Source-linked AI summary
Thou shalt not hate: Countering Online Hate Speech
Binny Mathew, Punyajoy Saha, Hardik Tharad, Subham Rajgaria, Prajwal Singhania, Suman Kalyan Maity, Pawan Goyal, Animesh Mukherje
TL;DR
Online hate moderation faces a tension between reducing harmful content and preserving free speech, while empirical counterspeech data has been lacking. The paper constructs and releases a manually annotated YouTube dataset, analyzes counterspeech patterns, and develops detection models. Counterspeech receives more likes than non-counterspeech, and binary detection reaches an F1-score of 0.71.
Problem
Online hate speech is growing, while blocking or suspending content can be ineffective and may threaten free speech; empirical counterspeech data is lacking.
Source: Introduction · p. 1, Current protocols to combat hate… · p. 1, Can countering hate speech be an… · p. 1Method
The paper defines counterspeech, releases a manually annotated YouTube dataset covering three target communities, analyzes its linguistic and engagement patterns, and trains classification models.
Source: Working definition of counterspeech · p. 2, Dataset annotation · p. 3Results
Counterspeech comments receive more likes than non-counterspeech comments, and counterspeech-versus-non-counterspeech detection achieves an F1-score of 0.71.
Source: Likes and comments · pp. 5–6, Conclusion and future works · p. 10Takeaways & Limitations
Counterspeech strategies and language differ across communities, and the released dataset supports empirical measurement and automated detection of counterspeech.
Source: Conclusion and future works · p. 10Future work is needed to develop detection models for other social media sites and study the effectiveness of counterspeech across communities.
Source: Conclusion and future works · p. 10
Abstract
from arXiv · showhide
Hate content in social media is ever-increasing. While Facebook, Twitter, Google have attempted to take several steps to tackle the hateful content, they have mostly been unsuccessful. Counterspeech is seen as an effective way of tackling the online hate without any harm to the freedom of speech. Thus, an alternative strategy for these platforms could be to promote counterspeech as a defense against hate content. However, in order to have a successful promotion of such counterspeech, one has to have a deep understanding of its dynamics in the online world. Lack of carefully curated data largely inhibits such understanding. In this paper, we create and release the first ever dataset for counterspeech using comments from YouTube. The data contains 13,924 manually annotated comments where the labels indicate whether a comment is a counterspeech or not. This data allows us to perform a rigorous measurement study characterizing the linguistic structure of counterspeech for the first time. This analysis results in various interesting insights such as: the counterspeech comments receive much more likes as compared to the non-counterspeech comments, for certain communities majority of the non-counterspeech comments tend to be hate speech, the different types of counterspeech are not all equally effective and the language choice of users posting counterspeech is largely different from those posting non-counterspeech as revealed by a detailed psycholinguistic analysis. Finally, we build a set of machine learning models that are able to automatically detect counterspeech in YouTube videos with an F1-score of 0.71. We also build multilabel models that can detect different types of counterspeech in a comment with an F1-score of 0.60.
Introduction
Online hate speech is difficult to moderate without risking free speech, motivating counterspeech as an alternative. This paper introduces the first counterspeech dataset and studies its patterns and automatic detection.
- Blocking or suspending hateful messages and accounts has been ineffective and raises concerns about restricting free speech.Source: Current protocols to combat hate… · p. 1
- Counterspeech is proposed as an alternative to blocking because it adds opposing speech rather than removing content.Source: Can countering hate speech be an… · p. 1, Current protocols to combat hate… · p. 1
- The paper defines counterspeech as a direct response to hateful or harmful speech and annotates comments targeting Jews, Blacks, and LGBT communities.Source: Working definition of counterspeech · p. 2
- The released dataset contains 6,898 counterspeech comments and 7,026 non-counterspeech comments, with counterspeech further labeled by type.Source: Our contributions and observations · p. 2
- Counterspeech comments receive more likes, differ linguistically from non-counterspeech comments, and vary in prevalence across target communities.Source: Our contributions and observations · p. 2
- XGBoost achieves an F1-score of 0.71 for counterspeech versus non-counterspeech classification, while cross-community classification reaches 0.62 - 0.65.Source: Our contributions and observations · p. 2
Related work
Prior counterspeech research spans social platforms and qualitative or simulated analyses, but this paper addresses the lack of empirical, scalable detection work.
- Earlier work examines counterspeech on Twitter, YouTube, and Facebook, including discourse change, prejudice reduction, and factors associated with success.Source: Related work · pp. 2–3
- Existing counterspeech literature is described as limited and largely qualitative or anecdotal rather than empirical.Source: Related work · p. 3
- The paper responds to the absence of automatic counterspeech detection by applying learning algorithms to the task.Source: Related work · p. 3
- Beyond binary classification, the work introduces multi-label classification for different counterspeech types.Source: Related work · p. 3
Dataset
The dataset is built from YouTube comments on hateful videos targeting three communities and annotated for counterspeech and strategy types. Its design separates direct responses from replies and distinguishes positive from hostile tone.
Collection
- The researchers scrape comments from manually selected hateful YouTube videos targeting Jews, Blacks, and LGBT communities.Source: Data collection from YouTube · p. 3, Dataset · p. 3
Dataset annotation
- Comments are annotated as counterspeech when they directly oppose hatred expressed in the video; replies are excluded because discussions often become noisy and off-topic.Source: Dataset annotation · p. 3
- The annotation yields 6,898 counterspeech comments and 7,026 non-counterspeech comments, with 90.23% agreement and Cohen’s κ of 0.804.Source: Dataset annotation · p. 3
- Counterspeech comments receive a second-level annotation for strategy types, with loose κ of 0.868 and strict κ of 0.743.Source: Dataset annotation · p. 3
- The dataset records 9,119 strategy assignments across 6,898 counterspeech comments because comments using multiple strategies are counted in each applicable type.Source: Table 1 · p. 3
Types of counterspeech
- The tone category is split into Positive tone and Hostile language to distinguish potentially de-escalatory and hostile forms.Source: Types of counterspeech · p. 4
- A Figure 1 example combines affiliation and empathy in response to claims that homosexuality is unnatural, detrimental, and destructive.Source: Figure 1 · p. 4
Detailed analysis
The analysis characterizes counterspeech strategies, audience engagement, community-specific language, and psycholinguistic differences between counterspeech and non-counterspeech comments.
Counterspeech categories
- 71.24% of counterspeech comments belong to exactly one category, indicating that most counterspeakers rely on a single strategy.Different communities attract different types of counterspeech.Source: Detailed analysis · p. 5
- 39.74% of counterspeech contains hostile language, the largest category across the analyzed classes.Source: Detailed analysis · p. 5
Community differences
- Different communities favor different counterspeech strategies: Jews more often receive positive tone, while LGBT targets more often receive humor and hypocrisy or contradiction pointing.For African-American targets, warning of consequences and denouncing hateful or dangerous speech receive higher average likes than other categories.Source: Detailed analysis · p. 5, Likes and comments · p. 6
Likes and replies
- 3.0 average likes for counterspeech comments versus 1.73 for non-counterspeech comments, while replies average 1.94 versus 1.50 and are not significantly different.The likes difference is reported as significant (p ∼0.0), whereas the replies difference is not (p > 0.1).Source: Likes and comments · p. 5
- 72.07% of counterspeech comments are replies, compared with 45.17% reply density for the full dataset.The paper relates this higher reply density to the controversial nature of the content.Source: Likes and comments · pp. 5–6
Psycholinguistic differences
- LIWC analysis finds affective processes stronger in counterspeech, while spoken language and personal concern are more pronounced in non-counterspeech; biological-process language is more dominant among counterspeakers.The analysis compares statistically significant differences across LIWC linguistic and cognitive categories.Source: Psycholinguistic analysis · p. 7
Classification model
The paper develops classifiers for binary counterspeech detection, multi-label prediction of counterspeech types, and cross-community generalization. Sentence vectors generally outperform BoWV and TF-IDF, while XGBoost with combined features provides the strongest reported binary model.
Classification tasks
- Three tasks predict counterspeech versus non-counterspeech, identify all counterspeech types, and classify comments across communities.The cross-community task trains on two communities and tests on the third.Source: Classification model · p. 7, Cross-community classification · p. 9
Preprocessing and features
- The pipeline preprocesses comments by removing URLs, numerals, stopwords, and punctuation, then lowercases and tokenizes them.Features include TF-IDF, bag-of-words vectors, and sentence vectors.Source: Classification model · p. 8
Evaluation
- Stratified 10-fold cross-validation with held-out validation tuning evaluates the classifiers using accuracy, weighted precision, recall, and F1-score.For multi-label classification, the paper uses multi-label stratified 10-fold cross-validation and additionally reports Hamming loss.Source: Counterspeech classification · p. 8, Counterspeech type classification · pp. 8–9
Binary classification
- 71.6% accuracy is achieved by the best binary model, XGBoost with SV+TF-IDF+BoWV features.Sentence vectors perform better than BoWV and TF-IDF in most cases; MLP and CatBoost are comparable.Source: Counterspeech classification · p. 8, Table 3 · p. 8
Multi-label classification
- The multi-label task predicts the counterspeech types present in an input counter comment, using GeneralB as a frequency-based baseline.GeneralB predicts only the most frequent label, “Hostile language,” as relevant for this dataset.Source: Counterspeech type classification · p. 8, Table 4 · p. 9
- Label-wise performance is good for “Hostile” and “Affiliation” but poor for “Warning of offline/online consequences” and “Pointing out Hypocrisy/contradiction.”The reported label-wise results use XGBoost with SV+TF-IDF+BoWV.Source: Counterspeech type classification · p. 9, Table 5 · p. 9
Cross-community classification
- Cross-community models train on two communities and test on the third, producing comparable results across each held-out community.This setting addresses cases where in-community training instances are scarce.Source: Cross-community classification · p. 9, Table 6 · p. 10
Discussion
The discussion shows that non-counterspeech often consists of agreement with hateful video content or hate directed at the target community. Community responses also vary by counterspeech strategy, and classifiers may support larger-scale effectiveness studies and automated responses.
Non-counterspeech
- Non-counterspeech mainly comprises agreement with the video’s content or hate speech toward the target community.The dominant forms vary across Jews, Blacks, and LGBT-targeted communities.Source: Discussion · p. 10
Community acceptance
- Counterspeech strategies receive different levels of community acceptance, with “Affiliation” and “Positive Tone” receiving little acceptance in the Blacks community.Replies often stated that positive tone and affiliation would not change hate speakers’ stances.Source: Discussion · p. 10
- “Hostile language” is prevalent but is opposed by target-community users in many replies.Users requested that counterspeakers refrain from using hateful language.Source: Discussion · p. 10
Applications
- Counterspeech classifiers could support larger-scale studies of strategy effectiveness and datasets for systems that automatically counter online hate.The paper also describes profiling hate speakers to provide counterspeech based on their profiles as a possible application.Source: Discussion · p. 10
Conclusion and future works
The paper releases a YouTube counterspeech dataset and reports differences in engagement, language, and community preferences. Future work extends detection across platforms and examines strategy effectiveness and influence on hate users’ mindsets.
Conclusion
- The released dataset contains counterspeech and non-counterspeech comments, supporting measurement of engagement, psycholinguistic differences, and community-specific preferences.The conclusion states that counterspeech comments receive more likes than non-counterspeech comments and that the resources are public.Source: Conclusion and future works · p. 10
Future work
- Future work includes automatic counterspeech detection on Facebook and Twitter and studying strategy effectiveness across communities.The authors also propose investigating whether counterspeakers change hate users’ mindsets.Source: Conclusion and future works · p. 10