Development of machine translation technology for assisting health communication: A systematic review

Objectives: To (1) characterize how machine translation (MT) is being developed to overcome language barriers in health settings; and (2) based on evaluations presented in the literature, determine which MT approaches show evidence of promise and what steps need to be taken to encourage adoption of MT technologies in health settings. Materials & methods: We performed a systematic literature search covering 2006-2016 in major health, engineering, and computer science databases. After removing duplicates, two levels of screening identified 27 articles for full text review and analysis. Our review and qualitative analysis covered application setting, target users, underlying technology, whether MT was used in isolation or in combination with human editing, languages tested, evaluation methods, findings, and identified gaps. Results: Of 27 studies, a majority focused on MT systems for use in clinical settings (n = 18), and eight of these involved speech-based MT systems for facilitating patient-provider communications. Text-based MT systems (n = 19) aimed at generating a range of multilingual health materials. Almost a third of all studies (n = 8) pointed to MT's potential as a starting point before human input. Studies employed a variety of human and automatic MT evaluation methods. In comparison studies, statistical machine translation (SMT) systems were more accurate than rule-based systems when large corpora were available. For a variety of systems, performance was best for translations of simple, less technical sentences and from English to Western European languages. Only one system has been fully deployed. Conclusions: MT is currently being developed primarily through pilot studies to improve multilingual communication in health settings and to increase access to health resources for a variety of languages. However, continued concerns about accuracy limit the deployment of MT systems in these settings. The variety of piloted systems and the lack of shared evaluation criteria will likely continue to impede adoption in health settings, where excellent accuracy and a strong evidence base are critical. Greater translation accuracy and use of standard evaluation criteria would encourage deployment of MT into health settings. For now, the literature points to using MT in health communication as an initial step to be followed by human correction.