Multilingual Surveillance Without Blind Spots: MT-Aware Bias Modeling for Public Health Signals

Herrera-Espejel & Rach’s scoping review (2023) makes it clear MT is underused for recruiting and two-way engagement, and almost no work evaluates legal/ethical or reception effects. We extend this by treating MT as a measurable, correctable source of bias in surveillance pipelines. Concretely, we: (1) attach uncertainty scores to MT outputs (using round-trip consistency, QE models, and human post-edit samples), (2) propagate these scores through topic/sentiment/intent classifiers to produce probabilistic labels, and (3) apply outcome-misclassification bias correction akin to Hubbard et al. (2020) to adjust epidemiologic associations and time-series estimates derived from these labels. This contrasts with the common practice of naive MT followed by standard NLP, which can systematically over- or under-estimate signals in underrepresented languages. The novelty is combining MT quality estimation with bias-aware inference so that, for example, a surge in Portuguese posts about rash symptoms contributes appropriately to a Zika signal with uncertainty-aware weighting. This could shrink geographic and linguistic blind spots and make cross-lingual surveillance ethically and statistically defensible.

References:

1. The Use of Machine Translation for Outreach and Health Communication in Epidemiology and Public Health: Scoping Review. P. S. Herrera-Espejel, Stefan Rach (2023). JMIR Public Health and Surveillance.
2. Reducing Bias Due to Outcome Misclassification for Epidemiologic Studies Using EHR-derived Probabilistic Phenotypes. R. Hubbard, Jiayi Tong, R. Duan, Yong Chen (2020). Epidemiology.