New paper on hate speech on WhatsApp @TheWebConf !!

We studied the prevalence of hate speech in the millions of WhatsApp messages we collected during the Indian elections in 2019.

We find that most hate is not explicit but in terms of fear!

Thread

1/n

When looking for hate speech, we realised that most hate speech against Muslims is in the form of stoking fear among the majority community. See examples below.

This seems to be a well planned tactic. Over 1 in 5 msgs that mention Muslims in our data has this fear rhetoric.

2/n

We manually annotated 27k messages (dataset publicly accessible below) & tried to understand the themes in these messages & build technology 2 detect such speech

You have all the usual suspects: Love Jihad, Muslims over populating, etc

Emojis are also used very effectively

3/n

Finally, we also used a cool new way to understand user attitudes conditioned on their prior beliefs. For instance, using a survey run through Facebook ads, we find that users who engage in fear speech are much more likely to be BJP supporters, oppose CAA, etc.

4/n

This is not a problem in India or just for WhatsApp. This is one of the most common anti-semitic tropes in use. But the context we study: India (BJP, hindu nationalism, etc) + WhatsApp (zero content moderation) is extremely important.

5/n

This has real implications for content moderation. There r no easy answers & could easily lead to free speech issues if not done well. But it’s not something to be swept under the rug. We’ve seen what could happen in Myanmar. It could happen in India if not properly handled.

6/n

Like this article points, Facebook might already be aware of this and has made changes in their terms in Myanmar. https://twitter.com/gvrkiran/status/1331049314085376000

Currently nothing is done & Facebook conveniently ignores such metrics in their reports abt reducing hate speech https://about.fb.com/news/2021/02/community-standards-enforcement-report-q4-2020/

7/n

Paper published at the Web Conference 2021 #thewebconf #thewebconf2021 #www2021
Paper: https://arxiv.org/abs/2102.03870 
Code and data: https://github.com/punyajoy/Fear-Speech-analysis

Check out the (anonymized) data. It's really high quality, including spread info on WhatsApp!

8/n

Joint work with folks from IIT Kgp @punyajoysaha @_BinnyM and @Animesh43061078

Lot of effort by the PhD students Punyajoy & Binny, including an extensive hand-coding of 27k loong WhatsApp messages (triple coded) by 9 annotators, which in itself was a massive undertaking

n/n