Quality of Neural Machine Translation of Public Health Information in Hong Kong (Part 1)
Timely and accurate translation of public health materials plays an important role in promoting health in bilingual or multilingual communities, such as Hong Kong, where not only Chinese and English but also ethnic minorities’ languages such as Hindi, Nepali and Urdu are spoken. The use of automatic translation tools may be useful in this regard, especially after the recent rapid development of neural machine translation systems, which have reportedly outperformed their predecessors, such as statistical machine translation systems, and have been adopted by major developers of online machine translation tools, including Google (Wu et al., 2016), Microsoft (Microsoft Translator, 2016), Systran (Crego et al., 2016), and Sogou (Wang et al., 2017).
Despite the growing popularity of neural machine translation, there have been few studies of its application to the public health domain, except a few papers on the feasibility of statistical machine translation and its incorporation into the translation procedure. This series of articles, based on a conference paper co-authored with S. C. Siu, therefore aims to study neural machine translation of public health materials in Hong Kong by selecting three online machine translation engines, namely Google Translate (n.d.), Microsoft Translator (2018a) and Sogou Translator (2018), and focusing on the evaluation of their machine translation outputs of selected press releases published by the Department of Health.
The structure of this article series is as follows: Part 1 discusses the publication of public health information, with a focus on press releases published by the Department of Health, and the global research trend in machine translation for public health. Part 2 explains the latest developments in the field of neural machine translation. Part 3 conducts automatic and manual evaluation of the three neural machine translation engines. Part 4 studies common issues in the machine outputs and suggests ways to make better use of neural machine translation in the public health domain.
Public Health Information and Research on its Machine Translation
Public health departments play a pivotal role in disseminating accurate and comprehensible health promotion messages to the public (Regidor et al., 2007), which is essential for not only educating the public on important health issues, but also ensuring the health of the nation at large (Maylahn Fleming & Birkhead, 2013).
Public health information is available in different formats including pamphlets, posters, press releases, web pages, TV announcements, video clips and even mobile applications. In Hong Kong, for example, such information is primarily made available by the Department of Health and the Hospital Authority. The former is the “Government’s health adviser and agency to execute health policies and statutory functions” (Chan, 2016) and its major service areas include health education, surveillance and epidemiology; many units have been set up within the Department, such as the Centre for Health Protection, Drug Office, Primary Care Office, and Office for Regulation of Private Healthcare Facilities. The latter is responsible for the management of public hospital and is accountable to the Hong Kong Government through the Secretary for Food and Health (Hospital Authority, 2018). In addition to government departments and statutory bodies, private medical organizations and hospitals also provide public health information (see, for example, Hong Kong Sanatorium & Hospital (n.d.)).
Given the wide variety of health promotion materials, ranging from environmental health, infection control and clinical service to primary care, the production of quality cross-lingual public health materials could be costly and time-consuming. In the light of the rapid development of machine translation technology, the global public health research community has an increasing interest in automatic translation of public health information. For example, Kirchhoff et al. (2011) investigated the feasibility of translating health promotion materials with the use of a freely available statistical machine translation website. Gathering documents from the websites of public health agencies in the United States and comparing the machine-assisted and human translations of the documents, the authors revealed the overall equivalency between machine-translated and manually translated materials. However, as the study only analyzed the translations of documents from English to Spanish, it did not conclude whether the proposed approach was applicable to other languages. To assess the feasibility of machine translation involving other language pairs, Turner et al. (2015) conducted follow-up research to examine the performance of Google Translate in translating English health promotion documents into Chinese. Experiments showed that more work was needed to improve the translation quality, given the many problems in the automatic translation results, such as errors of word sense and word order. The authors suggested that the less satisfactory results were due to the very divergent syntactic structures and linguistic differences between English and Chinese.
Stay tuned for Part 2 of this article series!
