Conversation

Sorry that the previous Y axis scale shifted between the comparisons. Here they are side by side with the same scale.

Dec 6, 2018

This is showing how the toxicity level doubles on #gab when comparing the average of all posts with the average of posts only containing the words "jew, jews and jewish." #dataviz

Replying to

Note that, at least in previous version of API, simple identity markers scored higher on toxicity, making for a biased distribution of false positives. Thus, it's possible that simple statement of Jewish identity or pro-Jewish posts will be misclassified engadget.com/2017/09/01/goo

Replying to

That's a good point. I know the model is sort of a "magic black box" to all of us right now -- but one thing that will help determine and address these biases is to do a parallel comparison with other social media platforms. I am going to score Reddit for the month of November ..

Replying to

and

And see what the baseline toxicity score is for Reddit. I will also run the "jew,jews,jewish" query to see the results for Reddit. If the model is flawed in a fundamental way, I would expect to see a likewise double of the toxicity percentage for Reddit ... But if Reddit gets ..

6:37 PM · Dec 6, 2018Twitter Web Client

Replying to

and

.. a baseline toxicity score of say 20% and then the jew toxicity avg is 24%, then that gives us more information and that the bias (if it exists) for those Jewish terms isn't so much such that it makes the Gab comparison unusable. There is a lot of data to score, but this will..

Replying to

and

get us to a position of being able to better understand the Perspective API and any flaws that are obvious, etc.