[Paper Review] Facilitating Access to Multilingual COVID-19 Information via Neural Machine Translation

Andy Way, Rejwanul Haque+4 more|arXiv (Cornell University)|May 1, 2020

Natural Language Processing Techniques21 references4 citations

TL;DR

This paper presents a suite of eight neural machine translation (NMT) systems for translating between English and four major European languages—German, French, Italian, and Spanish—specifically to improve access to multilingual COVID-19 health information. Trained on freely available data using state-of-the-art NMT techniques, the systems achieve performance comparable to or better than commercial engines like Google Translate and Bing Translator, and are publicly deployed online to support timely, privacy-preserving access to critical pandemic-related content for both medical professionals and the general public.

ABSTRACT

Every day, more people are becoming infected and dying from exposure to COVID-19. Some countries in Europe like Spain, France, the UK and Italy have suffered particularly badly from the virus. Others such as Germany appear to have coped extremely well. Both health professionals and the general public are keen to receive up-to-date information on the effects of the virus, as well as treatments that have proven to be effective. In cases where language is a barrier to access of pertinent information, machine translation (MT) may help people assimilate information published in different languages. Our MT systems trained on COVID-19 data are freely available for anyone to use to help translate information published in German, French, Italian, Spanish into English, as well as the reverse direction.

Motivation & Objective

• To address language barriers in accessing timely, accurate multilingual COVID-19 health information during the pandemic.
• To develop high-quality, publicly accessible neural machine translation (NMT) systems for key European languages affected by the outbreak.
• To ensure translation quality comparable to or better than commercial systems while preserving user privacy.
• To enable rapid, scalable access to multilingual public health content for medical professionals and the general public.

Proposed method

• Training end-to-end NMT models using only freely available, multilingual COVID-19 text data from diverse sources such as news outlets and health authorities.
• Employing a modular preprocessing pipeline including tokenization, sentence splitting, character normalization, spell-checking, and handling of special elements like URLs and tags.
• Implementing a distributed architecture with dedicated GPU servers for each of the eight language pairs to ensure low-latency, responsive online translation services.
• Using a webserver to manage user requests, distribute tasks to appropriate MT servers, and return translated results with proper post-processing.
• Applying advanced NMT techniques such as subword segmentation, truecasing, and compound word rejoining (especially for German) to improve fluency and accuracy.
• Validating system performance through automatic metrics and human evaluation against Google Translate and Bing Translator on real-world pandemic-related sentences.

Experimental results

Research questions

• RQ1 Can high-quality, low-cost NMT systems be rapidly developed to translate multilingual COVID-19 content into and from English, French, German, Italian, and Spanish?
• RQ2 How does the performance of these custom NMT systems compare to leading commercial MT platforms like Google Translate and Bing Translator in terms of fluency and adequacy?
• RQ3 To what extent can a modular, multilingual preprocessing pipeline improve translation quality for diverse language pairs in a crisis context?
• RQ4 Can publicly deployed, privacy-preserving MT systems effectively support non-expert users in accessing critical health information across language barriers?
• RQ5 What are the key strengths and weaknesses of custom NMT systems when applied to sensitive, rapidly evolving public health content?

Key findings

• The custom NMT systems achieved automatic evaluation scores comparable to or better than Google Translate and Bing Translator on the same test sets.
• Human evaluation revealed that the systems produced translations that were often more accurate and natural than those from commercial platforms, particularly in handling medical terminology and context.
• The system demonstrated strong performance in translating complex public health messages, including symptom identification, prevention, and treatment guidelines.
• The modular preprocessing pipeline significantly improved translation quality by correctly handling punctuation, special characters, URLs, and compound words—especially in German.
• The deployment of the systems via a distributed, privacy-preserving web architecture enabled low-latency, real-time access for users worldwide without data leakage risks.
• The systems successfully translated high-impact, time-sensitive content such as a German news headline from Die Welt (20 March 2020), confirming their practical utility in real-world crisis scenarios.