Anticipating Hazards in Machine Translations of Public Health Resources via Advanced Text Classification Pipelines

Abstract

Public health educational resources developed by health institutions aim for high accessibility of information. The translation of these resources, which provide the public with a basic understanding of health risks and diseases, is commonly conducted by professional translators to cater to diverse linguistic and cultural backgrounds. In recent years, the global advancement of information technology has broadened the use of Machine Translation (MT) in online health education and promotion. MT tools such as Google Translate, DeepL, and ChatGPT have significantly improved performance, yet they face challenges posed by the language complexity, content complexity, and formality of professional medical resources.

In this study, we leverage Natural Language Processing (NLP) and Machine Learning (ML) tools to harness the power of text classification, a vital task that assigns text to one or more predefined categories. We aim to develop machine learning classifiers within our newly proposed Multi-Dimensional Text Classification Pipeline (MD-TCP) framework. As a risk-prevention mechanism, MD-TCP assists medical professionals with limited knowledge of the patient's language and helps patients who wish to self-navigate. Our model predicts the likelihood of clinical mistakes or incomprehensible machine translation outputs based on the features of English source input to the machine translation systems.

MD-TCP is a new, comprehensive pipeline for data mining and feature extraction that we developed to achieve this goal. The pipeline has demonstrated significant improvements in both of our datasets. Regarding Accuracy, AUC, Sensitivity, Precision, and Specificity, our method improved by 24% - 33% compared to baseline methods. This underscores the potential of machine learning, mainly when implemented through MD-TCP, in predicting translation errors, thereby ensuring more accurate and understandable translations for health resources across diverse populations.