Hate Lingo: A Target-Based Linguistic Analysis of Hate Speech in Social Media

Mai ElSherief, Vivek Kulkarni, Dana Nguyen, William Yang Wang, Elizabeth M. Belding

TL;DR

This paper introduces a novel linguistic and psycholinguistic analysis of hate speech, differentiating between "Directed" (individual-targeted) and "Generalized" (group-targeted) forms. By analyzing these distinct types, the authors uncover unique linguistic markers and psychological patterns, paving the way for more nuanced hate speech detection and understanding.

What Problem Does This Paper Address?

Social media platforms, while fostering expression, also host a significant amount of harmful content, including hate speech. Existing research often treats hate speech monolithically, failing to capture crucial nuances that impact its social implications and detection. This paper addresses this gap by focusing on the target of hate speech, distinguishing between attacks on specific individuals ("Directed hate") and those against entire groups ("Generalized hate"). This distinction is vital because it can influence legal interpretations (e.g., free speech rights) and requires different approaches for effective mitigation.

Methods

The authors employ a multi-faceted approach:

  • Data Curation: They build upon existing hate speech datasets and augment them with new collections derived from Twitter using keyword-based and hashtag-based methods. Crucially, they use the Perspective API (a Jigsaw/Google tool) to filter for toxicity and attack on commenter scores, ensuring higher quality data.
  • Linguistic Analysis:
    • SAGE (Sparse Additive Generative Models of Text): Used to identify salient keywords and topic models for each hate speech category, revealing distinct vocabulary for Directed vs. Generalized hate.
  • T-NER (Named Entity Recognition): Applied to identify entities (persons, organizations, etc.) mentioned in tweets, highlighting differences in the types of entities targeted.
  • SEMAFOR: Utilized for frame-semantic analysis to understand the underlying concepts and situations evoked by words in hate speech.
  • Psycholinguistic Analysis:
    • LIWC2015 (Linguistic Inquiry and Word Count): A text analysis tool used to measure psychological dimensions like clout, analytical thinking, authenticity, emotional tone, informality, social components, and temporal focus.

Datasets

The paper introduces a curated dataset of 28,318 Directed hate speech tweets and 331 Generalized hate speech tweets, contributing to the existing public hate speech corpus. This dataset was built using:

  • Keyphrase-based dataset: Derived from Twitter's public stream, filtered using Hatebase keywords and the Perspective API.
  • Hashtag-based dataset: Collected from popular hate speech hashtags on Twitter.
  • Public datasets: Incorporating data from Waseem and Hovy (2016) and Davidson et al. (2017).
  • General dataset (Gen-1%): A random sample of 85,000 English tweets from Twitter's public stream for comparison.

Key Insights and Ideas

  • Distinct Linguistic Fingerprints: Directed and Generalized hate speech use remarkably different vocabularies. Directed hate employs more informal, angry language, often involving name-calling and words suggesting authority or hindering actions. Generalized hate leans towards religious and ethnic slurs, lethal terms (murder, kill), and quantity words.
  • Targeted Entities Differ: Directed hate speech disproportionately targets "persons," reflecting its individual focus. Generalized hate speech, conversely, often targets "other" entities, frequently religious groups (Islam, Jews, Christians), aligning with its group-based nature.
  • Psychological Underpinnings: Directed hate speech is characterized by higher "clout" and lower "analytical thinking," suggesting a more impulsive and dominant tone. Generalized hate speech, while less "clouty," shows a stronger "Us vs. Them" mentality, emphasizing third-person plural pronouns.
  • Emotional Tone and Intent: Directed hate speech is significantly angrier than Generalized hate speech, which is in turn angrier than general tweets. Directed hate also evokes frames related to intentional acts and hindering, while Generalized hate focuses on killing, religion, and quantity.

Key Results

  • Directed Hate: Exhibits higher clout (70.7 vs. 48.5 for Generalized), lower analytical thinking (43.9 vs. 68.9 for Generalized), and higher anger scores (7.6 vs. 3.6 for Generalized). It also shows a higher proportion of "person" entities (55.8% vs. 42.1% for Generalized).
  • Generalized Hate: Dominated by religious themes and terms like "kill," "murder," "exterminate," and quantity words like "million" and "many." It shows a higher proportion of "People_by_religion" frames.
  • General Tweets (Gen-1%): Exhibit higher authenticity and emotional tone, and are more focused on cardinal numbers and calendric units.

Limitations and Future Work

  • Platform Specificity: The study primarily focuses on Twitter, and findings may not generalize to other social media platforms with different affordances and user bases.
  • Keyword Limitations: While efforts were made to mitigate keyword limitations (e.g., misspellings, context), some hate speech might still be missed.
  • Future Work: The authors suggest exploring other platforms, developing more robust detection methods that account for these nuances, and investigating the societal implications of these distinctions further, particularly in relation to policy and counter-speech strategies. They also highlight the need to move beyond simple keyword-based detection.

Summary updated

Abstract

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.
BibTeX
@inproceedings{elsherief2018hate,
  author = {Mai ElSherief and Vivek Kulkarni and Dana Nguyen and William Yang Wang and Elizabeth M. Belding},
  title = {{Hate Lingo: A Target-Based Linguistic Analysis of Hate Speech in Social Media}},
  booktitle = {ICWSM},
  year = {2018},
  url = {https://openreview.net/forum?id=IGow92I6uSm},
}
Figure