MULTILINGUAL XLM-R FOR MENTAL HEALTH SIGNAL DETECTION IN LOW-RESOURCE TAMIL LANGUAGE THROUGH CROSS-LINGUAL AND BILINGUAL LEARNING

Abstract

Mental-health NLP (Natural Language Processing) remains strongly English-centric, limiting language-inclusive digital health technologies for low-resource languages such as Tamil. This study investigates whether multilingual representation learning can transfer a specialised mental-health classification task to Tamil and whether bilingual English–Tamil training improves performance over Tamil-specific learning. To experiment this, two publicly available mental-health datasets were combined into four categories: Normal, Anxiety, Depression, and Suicidal Ideation. English samples were translated into Tamil, with 1,000 translations manually validated. Eight experimental setups compared TF-IDF-based classifiers and XLM-R models across English, Tamil, cross-lingual transfer, and bilingual training configurations. Accuracy, macro-F1, class-wise performance, confusion patterns, paired predictions, and explainability were analysed. McNemar’s test compared Tamil-specific and multilingual XLM-R predictions. Tamil-specific XLM-R achieved 73.20% [mean ± SD] accuracy and 0.7166 [mean ± SD] macro-F1, whereas Multilingual XLM-R achieved 73.67% [mean ± SD] accuracy and 0.7231, respectively. English-only cross-lingual transfer produced lower performance (64.78% accuracy; 0.6110 macro-F1), indicating that multilingual pretraining alone did not ensure effective transfer of the specialised task to Tamil. Bilingual training showed a modest numerical improvement over Tamil-specific training, including a 4.38-percentage-point increase in Suicidal Ideation-class recall, but the overall difference was not statistically significant (χ² = 2.113, p = 0.1461). Explainability analysis showed that some highly attributed tokens were linguistically ambiguous or difficult to interpret. Multilingual language capability does not necessarily ensure effective specialised mental-health NLP for low-resource Tamil. Tamil task-specific learning substantially improved cross-lingual performance, while bilingual training provided only a modest additional benefit. The findings highlight the importance of language-specific adaptation and rigorous cross-lingual evaluation for inclusive mental-health information-processing systems.