Source-linked AI summary
Analyzing the Targets of Hate in Online Social Media
Leandro Silva, Mainack Mondal, Denzil Correa, Fabricio Benevenuto, Ingmar Weber
TL;DR
Online hate speech is widespread, but its targets in social media remain insufficiently understood. The paper analyzes one year of Whisper and Twitter data using a validated sentence-structure method, finding recurring target categories and patterns that broaden characterization of the phenomenon. Its dataset and methodology are intended to support monitoring and detection of novel hate-speech forms.
Problem
The paper addresses the limited broad understanding of online hate speech in popular social media systems, beyond specific groups, forums, or manifestations.
Method
The authors collect one year of Whisper and Twitter data and use validated sentence-structure templates, supplemented by Hatebase words, to identify hate-speech posts and targets.
Results
Race, behavior, and physical categories are the top three hate categories on both platforms, covering 89% of Twitter hate tweets and 69% of Whisper hate whispers.
Takeaways & Limitations
The study provides a broader view of online hate, including soft targets outside categories commonly documented in offline hate crimes, and offers directions for monitoring and detection.
Takeaways & Limitations
The detection method is not designed to capture all or most online hate speech and is intended to identify main targets rather than provide complete detection.
Abstract
from arXiv · showhide
Social media systems allow Internet users a congenial platform to freely express their thoughts and opinions. Although this property represents incredible and unique communication opportunities, it also brings along important challenges. Online hate speech is an archetypal example of such challenges. Despite its magnitude and scale, there is a significant gap in understanding the nature of hate speech on social media. In this paper, we provide the first of a kind systematic large scale measurement study of the main targets of hate speech in online social media. To do that, we gather traces from two social media systems: Whisper and Twitter. We then develop and validate a methodology to identify hate speech on both these systems. Our results identify online hate speech forms and offer a broader understanding of the phenomenon, providing directions for prevention and detection approaches.
1. INTRODUCTION
Social media expands low-cost, large-scale publication while creating challenges for balancing expression with human dignity. The paper addresses limited broad understanding of online hate speech by characterizing its main targets across Whisper and Twitter.
- Near-zero-cost broadcasting lets users publish content and reach millions quickly, democratizing information access and artistic expression.
- Online hate speech illustrates the tension between freedom of expression and protecting human dignity, prompting responses from authorities and organizations.
- Existing studies often examine known hate groups, radical forums, or specific manifestations such as racism, limiting a broader view of popular social media.
- The paper gathers one year of Whisper and Twitter data, validates sentence-structure detection, and quantitatively identifies hate-speech targets.
2. DATASETS
The study assembles one-year datasets from Whisper and Twitter, combining anonymous and non-anonymous social-media posts to examine online hate speech. After filtering, the Whisper collection contains 27.55 million English, geolocated whispers, while the Twitter dataset contains 512 million English tweets.
- The researchers collected data from two social-media systems, Whisper and Twitter, over the same June 2014–June 2015 period.
- Whisper’s anonymous posting environment and popularity made it a suitable setting for studying online expression and hate speech.
- 48.97 million whispers were collected over one year, of which 27.55 million English whispers with valid location information formed the final Whisper dataset.
- Twitter data came from a 1% random sample of publicly available posts collected through the streaming API for one year.
- 1.6 billion tweets were collected, yielding a resulting dataset of 512 million English tweets with and without location information.
3. MEASURING HATE SPEECH
The paper defines hate speech operationally and detects it with sentence-structure templates designed to identify hate directed at groups. It validates the approach, categorizes prevalent targets, and acknowledges that the method prioritizes high precision over coverage.
- Defining hate speech: Hate speech is defined as an offense partly or wholly motivated by bias against an aspect of a group.The definition includes race, religion, disability, sexual orientation, ethnicity, gender, behavioral, and physical aspects.
- Sentence-structure detection: The detection method searches for first-person expressions combining intensity, a hate-related user intent, and a hate target.The basic expression is “I <intensity> <user intent> <hate target>,” with hate or synonyms used for user intent.
- Target identification: Hate targets are constrained to groups of people using templates such as “<one word> people,” with exclusion words reducing false positives.A second template accounts for hate words that do not occur with “people.”
- Target identification: Hatebase contributes 1,078 hate words across eight categories, while words with offensivity greater than 50% are used as hate-target templates.The selected list contains 116 such hate words.
- Validation and dataset construction: 20,305 tweets and 7,604 whispers containing hate speech were identified, and the “people” template accounted for 65% of Twitter matches and 99% of Whisper matches.The authors report that manual review classified sampled matched posts as hate speech when the poster expressed hate against somebody.
- Methodological scope: The methodology is not designed to capture all or most social-media hate speech, because the study aims to identify its main targets.This scope boundary frames the resulting datasets as suitable for target characterization rather than exhaustive detection.
4. TARGETS OF ONLINE HATE SPEECH
Race, behavior, and physical categories are the most prevalent hate targets on both Twitter and Whisper, though they cover different shares of hate posts. Behavior and physical targets often concern soft targets outside traditional offline hate-crime categories.
- Race, behavior, and physical are the top three hate categories on both Twitter and Whisper.
- 89% of Twitter hate tweets versus 69% of Whisper hate whispers fall within these three categories.
- Negative-connotation hate is more common than hate expressed as a response to hate.
- Behavior and physical categories commonly target people described through traits such as being fat or stupid.
5. CONCLUSION
The study provides an overview of how online hate speech manifests and identifies forms that may harm people without necessarily being crimes. Its dataset and methodology are intended to support monitoring and detection of novel hate-speech keywords.
- The measurement study offers an overview of how online hate speech manifests in contemporary society.
- It identifies new forms of online hate that are not necessarily crimes but can still be harmful to people.
- The dataset and methodology may help monitoring systems and detection algorithms identify novel hate-speech keywords.