QZ AI Studio Demo: Automatically Labeling Hate Incident Tips
Many news organizations receive tips from their readers, alerting journalists to incidents or trends. But newsrooms can be opaque and tips can sometimes end up in a giant inbox... and not read by the right reporter.
We hypothesize that machine-learning can help, by learning from examples of categorized tips.[Where else does AI fit into reporting and newsgathering? Read our first post, How you're feeling when machine learning might help..]
As an experiment, we used tips from ProPublica's Documenting Hate project. Tips submitted to that project included a free-text description of a hate incident and, optionally, a closed-class categorization of why the victim of the hate incident was targeted, like race or religion.
We trained a model to predict from that free-text description where it's left blank.
This model probably won't be directly useful to other journalists, but we hope to make the methodology clearer and easier to use -- there's nothing about it that's specific to hate crimes. It should hopefully work just as well categorizing general tips into beats, so the sports tips can be sent to the sports reporter, education tips to the schools reporter, etc.
We can also ask the model what most influenced its decision to classify the tip as a particular type of hate incident or not.
So why that prediction?
To figure out what contributes to the prediction, we remove each bigram (2-word phrase) one by one and run the modified tip through the classifier. The bigrams that, when removed, cause the largest change in the model's guess are the ones that have the biggest effect. This is kind of a hack: some models types might take into account additional context than just trigrams or do so in different ways. But this gives us a sense of what the model is basing its decision on.
You'll find some times when the model makes the right call... and other times where it has mistakenly learned that an phrase that ought to be irrelevant has a big effect. (Can you think of why that might happen?)