📄 Cross-Lingual Mental Health Ontologies for Indian Languages: Bridging Patient Expression and Clinical Understanding through Explainable AI and Human-in-the-Loop Validation

Summary

The paper examines the gap between how patients in India describe psychological distress in their native languages and how it is captured by the Western clinical classifications DSM-5 and ICD-11. NLP tools trained predominantly on English-language data barely recognize idioms (for example, Hindi man ka bhoj), bodily metaphors, and code-mixing with English. The authors propose the conceptual framework CL-PDE (Cross-Lingual Graphs of Patient Distress Expressions). First, a corpus of patient narratives is collected from consultations, helplines, forums, and social media and annotated with linguistic and cultural labels. Next, a heterogeneous graph is built in which expressions in different languages are linked to one another and to clinical concept nodes. Candidate links are proposed by multilingual encoders and LLMs, and experts verify them in a human-in-the-loop cycle. A separate explanation layer makes each mapping understandable to the clinician. This is a position paper: it contains no empirical results, only an evaluation plan. Its value lies in a methodological template for a culturally sensitive cross-lingual ontology for low-resource languages.

Key points

• Type of paper: a framework description without experiments. Language examples are drawn from Hindi, Kannada, and Marathi, as well as from code-mixed Hindi-English utterances. Context: 22 officially recognized (scheduled) languages with more than 1.17 billion speakers, and 1369 languages and dialects according to the 2011 census.
• The graph is built in four steps:
  1. Expression nodes are extracted using NER and phrase mining.
  2. Intra-lingual and cross-lingual edges are determined using mBERT / XLM-R embeddings and cosine similarity.
  3. An LLM proposes links to ICD-11, DSM-5, or cultural categories, adding a rationale and an uncertainty estimate.
  4. Edges are enriched with metadata: relation type, provenance, validator confidence.
• Data are planned to come from two sources. Primary: anonymized transcripts from mental health organizations in five states. Secondary: forums, support groups, and social media (subject to ethical approval).
• Explanations are generated in five ways: annotation-based, via attention visualization, rule-based, via similar examples, and contrastively. The HITL cycle includes active learning, batch validation, feedback for updating thresholds, and resolution of disagreements between annotators.
• Evaluation covers five dimensions: intrinsic graph metrics, downstream tasks, explainability, HITL efficiency, and cultural validity. The target inter-annotator agreement is κ > 0.7.