Goal: prototype two content tagging models. Prototype is an ambiguous term and the different models are at different stages. But the goal is each will have a fully-working language-agnostic model and plan for additional improvements / language-specific tweaks.
These will be:
| Status | Subtype | Assigned | Task | ||
|---|---|---|---|---|---|
| Open | Isaac | T293478 Content Tagging Models | |||
| Resolved | Isaac | T293480 Content Tagging Models: Prototype two |
Weekly update:
Weekly update: talked with Miriam and this work will move more slowly for the rest of the quarter while I focus on some other projects but pick back up in Q3
Weekly update: began work on expanding feature set in quality model per Meta page: https://meta.wikimedia.org/wiki/Research:Prioritization_of_Wikipedia_Articles/Language-Agnostic_Quality
Weekly update:
Weekly updates:
Weekly updates:
Hello Issac, @diego,
I have the following queries for the model V2-
-Best,
Paramita
Hey @paramita_das: you can see most of these details in the write-up and attached notebook. Pointers to your specific questions below:
How are the features (e.g., page length, references, etc.) weighted in the model?
The exact weights are in the meta page and predictQuality function in the notebook but they were derived from a groundtruth dataset from English, French, and Arabic articles. I don't think I have that notebook public anywhere at the moment but it's based on a sample of articles from that were rated for quality in the last month with a little bit of balancing so English doesn't dominate the sample.
How did you set the minimum wiki thresholds? I understand that these are determined empirically, but is it considering all the wikis? or a few selective ones?
This was based on eye-balling the data for all wikis. You can see some comments on this where the minimum threholds are set in the notebook and then later on the raw data is under a cell labeled Data to help in setting min thresholds if you want to get a sense of the practical impact of these thresholds.