Isaac added a comment.

  Weekly update:
  
  - I cleaned up the results notebook 
<https://public.paws.wmcloud.org/User:Isaac_(WMF)/Annotation%20Gap/eval_wikidata_quality_model.ipynb#Results>.
 The original ORES model does better on the labeled data than my initial model. 
This isn't a big surprise -- it was trained directly on them and uses many more 
features. A few takeaways:
    - I think one salient thing in comparing feature lists to take from the 
ORES model is boosting the importance of having an image if that's a common 
property for similar items.
    - The real perceived benefit of this new model will be its simplicity and 
flexibility. If we had updated test data, I think the new model would perform 
much better comparatively because it shouldn't go stale in the same way the 
ORES model would go because I'm not hard-coding lots of rules but allowing the 
model to adapt and learn from the current state of Wikidata.
    - The ordinal logistic regression approach that I used might also not be 
working well. I never really planned to keep it even though it's a good 
theoretical match for the data because I think a simpler classification or 
linear regression model w/ cut-offs would be just as reasonable. I also only 
trained it on about 200 items so I'd have plenty of test data so certainly 
plenty of room to scale that up.
    - My model includes no features regarding the actual number of statements. 
They are implicitly included in the completeness proportions (e.g., what 
proportion of expected claims exist) but I suspect humans in labeling items pay 
much more attention to the sheer quantity of statements regardless of what's 
actually expected for an item of a given type. Not sure if this is a drawback 
or not but I like that it theoretically allows for an item to be high quality 
even if it only has a few statements.
  - Other big next step will be considering how to scale up the model so it 
could potentially run on LiftWing if that's desired. It has a few semi-large 
data dependencies and that might pose a challenge.

TASK DETAIL
  https://phabricator.wikimedia.org/T321224

EMAIL PREFERENCES
  https://phabricator.wikimedia.org/settings/panel/emailpreferences/

To: Isaac
Cc: Lydia_Pintscher, diego, Miriam, Isaac, Astuthiodit_1, karapayneWMDE, 
Invadibot, Ywats0ns, maantietaja, ItamarWMDE, Akuckartz, Nandana, Abdeaitali, 
Lahi, Gq86, GoranSMilovanovic, QZanden, LawExplorer, Avner, _jensen, 
rosalieper, Scott_WUaS, Wikidata-bugs, aude, Capt_Swing, Mbch331
_______________________________________________
Wikidata-bugs mailing list -- wikidata-bugs@lists.wikimedia.org
To unsubscribe send an email to wikidata-bugs-le...@lists.wikimedia.org

Reply via email to