As in the opposite of a personal attack.
Signal for modeling: We can get better signal for "attack" if we also have signal for "polite". We can use this in modeling detox
Politeness as measurement: We can use this to look at temporal dynamics of teahouse responses. It would be interesting to see if polite language usage changes over time or as people integrate into the Teahouse. Similarly, we could look for polite responses to newcomers elsewhere in Wikipedia.
Is it worth the time?
We're really just looking to measure the effect of the teahouse with more nuance? Signal for polite would be good for detox.
Is politeness a goal or a means to something different? Maybe we should just model the kind of interactions that people have with newcomers that result in retention. Maybe we could use that to nudge people towards certain patterns. Could be an interface intervention.
Doesn't have to be just the Teahouse. This could inform Teahouse-like initiatives.
The harassment project has a lot of traction: board proposal, multiple team collabs, and grant proposal being worked out for actual product design.
The detox dataset contains 100k comments labelled on civility that we could use for training models.
We can't just inflict an intervention on the Teahouse. But we can provide the signal as a service and interate with them on determining how they'd like to use it. The Teahouse is interested in experimenting on things like this. Even if we do no intervention, measurements we can do will be interesting. Also might inform other modeling work.