Hate Speech Detection in an Indian Context

Anjali Bhavan

What is Hate Speech?

• Multiple definitions by authors
   
• Widely accepted definitions?
   
• Protected categories covered under hate speech
   
• What factors are considered while defining hate speech?

Why this topic for my thesis?

• West-centric research mostly - majority of academic publishing, industrial advancements
   
• Colonialism in AI
   
• India-specific issues when it comes to hate speech detection

Issues with Western focus in ML R&D

• Most popular internet applications developed in US/Europe
  - narrow intended audience
  - lack of resources for other cultures/countries
  - similar situation in academic research
   
• Perspective matters in hate speech - who gets a say?
  - Depends on who holds power in global landscape
  - Easy to dictate terms when you're the default arbiter of everything
   
• Highly subjective nature of hate speech

India-specific issues

• Consequences of Western orientation:
  - significant population without English-medium education
  - other factors of discrimination
   
• Casteist hate speech/caste as a protected category
  - What is caste? Why does it matter?
   
• Sociopolitical landscape: rise in communal violence, hate crimes
   
• Code-mixed language/multilinguility

Other issues

• Instances of racist bias in hate speech models: recent work demonstrating bias against African-American English in models
   
• Lack of non-English data: serious setback

What I hope to do

• Study how hate speech models built with a US-centric idea of hate and racism function in a non-Anglophone context
   
• Specifically how do they work with Indian data?
   
• Evaluation/auditing: running state-of-the art hate speech detection models on Indian data and analyzing performance
   
• A thorough literature review

What I hope to do (cont.)

• Write about categories in hate speech: extreme speech, dangerous speech, fear speech etc.
   
• (Misc.) A commentary on caste in computing (particularly casteist speech), how it manifests on social media: linguistic markers etc.
   
• Some more focus on WhatsApp and its part in spreading inflammatory, hateful content and instigating communal violence in India

Some takeaways

• A conversation about hate speech is a conversation about society.
   
• It is also a conversation about the concentration of power - who gets to express their sentiments, who gets to air their hateful opinions
   
• Subjectivity: a lot of nuances to hate speech, which we can barely begin to capture