An In-depth Analysis of Implicit and Subtle Hate Speech Messages

Abstract

The research carried out so far in detecting abusive content in social media has primarily focused on overt forms of hate speech. While explicit hate speech (HS) is more easily identifiable by recognizing hateful words, messages containing linguistically subtle and implicit forms of HS (as circumlocution, metaphors and sarcasm) constitute a real challenge for automatic systems. While the sneaky and tricky nature of subtle messages might be perceived as less hurtful with respect to the same content expressed clearly, such abuse is at least as harmful as overt abuse. In this paper, we first provide an in-depth and systematic analysis of 7 standard benchmarks for HS detection, relying on a fine-grained and linguistically-grounded definition of implicit and subtle messages. Then, we experiment with state-of-the-art neural network architectures on two supervised tasks, namely implicit HS and subtle HS message classification. We show that while such models perform satisfactory on explicit messages, they fail to detect implicit and subtle content, highlighting the fact that HS detection is not a solved problem and deserves further investigation.

Unveiling the Implicit Toxicity in Large Language Models

“…Implicit toxicity has emerged as a main challenge in the field of toxicity detection owing to its nuanced nature (ElSherief et al, 2021). Rather than overtly abusive language such as swearwords, implicit toxicity is conveyed through a variety of linguistic features (such as euphemism (Magu and Luo, 2018), sarcasm (Frenda et al, 2022), circumlocution (Gao and Huang, 2017), and metaphor (Lemmens et al, 2021)) and extralinguistic knowledge (such as commonsense knowledge (Sridhar and Yang, 2022), world knowledge (Ocampo et al, 2023), and social norm (Jiang et al, 2021)).…”

“…Diverse Linguistic Features To demonstrate that LLMs can employ diverse linguistic features to express toxicity, we provide multiple qualitative examples in Appendix C. We can see that LLMs use diverse linguistic features such as circumlocution, euphemism, sarcasm, metaphor, rhetorical question (Frank, 1990), antithesis (Ruzibaeva, 2019), and visual signs (Ocampo et al, 2023). Moreover, LLMs often combine multiple features in their toxic outputs, posing a greater challenge for reasoning over compositional linguistic features.…”

K-HATERS: A Hate Speech Detection Corpus in Korean with Target-Specific Ratings

“…Most research focused on overt forms of hate speech, but explicit hate is more easily identifiable, e.g., by lexicon-based methods (Davidson et al, 2017). Recent research focused on implicitness of hate speech and proposed new datasets (Jurgens et al, 2019;ElSherief et al, 2021;Wiegand et al, 2021;Ocampo et al, 2023;Nejadgholi et al, 2022;Hartvigsen et al, 2022). A study developed a taxonomy of implicit hate and provided a labeled dataset (ElSherief et al, 2021), but its annotation scheme is not generalizable in the Korean context (e.g., White Grievance).…”

“…(2) Implicitness: The majority of previous research conducted on this subject has primarily concentrated on explicit manifestations of offensive language. A few studies investigated the implicit form of hateful content (ElSherief et al, 2021;Ocampo et al, 2023;Hartvigsen et al, 2022), most of which targeted English. (3) Biased and noisy labels: Manual annotation is a predominant approach for constructing labeled data.…”