Machine translation (MT) seems to be the earliest goal of computer applications, not only used for private purposes but especially by governments. Despite extensive research carried out in the area, translation quality still remains unsatisfactory. Storing all the users’ utterances occurring in the web on MT services overcomes many linguistic problems, but one of the factors determining translation quality is a language pair phenomena.

Following this fact, the authors present the results of their experiments using very common and quite old tools like the Moses system, the BM25 retrieval model, and the Lucene search engine, in order to test Czech, German, and French into English medical CLEF eHealth 2013 collections. Adapted are synonyms as translation variants, compound splitting, and data selection to achieve over 50 percent higher translation precision measured with the BLEU evaluation metric.

In conclusion, there is a lack of correlation between translation quality and information retrieval results in the medical domain. This seems rather unusual, as it indicates that monolingual retrieval can be comparable to multilingual retrieval, and hence translation quality has no impact on retrieval results.

Both the techniques and the tools are described in great detail. Therefore, the paper may attract the attention of students rather than professionals working in the field. However, professionals can make some use of the experimental results, so I would recommend it both to those in academia and practitioners.

In my opinion, the paper is quite interesting and well written, but the research group should investigate other factors affecting translation quality; then the results would be more reliable and not that outstanding. The references are pretty good, but 135 items cited seems a little bit too many for an 18-page paper.