yet been completed, this article does not report empirical model results. Instead, it prespecifies the classification task, corpus-matching principles, annotation procedures, baseline and transformer models, culturally informed feature streams, validation regimes, evaluation metrics, ablation strategy, explainability procedures, and ethical boundaries to be implemented after corpus completion. The protocol compares corpus-linguistic and classical machine-learning baselines with multilingual BERT and XLM-RoBERTa, while testing whether structured linguistic and culturally situated features add information beyond contextual representations. Evaluation prioritizes discrimination, calibration, source-held-out generalization, cross-linguistic and cross-script transfer, subgroup stability, and interpretability. The central argument is that multilingual suicide-discourse modelling should be evaluated not by predictive performance alone but by linguistic validity, cultural calibration, source generalization, transparent error analysis, and strict separation between computational risk indication and clinical judgement.