It didn't help that were all Swedes and the vast majority of the users were Russian/Indonesian/Nigerian/Indian. Language-specific ML could definitely have helped nowadays, but this was 2005.
Another aspect:
The content was so "inappropriate" that we felt that we coulnd't really recruit young/cheap people with the proper language skills to watch these feedback channels. It just felt.. well, wrong.
Here's a lesson: if you build a browser and target it to people who have not been exposed to the Internet before, most users will use it to access porn. Not wikipedia.org, un.org etc. Very, very hard-core porn.
(We did notice that they actually started looking for things besides porn after a couple of years, though! I guess it's something people have to go through...)
In the beginning, I mean. Quite soon we also had norwegian team members, and a few years later some polish team members, and so on. Language mis-match still was most certainly an issue though.