Health in my Language: Are health domain adapted machine translations of Cochrane Plain Language Summaries more useful to readers than having no translation in their language?

Article type
Authors
Ried J1, Hassan H1, For the HimL consortium N2
1Cochrane
2University of Edinburgh, Charles University Prague, Ludwig Maximilian University of Munich, NHS 24, Cochrane, and Lingea
Abstract
Background:
Health in my Language (HimL) was an EU-funded, three-year project. It aimed to address the need for reliable and affordable translation of public health content into different languages via fully automatic machine-translation (MT) systems, focusing on translation from English into Czech, German, Polish, and Romanian. Recent advances in MT were used to produce health adapted MT engines. Health information from Cochrane and NHS24 served as the test cases to evaluate the obtained translations.

Objectives:
To show the usefulness and acceptability of HimL MT for non-native English users of Cochrane Plain Language Summaries (PLS).

Methods:
Randomly selected PLS were translated into the HimL languages using HimL MT and published on cochrane.org. A short anonymous survey appeared when users viewed those translations. A similar survey was added to randomly selected Polish and German human translations for comparison.
The survey included three questions, which had been translated into the HimL languages:
1) The translation below was generated using machine translation. How easy is it to understand? (Rating out of 5)
2) Is this translation more useful for you than only seeing the original English text? (yes/no)
3) If you prefer to read the English version, please explain why. (free text)

Results:
HimL MT was rated with 3 out of 5 stars on average, compared to 5 stars for human translations. Czech, German and Romanian MT reached ratings of or beyond 3 stars, while Polish MT only received 2 stars on average. 40% of all respondents found HimL MT more useful than only seeing the English text: 75% of German, 40-50% of Czech and Romanian, but only 6% of Polish respondents. The most common reason given for why the MT was not more useful, was low quality of the MT. Some respondents found the English text easier to understand, which indicates that they had good English skills.

Conclusions:
A substantial number of Cochrane PLS users would benefit from domain adapted MT, where no human translations are available. Users who are less comfortable reading English are likely to benefit most. Some languages may require more research.

Patient or healthcare consumer involvement:
The anonymous survey was completed by users of the Cochrane website, which includes patients and healthcare consumers.