Outside of the code base, there are many informal descriptions of code elements, like on StackOverflow, on mailing lists or in issue trackers. An example:
The process to get them is as follows:
To evaluate the approach Peter compared 100 manually created summaries with the auto generated summaries and this resulted in a precision of about 50%.
Unfortunately I was not able to find a preprint, and I could not follow all details of the evaluation.
Update: I got tweeted a preprint: http://latifaguerrouj.ca/pubs/ACM-ICSE-2015.pdf