Page MenuHomePhabricator

Create follow-up edit quality campaign for ptwikipedia
Closed, ResolvedPublic

Description

Goal: Add new labeled data to the train/test set.

We don't need to have the same amount of data. We could try smaller sample. Maybe 10k which would mean we'd need ~2k labels.

Event Timeline

Restricted Application added a subscriber: Aklapper. · View Herald Transcript

Hi @GoEThe! We already have some edit quality models for ptwiki. See https://ores.wikimedia.org/v3/scores/ptwiki See also https://www.mediawiki.org/wiki/ORES#Edit_quality

However, these models are trained on old data from 2015. Would you like us to start a new labeling campaign to retrain the models with fresh data?

Hi. If you think the old data is still valid, no need to start a new labelling campaign.

Halfak renamed this task from Edit quality campaign for ptwikipedia to Create follow-up edit quality campaign for ptwikipedia.Mar 23 2020, 4:52 PM
Halfak claimed this task.
Halfak triaged this task as Medium priority.
Halfak updated the task description. (Show Details)

I think it would be valuable to update our dataset with new labels. This will let us check on the performance of the model trained on 2015 data and it will also let us improve the performance.

I gathered a random sample of 10k revision, but it looks like almost 41% of those edits were:

  • Saved by an anonymous editor
  • Reverted at some point by someone
  • Saved by a registered newcomer

So we'd need to get labels for 4100 edits in order to boost the the labeled data by 10k observations. When working quickly, it takes about 5s to review each edit. That means we're looking at ~6 hours of work distributed across a set of editors. If we had 6 editors labeling, they could each do an hour of work. If we had 12, each could do 30 minutes.

Does this sound reasonable? Do you think we could get enough Wikipedians together to do this work?

I think it sounds reasonable. I can recruit some Wikipedians to do this on the Village Pump and on the Telegram chat.

OK I'll get a campaign loaded to wiki labels and will ping here when it is ready. Oh! In the meantime, what do you think it should be called? I think some Portuguese translation of "Edit quality 4k sample (2020)" would be great.

"Amostragem da qualidade de 4 mil edições (2020)" seems a reasonable translation to Portuguese.

Sorry for the delay on this one. I'd been focusing on working on the articlequality model so I forgot about it. See the new campaign here: https://labels.wmflabs.org/ui/ptwiki/

We have 4000 labels. I expect this labeling work to take about 7 hours. This assumes the average edit takes ~10 seconds to review. I think that's pretty conservative and most edits will take less than 5 seconds to review. 3.5 hours of work is possible. The more people who work on labeling, the faster it goes!

Progress (100% done, -19 labels left):

Screenshot from 2020-06-05 11-00-48.png (768×1 px, 91 KB)

https://labels.wmflabs.org/stats/ptwiki/93

The negative number is possibly related to https://github.com/wikimedia/wikilabels/issues/68

Thanks! Rebuilding the model.