[Paper Review] Are Chess Discussions Racist? An Adversarial Hate Speech Data Set

Rupak Sarkar, Ashiqur R. KhudaBukhsh|arXiv (Cornell University)|Nov 20, 2020

Hate Speech and Cyberbullying Detection9 references4 citations

TL;DR

This paper investigates whether off-the-shelf hate-speech classifiers misclassify benign chess discussions as hate speech due to racial terminology in the game's lexicon. Using a dataset of 681,995 YouTube comments from chess channels, the authors identify 1,000 human-verified non-hate-speech comments falsely flagged by classifiers—highlighting a critical failure in domain robustness and revealing systemic color polysemy in NLP models.

ABSTRACT

On June 28, 2020, while presenting a chess podcast on Grandmaster Hikaru Nakamura, Antonio Radić's YouTube handle got blocked because it contained "harmful and dangerous" content. YouTube did not give further specific reason, and the channel got reinstated within 24 hours. However, Radić speculated that given the current political situation, a referral to "black against white", albeit in the context of chess, earned him this temporary ban. In this paper, via a substantial corpus of 681,995 comments, on 8,818 YouTube videos hosted by five highly popular chess-focused YouTube channels, we ask the following research question: \emph{how robust are off-the-shelf hate-speech classifiers to out-of-domain adversarial examples?} We release a data set of 1,000 annotated comments where existing hate speech classifiers misclassified benign chess discussions as hate speech. We conclude with an intriguing analogy result on racial bias with our findings pointing out to the broader challenge of color polysemy.

Motivation & Objective

• To investigate whether hate-speech classifiers misclassify benign chess-related discussions as hate speech due to racial terminology.
• To examine the robustness of off-the-shelf hate-speech classifiers when applied to out-of-domain content, specifically chess commentary.
• To analyze the role of lexical ambiguity—particularly 'white' and 'black'—in triggering false positives in hate-speech detection.
• To explore broader implications of word embedding biases through analogy tests in different textual domains.
• To release a curated dataset of 1,000 misclassified chess comments for future research on adversarial examples and bias in NLP.

Proposed method

• Collected 681,995 comments from 8,818 YouTube videos across five popular chess channels using the YouTube API.
• Applied two pre-trained hate-speech classifiers: one fine-tuned on Twitter data ($\mathcal{M}_{\mathit{twitter}}$) and another on a white supremacist forum dataset ($\mathcal{M}_{\mathit{stormfront}}$).
• Manually annotated 1,000 randomly sampled comments flagged as hate speech by at least one classifier to determine true labels.
• Conducted analogy tests (e.g., black : slave :: white : ?) on word embeddings from chess, Fox News, and CNN comment datasets to assess semantic bias.
• Quantified false positive rates and evaluated classifier performance across domains using human-annotated gold standards.
• Analyzed linguistic triggers such as 'attack', 'threat', 'capture', and color terms to identify patterns in misclassification.

Experimental results

Research questions

• RQ1 To what extent do off-the-shelf hate-speech classifiers misclassify benign chess discussions as hate speech?
• RQ2 How does the performance of hate-speech classifiers vary across domains, particularly when applied to chess-related content?
• RQ3 What linguistic features—especially color terms and combat metaphors—trigger false positives in hate-speech detection?
• RQ4 How do word embeddings from different domains (e.g., chess vs. news) reflect or distort racial stereotypes in analogical reasoning?
• RQ5 To what extent does the context of 'white' and 'black' in chess terminology contribute to racial bias in NLP models?

Key findings

• 82.4% of comments flagged as hate speech by classifiers were actually benign, indicating a high false positive rate of 82.4% on chess discussions.
• The BERT-based classifier trained on a white supremacist forum ($\mathcal{M}_{\mathit{stormfront}}$) had a slightly lower false positive rate than the Twitter-trained model ($\mathcal{M}_{\mathit{twitter}}$), suggesting better domain alignment for racial hate speech.
• Despite being trained on extreme hate content, $\mathcal{M}_{\mathit{stormfront}}$ still misclassified 82.4% of chess comments as hate speech, demonstrating poor out-of-domain generalization.
• The analogy test 'black : slave :: white : ?' yielded 'slave' in the chess dataset, while Fox News and CNN datasets predicted 'slavemaster', revealing that chess contexts normalize racial power dynamics in embeddings.
• Terms like 'white', 'black', 'attack', 'threat', and 'capture' were primary triggers for false positives, indicating lexical ambiguity as a core cause of misclassification.
• The study demonstrates that even in a contextually neutral domain like chess, where 'white' and 'black' are purely symbolic, NLP models still associate them with racial power dynamics due to polysemy and biased pretraining data.