Source-linked AI summary
Latent Hatred: A Benchmark for Understanding Implicit Hate Speech
Mai ElSherief, Caleb Ziems, David Muchlinski, Vaishnavi Anupindi, Jordyn Seybolt, Munmun De Choudhury, Diyi Yang
TL;DR
Implicit hate speech remains under-addressed because coded and indirect language is difficult to detect and spans diverse linguistic forms. The paper introduces a theoretically grounded taxonomy and benchmark corpus with fine-grained labels and implied statements, then evaluates baseline classifiers and generators. Results show useful high-level categorization and explanation baselines, while detailed categories and implied statements remain challenging.
Problem
Prior hate-speech research has focused mainly on overt abuse, leaving the diverse coded and indirect forms of implicit hate insufficiently addressed.
Method
The paper builds a social-science-grounded taxonomy and Twitter benchmark with fine-grained labels, implied statements, and baseline detection and generation systems.
Results
Neural baselines effectively categorize hate speech, but state-of-the-art models are not effective at detailed categories and explanations of implied messages.
Takeaways & Limitations
The corpus provides a benchmark for understanding implicit hate speech and supports research on online hate, radicalization, and societal bias.
Takeaways & Limitations
Releasing an implicit-hate dataset carries risks associated with potentially offensive or upsetting content.
Abstract
from arXiv · showhide
Hate speech has grown significantly on social media, causing serious consequences for victims of all demographics. Despite much attention being paid to characterize and detect discriminatory speech, most work has focused on explicit or overt hate speech, failing to address a more pervasive form based on coded or indirect language. To fill this gap, this work introduces a theoretically-justified taxonomy of implicit hate speech and a benchmark corpus with fine-grained labels for each message and its implication. We present systematic analyses of our dataset using contemporary baselines to detect and explain implicit hate speech, and we discuss key features that challenge existing models. This dataset will continue to serve as a useful benchmark for understanding this multifaceted issue.
1 Introduction
Implicit hate speech uses coded or indirect language that is harder to detect than explicit hate. The paper addresses this gap with a theoretically grounded taxonomy, annotated dataset, and baseline systems for classification and explanation.
- Implicit hate uses coded or indirect language, whereas explicit hate relies on direct wording and specific keywords.
- Because implicit hate lacks clear lexical signals, keyword systems can miss it and models may struggle without relevant training data.
- Implicit hate spans sarcasm, humor, euphemisms, circumlocution, symbolic language, stereotypes, threats, intimidation, and incitement to violence.
- The paper develops a social-science-grounded 6-class taxonomy and uses it to annotate a Twitter corpus with fine-grained labels and implied statements.
- Baseline classifiers detect implicit hate, but state-of-the-art neural models remain ineffective at detailed category classification and explanation.
2 Related Work
Prior hate-speech datasets and detection efforts largely emphasize explicit lexical signals, leaving implicit hate insufficiently labeled and potentially subject to topic bias. The paper extends earlier resources with a theoretically grounded, fine-grained implicit-hate benchmark.
- Most existing datasets were seeded with hate lexicons, racial identifiers, or explicit hateful phrases, biasing them toward overt abuse.
- Neutrally seeded datasets may contain implicit hate, but generally lack implicit-hate labels and have unknown distributions.
- Datasets focused on controversial events or specific targets may introduce topic bias and inflate performance on implicit examples.
- The paper extends the Gab Hate Corpus with a theoretically grounded taxonomy, fine-grained labels, and free-text pragmatic implications.
3 Taxonomy of Implicit Hate Speech
The paper defines implicit hate as coded or indirect disparagement and organizes it into theoretically motivated categories. The taxonomy covers 98.6% of implicit hate in a representative sample, although categories are not necessarily exhaustive or mutually exclusive.
- Implicit hate uses coded or indirect language, including sarcasm, metaphor, and circumlocution, to convey prejudicial or harmful views about protected groups or individuals.
- The taxonomy represents principal axes of implicit hate, is not necessarily mutually exclusive, and covers 98.6% of a representative sample.
- White Grievance: White Grievance frames majority groups as victims of racism and minorities as possessing perceived privilege.
- Incitement to Violence: Incitement to Violence elevates hate groups or ideologies and flaunts in-group unity or power.
- Inferiority Language: Inferiority Language portrays a group or individual as inferior, including through dehumanization, toxification, or in-group superiority.
- Irony: Irony uses sarcasm, humor, or satire to attack or demean a protected class or individual.
- Stereotypes and Misinformation: Stereotypes and Misinformation associate protected classes with negative attributes and include misinformation that reinforces such stereotypes.
- Threatening and Intimidation: Threatening and Intimidation includes implicit violations of rights and freedoms, removal of opportunities, and subtler forms of intimidation.
4 Data Collection and Annotation
The corpus was collected from prominent U.S. hate-group Twitter accounts, filtered to emphasize implicit content, and annotated at both high-level and fine-grained resolutions. The study also expands minority categories with bootstrapped and out-of-domain samples.
- Data Collection: Sampling covers eight major U.S. hate-group ideologies, including Black Separatist, White Nationalist, Neo-Nazi, Anti-Muslim, Anti-LGBT, and Anti-Immigrant clusters.
- Data Collection: The source corpus contains 4,748,226 tweets collected from selected hate-group accounts between January 1, 2015 and December 31, 2017.
- Data Collection and Filtering: Filtering used ideology-specific salient terms, HateSonar classifications, and explicit-keyword removal to reduce overtly hateful content.
- Annotation: Annotation proceeded from three-way high-level labels to expert fine-grained taxonomy labels for implicit-hate tweets.
- Annotation: Expert annotators reached Fleiss’ Kappa of 0.61 on an independent set of 150 tweets.
- Class Expansion: The study expanded minority classes using bootstrapped in-domain predictions and out-of-domain samples to address extreme class imbalance.
- Target and Implication Annotation: For 6,346 implicit-hate tweets, two annotators supplied target demographics and free-text implied statements formatted as Hearst-like patterns.
5 Implicit Hate Speech Classification
The paper evaluates binary and six-way implicit hate classification using SVM and BERT baselines, finding stronger neural performance but persistent linguistic challenges and limited gains from augmentation or external knowledge.
- The experiments address binary implicit-hate detection and six-way categorization using the proposed taxonomy.
- SVM baselines reach F1 scores up to 64.4, while fine-tuned neural models gain up to 6 additional points.
- BERT-base outperforms linear SVMs, achieving macro precision of 72.1 versus at most 61.4.
- Augmentation and Wikidata or ConceptNet integration do not significantly improve performance, suggesting that entity-aware compositional reasoning remains difficult.
- Challenges in Detecting Implicit Hate: False negatives involve coded symbols, discourse relations, entity framing, commonsense, metaphor, colloquial language, and irony, while false positives reflect identity-term bias.
- The dataset contains diverse linguistic phenomena that challenge current baselines and support its use as a benchmark for future work.
6 Explaining Implicit Hate Speech
The paper formulates explanation generation as a conditional task that produces a tweet’s target group and implied meaning. GPT-2 performs better than GPT, although lexical-overlap metrics may undervalue semantically aligned generations.
- The generation system produces both the intended target group and hidden implied meaning from a post’s content.
- Task Formulation: The model is trained conditionally with target and implied-statement tokens, then generates explanations from post tokens using greedy, beam, or top-p decoding.
- Training uses annotated target and implication text for 6,346 implicit-hate tweets, split into training, validation, and test sets without separating posts from their annotations.
- Generation Results: GPT-2 outperforms GPT for both target-group and implied-statement generation.
- Generation Results: Target-group generation scores higher than implied-statement generation, with example BLEU scores of 83.9 and 75.3, respectively.
- Generation Results: Beam search achieves the highest performance for both models, followed by top-p sampling.
- Generation Results: BLEU and ROUGE-L measure word overlap, so semantically aligned generations can receive pessimistic scores when their wording differs.
7 Conclusion
The paper introduces a theoretical taxonomy and benchmark corpus for fine-grained understanding of implicit hate speech, with baselines for detection and explanation. It identifies eight challenges and positions the corpus as a benchmark for studying implicit hate online and related applications.
- The work introduces a theoretical taxonomy and large-scale benchmark corpus with fine-grained labels for implicit hate speech and its implications.
- Neural models effectively categorize hate speech and generate more fine-grained explanations of implicit hateful messages.
- The study identifies eight challenges: coded symbols, discourse relations, entity framing, commonsense, metaphor, colloquial speech, irony, and identity-term bias.
- The corpus is presented as a benchmark for understanding implicit hate speech and as relevant to countering hate, detecting radicalization, and modeling societal biases.
Ethical Considerations
The study reports IRB approval and acknowledges risks from releasing an implicit hate dataset. It describes safeguards for annotators and weighs the dataset’s potential benefits against release risks.
- The study was approved by the researchers’ Institutional Review Board.
- Annotators received warnings about potentially upsetting content and could stop labeling if they felt overwhelmed.
- The authors acknowledge risks associated with releasing an implicit hate dataset but judge its benefits to outweigh those risks.
A Data Collection Details
Data collection used staged annotation: broad labels were defined and collected first, followed by fine-grained expert annotation and implied-statement collection. The materials included worker instructions, examples, ideology-specific salient terms, and classifier-based filtering.
- Annotation stages: The first annotation stage defined explicit hate, implicit hate, and non-hateful content using Twitter’s hateful conduct policy and examples.
- Annotation stages: The second stage used expert annotation for fine-grained implicit-hate categories because those distinctions were too subtle for MTurk workers.
- Annotation interfaces: Figure 2 shows the MTurk interface for collecting ternary explicit-hate, implicit-hate, and not-hate labels.
- Annotation interfaces: Figure 3 shows collection of each implicit-hate post’s hate target and implied statement.
- Annotation interfaces: Figure 4 presents worker instructions and examples grounded in social media communities’ rules.
- Evaluation: Table 6 reports fine-grained classification averaged across five random seeds and compares BERT-based models, linear SVMs, augmentation, and knowledge-graph integration.
- Data sampling: Ideology-specific salient terms were identified with log odds ratios for nouns, hashtags, and adjectives across six hate-group ideologies.