Newsletter 4/2016

End of Year 1 – Project progress

The EU-funded Health in my language project (HimL) started in February 1st, 2015, and has duration of 36 months. Its main objective is to make public health information available in a wider variety of languages, using automatic translation. After the first year of running we successfully achieved its first milestones. We deployed translation systems to translate content from both project partners - NHS 24 and Cochrane. Currently they are serving prototype websites while we assess their performance:

  • The deployed systems are integrated with the Cochrane content management system (CMS)
  • At NHS 24, they are in the process of migrating the CMS so the translation systems are not fully integrated, but we have been planning and evaluating the best way for this integration to proceed.

As we focus on producing high accuracy machine translation systems we have had to deal with several tasks in areas, such as semantics and morphology.

In the first year we have built state-of-the-art systems based on those submitted to annual shared translation tasks by members of consortium. We also have been developing techniques to improve the semantic accuracy of these baseline systems, focusing on negation, semantic roles, and the core fidelity of the rule sets in the models, while we have addressed data sparsity issues caused by our morphologically rich target languages.

Furthermore, we have developed a novel human evaluation strategy for measuring MT accuracy and initiated research on developing an automatic accuracy metric. We have also identified human accuracy and user acceptance testing as we have developed an evaluation plan closely tied to the project goals and resources of NHS 24 and Cochrane.

Once we have gathered reliable human evaluation data, we will use it to compare the performance of a number of automatic metrics, to determine which of the existing metrics are most suited to evaluating accuracy.