Yusuke Matsubara and I ran a Mechanical Turk study to classify WikiLove messages a year or two ago.
Not what your colleague is asking for, but might give a couple pointers for how someone could perform that study... I think the dataset is still around somewhere too.
- about a quarter of WikiLove messages aren't "Hey, great job!". They're help requests, responses to help requests, offers of condolences, welcomes to new users, even some warnings.
- the context and content of a message matter, so if you want to track the influence of WikiLove (on # edits? length of participation?) these messages are likely to be "noise" that confounds your analysis. So, if you're interested in the effect of praise or thanks specifically, you'll want to filter out messages like these.