Online Hate Ratings Vary by Extremes: A Statistical Analysis

By Joni O. Salminen, Hind Almerekhi, Ahmed Mohamed Sayed Kamel, Soon-gyo Jung, Bernard Jim Jansen. 2019

55 citations recorded by the source corpus.

Analyzing 5,665 crowd ratings on 1,133 social media comments, we find that individuals tend to agree on the extremes of a hate rating scale more than in the middle when evaluating the hatefulness of online comments. The agreement is higher for less hateful comments and lowest on moderately hateful comments. The results have implications for researchers developing machine learning models for online hate processing, as the extreme classes are likely to require fewer annotations for reaching statistical stability. Our findings suggest that the models developed in this domain should consider the distributions of hate ratings rather than average hate scores.

Subfields: Artificial Intelligence, Communication, Social Psychology, Statistics and Probability, AI Ethics, Fairness & Society

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

1. Bertie Vidgen
2. Mohammad Shoeybi
3. Bryan Catanzaro
4. Vinodkumar Prabhakaran
5. Helen Z. Margetts
6. Anima Anandkumar
7. Dong Nguyen
8. Dylan Baker

In Communication

1. Dong Nguyen
2. Helen Z. Margetts
3. Dylan Baker
4. Bertie Vidgen
5. Vinodkumar Prabhakaran

In Social Psychology

1. Dong Nguyen
2. Helen Z. Margetts
3. Bertie Vidgen
4. David Jurgens
5. Vinodkumar Prabhakaran
6. Shrimai Prabhumoye

In Statistics and Probability

1. Dong Nguyen
2. Helen Z. Margetts
3. Mohammad Shoeybi
4. Anima Anandkumar
5. Patrícia G. C. Rossini
6. Bryan Catanzaro
7. Bertie Vidgen
8. David Jurgens

In AI Ethics, Fairness & Society

1. Dong Nguyen
2. Helen Z. Margetts
3. Bertie Vidgen
4. David Jurgens
5. Vinodkumar Prabhakaran
6. Rebekah Tromble
7. Shrimai Prabhumoye

Citations in the corpus

• Detecting Cross-Geographic Biases in Toxicity Modeling on Social Media (2021, background): “Moreover, prior work shows that interpretation of hate varies significantly across different cultural contexts (Salminen et al., 2018, 2019) and reveals that the performance state-of-the-art models such as BERT (Devlin et al., 2019) on various tasks degrades significantly when used on noisy text…”
• Overview of the HASOC track at FIRE 2019: Hate Speech and Offensive Content Identification in Indo-European Languages (2019, background): “in which humans agreed much on obvious Hate Speech tweets [35], there seems to be less agreement by systems.”
• Introducing CAD: the Contextual Abuse Dataset (2021, background): “Low agreement is partly because abuse is inherently ambiguous and subjective, and individuals can perceive the same content very differently (Salminen et al., 2019, 2018).”
• Few-shot Instruction Prompts for Pretrained Language Models to Detect Social Biases (2021, methodology): “The LDNOOBW dataset (Shutterstock, 2013) contains 403 banned English words and has been used in prior research (Salminen et al., 2019; Simonite, 2021).”