Selection of mental-health target words; collection of contextualized usages from everyday online discourse, clinical and scientific literature, and LLM-generated text; pilot and validate annotation protocols for semantic proximity, valence and arousal.
A validated multilingual annotation protocol and curated contextualized usage dataset.
Collect human annotations of contextualized usages and usage pairs. Compare participants with and without self-reported lived experience of mental illness. Construct separate Word Usage Graphs for semantic proximity, valence and arousal and compare clusters across groups and sources.
A human-annotated multilingual benchmark characterizing meaning variation across contexts and participant groups.
Evaluate contextualized representation models and LLMs as computational annotators. Compare predictions with human judgments at Word-in-Context (WiC), Word Sense Induction (WSI) and Graded Change Detection (GCD) levels, including affective dimensions.
Systematic evaluation of how models encode and recover denotational and connotational meaning.
Assess existing guardrails for clinical–colloquial disambiguation, over-pathologization and under-recognition of distress. Develop and evaluate targeted hybrid safeguards and guidance for input-, generation- and output-stage interventions.
Prototype guardrails and design guidelines for safer human–AI interaction.