Source-linked AI summary
Hate Lingo: A Target-based Linguistic Analysis of Hate Speech in Social Media
Mai ElSherief, Vivek Kulkarni, Dana Nguyen, William Yang Wang, Elizabeth Belding
TL;DR
Online hate-speech research faces persistent detection challenges and limited understanding of how hate targets differ. This paper analyzes Directed and Generalized hate speech using linguistic and psycholinguistic methods, finding distinct markers between them and implications for understanding and detecting hate speech.
Problem
Accurate hate-speech detection remains challenging, while research has largely lacked analysis of hate speech by target type.
Method
The authors conduct linguistic and psycholinguistic analyses of Directed and Generalized hate speech and curate datasets of 28,318 Directed and 331 Generalized hate-speech tweets.
Results
Directed hate speech is more personal, informal, angrier, and higher in clout, whereas Generalized hate speech emphasizes religious, quantity, and lethal words.
Takeaways & Limitations
Distinguishing Directed from Generalized hate speech has societal implications and can support further research on hate-speech understanding, detection, and counter-speech.
Takeaways & Limitations
The study notes sample-quality and Twitter Streaming API issues, along with the need to move beyond keyword-based methods.
Abstract
from arXiv · showhide
While social media empowers freedom of expression and individual voices, it also enables anti-social behavior, online harassment, cyberbullying, and hate speech. In this paper, we deepen our understanding of online hate speech by focusing on a largely neglected but crucial aspect of hate speech -- its target: either "directed" towards a specific person or entity, or "generalized" towards a group of people sharing a common protected characteristic. We perform the first linguistic and psycholinguistic analysis of these two forms of hate speech and reveal the presence of interesting markers that distinguish these types of hate speech. Our analysis reveals that Directed hate speech, in addition to being more personal and directed, is more informal, angrier, and often explicitly attacks the target (via name calling) with fewer analytic words and more words suggesting authority and influence. Generalized hate speech, on the other hand, is dominated by religious hate, is characterized by the use of lethal words such as murder, exterminate, and kill; and quantity words such as million and many. Altogether, our work provides a data-driven analysis of the nuances of online-hate speech that enables not only a deepened understanding of hate speech and its social implications but also its detection.
Introduction
The paper distinguishes hate speech by its target, separating attacks on individuals or entities from attacks on communities or groups. It analyzes these forms and identifies linguistic and psycholinguistic differences with implications for understanding and detecting online hate speech.
- Research gap: Prior work largely reduced hate speech detection to a hate-versus-non-hate distinction, overlooking the target of the hate.This omission misses distinctions that can influence free-speech policy and have First Amendment implications.
- Target-based typology: Directed hate speech targets an individual entity, whereas Generalized hate speech targets a particular community or group.Figure 1 illustrates the two target-based categories.
- Linguistic distinctions: Directed hate speech is more personal and invokes intentional action and words explicitly hindering the target’s action, including name calling.The paper also characterizes it as more informal, angrier, and higher in clout than Generalized hate speech.
- Linguistic distinctions: Generalized hate speech is dominated by religious and ethnic terms, quantity words such as million and many, and lethal words such as murder, killed, and exterminate.Religious hate dominates categories such as Nationality, Gender, and Sexual Orientation.
- Contributions: The study contributes the first extensive target-based analysis, lexical and semantic comparisons, and a dataset containing 28,318 Directed and 331 Generalized hate speech tweets.The dataset extends an existing public hate speech corpus.
Related Work
Related work covers abusive-language detection, automated hate-speech moderation, and research on hate-speech targets. The paper extends this literature through an open-vocabulary characterization of Directed and Generalized hate speech.
- Anti-social behavior detection: Earlier studies used machine learning to detect abusive messages, cyberbullying, personal insults, and offensive language across social-media platforms.Examples include Twitter and YouTube studies.
- Hate speech detection: Hate-speech detection systems have used surface, lexical, sentiment, linguistic, knowledge-based, metadata, and multimodal features.Automated detection is commonly paired with social-content moderation.
- Hate speech targets: Research on hate-speech targets identified groups targeted for ethnicity, behavior, physical characteristics, sexual orientation, class, and gender.Prior target studies searched for sentence structures resembling “I hate targeted group.”
- This paper’s extension: This paper differentiates Directed and Generalized hate speech without limiting data curation to one sentence structure.It uses an open-vocabulary approach to characterize both categories.
Data, Definitions and Measures
The study defines hate speech and constructs target-specific Twitter datasets through multiple collection and filtering strategies. It combines lexical resources, hashtags, public datasets, stance-sensitive classification, and human evaluation.
- Definitions: Hate speech is defined as direct and serious attacks on protected categories based on characteristics including race, ethnicity, religion, sex, gender, and disability.The paper adopts a prior typology distinguishing Directed and Generalized language.
- Dataset challenges: The authors identify dataset challenges including scarce hate tweets, time-consuming annotation, definitional disagreement, contextual ambiguity, misspellings, and abbreviations.These issues motivate multiple collection strategies and comprehensive human evaluation.
- Data collection: The keyphrase-based dataset begins with a 1% Twitter public-stream sample from January 2016 through July 2017 and Hatebase’s categorized English hate terms.The resource includes terms spanning archaic, class, disability, ethnicity, gender, nationality, religion, and sexual orientation categories.
- Filtering and measures: Perspective API toxicity and attack-on-commenter scores help filter obscure context and stance after keyword filtering.The toxicity model estimates whether a comment may make people leave a discussion, while attack-on-commenter estimates attacks on fellow commenters.
- Filtering and measures: Directed hate tweets are retained only when they mention another account and contain second-person pronouns, producing 28,318 high-precision instances.The retained tweets contain explicit Hatebase expressions directed toward target accounts.
- Data collection: The broader collection combines keyphrases, hashtags, public datasets, No Hate Speech Movement reports, and an 85,000-tweet random general-Twitter sample.The general sample is drawn from a 1% stream collected during the same 18-month window.
- Dataset evaluation: Human evaluation labeled 95.6% of sampled Generalized tweets and 97.8% of sampled Directed tweets as hate speech.The corresponding target labels were 87.5% for group-directed tweets and 94.3% for individual-directed tweets.
Analysis
The analysis uses SAGE, entity recognition, LIWC, and frame analysis to identify linguistic patterns distinguishing Directed and Generalized hate speech. Directed hate is more personal, informal, influential, and action-oriented, whereas Generalized hate emphasizes religious entities, lethal concepts, and quantity.
- Lexical Analysis: SAGE identifies distinctive lexical domains for hate-speech classes, with minimal overlap among categories.The model treats each tweet as a document and retains words appearing at least five times in the corpus.
- Lexical Analysis: Salient words for #whitepower center on white-nationalist and political themes, while #nomuslimrefugees emphasizes anti-Islam, exclusion, and terrorism-related terms.Examples include #whitepride, nazi, #kkk, #stopislam, #muslimban, and #sendthemback.
- Lexical Analysis: Directed hate contains 55.8% PERSON mentions, compared with 42.1% in Generalized hate and 46.4% in Gen-1%.This distribution is consistent with personal attacks on specific individuals.
- Lexical Analysis: Generalized hate features religious and ethnic entities, whereas Directed hate includes more common person names and explicit target references.Generalized examples include Jews, Muslims, Christians, Hindus, Hamas, Palestine, and Israel; some celebrity names occur only in Gen-1%.
- Semantic Analysis: Directed hate most strongly evokes intentional acts, statements, hindering, and obligation, with intentionally act frames at 0.05 versus 0.03 for Generalized hate.Words such as do, doing, did, retard, and retarded explicitly call out actions or attack the target.
- Semantic Analysis: Generalized hate has the highest proportions of People by religion, Killing, Color, and Quantity frames, including 0.06 for People by religion and 0.03 for Killing.These values exceed the corresponding Directed hate proportions of 0.002 and 0.006.
- Semantic Analysis: General tweets have the highest proportions of Cardinal Numbers and Calendric Units, at 0.03 and 0.031 respectively.These exceed Directed hate proportions of 0.016 and 0.01.
Discussion and Conclusion
Distinguishing Directed from Generalized hate speech has implications for law, public policy, society, and hate-speech detection. The analysis also faces sample-quality and Twitter Streaming API challenges.
- Distinguishing emotional harm to private individuals, public political figures, and general communities raises legal and policy questions about free speech.
- Generalized hate speech can reach and potentially mobilize larger numbers of people, with potentially devastating societal consequences.
- Modeling hate-speech nuances beyond hate-versus-non-hate classification is presented as critical for effectively combating online hate speech.
- The approach is constrained by sample-quality issues and other problems associated with the Twitter Streaming API.
- Keyword-based methods remain insufficient because they can miss many instances of hateful speech.