By Max Sogin and Andrey Akselrod
Many of our clients have the need to download a few translated files from our Smartling API every so often. But for one client in particular that continuously translates 500 files into 10 different languages, downloading in one sitting became a bit of a time-waster.
In their case, downloading 5,000 files at once resulted in 1.5 hours of download time! This was way too long – so we decided to do a little investigating.
Potential reasons for long download times:
- Server processing time of the translated file
- Time to download the file over the network
- Connection time over SSL connection
Our test case:
- 5,000 files to download
- Average file size is 12Kb
Amount of time spent on a single file download request:
It was a bit of a surprise to us that establishing a connection for this client took the same amount of time as generating a translated a file itself. Download time wasn’t an issue and server-side processing scales nicely with Smartling as we keep a lot of spare capacity on our API download servers. We also have the ability to scale horizontally by adding more hardware. But just imagine establishing that expensive connection 5,000 times – Ouch.
There are a couple of benefits that are apparent right away when files are downloaded in parallel:
- Single file server processing stays constant – you can download many files in parallel in the same amount of time it takes to download a single file.
- Time to establish a connection is significantly reduced across multiple files – If we download 20 files at the same time, we are cutting the overall time it takes to make a connection 20 times.
- The network is utilized much more efficiently
Downloading files in parallel allowed us to reduce download time from 1.5 hours to 6 minutes.
Here’s the breakdown (ballpark):
- 1.7 min to establish connections
- 1.7 min for server-side processing
- 2.6 minutes to download at roughly 3.1Mb/s
There are a number of ways to implement parallel downloads. We used gnu parallel. Download the sample bash script at our ZenDesk and try it yourself!
To continue the conversation, click here to leave a comment or tweet Andrey at @chelya!
by Nataly Kelly
A recent article in the Journal of Consumer Research reveals an important finding – that survey results can be biased if the translations are word-for-word instead of meaning-for-meaning. “If the response category labels used in different languages are not equivalent, this could bias survey results,” explain the survey authors, who hail from Ghent University, Vlerick Business School, and Pennsylvania State University.
For example, the researchers found that in one consumer survey of French speakers, the response categories were more likely to be chosen if the translations were more common expressions – such as tout à fait d’accord (completely agree) versus extrêmement d’accord (extremely agree). Likewise, a survey carried out in Dutch showed that response rates increased when the translation used the Dutch equivalent of terms such as “completely agree” instead of less common terms like “strongly agree.”
This finding won’t come as a surprise to most language professionals. I recall as an interpreter that one of the hardest things to interpret was when doctors instructed patients, “Rate your pain on a scale of 1 to 10.” While this is not difficult to render from one language into another, culturally, it does not always make sense. Many patients had no idea how to rate their pain using this method. They had never been asked to do such a thing in their home countries, and the idea of giving a numerical score to pain seemed not only foreign, but completely strange.
One thing that designers of surveys destined for multilingual audiences can do is adapt their response choices instead of merely translating them. If they use similar scales (a five-point scale, for example) and look at the same factor (agreement or satisfaction, for example), the data can still be compared across language groups.
In the future, researchers may find themselves including a discussion of cultural and linguistic adaptation for each language or country in their methodology write-up, to prove that they are using terminology that is equally recognizable and acceptable in each market and language they are surveying. For that, they’ll certainly need help from their translation partners. It might not be enough for them to simply throw their content over the wall and wait for the translation. Going forward, they might also need a description of why certain terms were selected in lieu of others, in order to prove that the translations – and the research findings that result from them – can be trusted.