Online Hate Interpretation Varies by Country, But More by Individual: A Statistical Analysis Using Crowdsourced Ratings

By Joni O. Salminen, Fabio Veronesi, Hind Almerekhi, Soon-gyo Jung, Bernard Jim Jansen. 2018

46 citations recorded by the source corpus.

Hate is prevalent in online social media. This has resulted in a considerable amount of research in detecting and scoring it. Most computational efforts involve machine learning with crowdsourced ratings as training data. A prominent example of this is the Perspective API., a tool by Google to score toxicity of online comments. However., a major issue in the existing approaches is the lack of consideration for the subjective nature of online hate. While there is research that shows the intensity of hate varies and the hate depends on the context., there is no research that systematically investigates how hate interpretation varies by country or individual. In this exploratory research, we undertake this challenge. We sample crowd workers from 50 countries, have them score the same social media comments for toxicity and then evaluate the differences in the scores., altogether 18.,125 ratings. We find that the interpretation score differences among countries are highly significant. However., the hate interpretations vary more by the individual raters than by countries. These findings suggest that hate scoring systems should consider user-level features when scoring and automating the processing of online hate.

Subfields: Artificial Intelligence, Communication, Information Systems, 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

In Communication

In Information Systems

In Statistics and Probability

In AI Ethics, Fairness & Society

Citations in the corpus

Challenges and frontiers in abusive content detection (2019, background): “This is reflected by research into inter-annotator agreement, whereby annotators often vary considerably in what they consider to be hateful or abusive, even with training and guidance (Salminen & Almerekhi, 2018).”
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…”
Directions in abusive language training data, a systematic review: Garbage in, garbage out (2020, background): “The risk here is that authors’ perceptions may differ considerably; Salminen et al. show that online hate interpretation varies considerably across individuals [114].”
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).”
ConvAbuse: Data, Analysis, and Benchmarks for Nuanced Abuse Detection in Conversational AI (2021, background): “This is because perception of phenomena such as hate, varies both across individuals and culturally (Salminen et al., 2018).”
Online Hate Ratings Vary by Extremes: A Statistical Analysis (2019, background): “However, interpretation of hatefulness varies among individuals [19], hindering the use of approaches that assign one hate score to an online comment.”
Detecting weak and strong Islamophobic hate speech on social media (2018, background): “(Salminen et al., 2018, p. 331) Most existing research into multi-class classification has focused on distinguishing between different targets of hate rather than different strengths (Burnap & Williams, 2016; Park & Fung, 2017; Saleem, Dillon, Benesch, & Ruths, 2017; Salminen et al., 2018; Silva,…”
Directions in Abusive Language Training Data: Garbage In, Garbage Out (2020, background): “The risk here is that authors’ perceptions may differ considerably; Salminen et al. show that online hate interpretation varies considerably across individuals [93].”