Text Genre, Not Platform Identity, Predicts Transfer Failure in Mental Health Natural Language Processing: A Five-Axis Deployment Audit Across Five Corpora
A five-axis pre-deployment audit of mental-health text classifiers moved across platforms and corpora.
The paper in five minutes
Mental-health NLP models are almost always validated on the platform they were trained on. This work proposes a pre-deployment audit on five axes, discrimination, statistical significance, prediction equity, calibration and attribution stability, and asks what actually predicts transfer failure when those models meet new platforms and corpora.
The research question
When a mental-health text classifier moves to new platforms and corpora, what predicts where it fails?
Resources
Not public yetResults and code are withheld while this manuscript is prepared for double-blind review.
- NLP · LLMs
- Fairness
- Healthcare AI
- Deployment Shift
Citation
Manuscript. A draft is available on request. Reach me at rajveerpall04@gmail.com.