Source-linked AI summary
CONAN -- COunter NArratives through Nichesourcing: a Multilingual Dataset of Responses to Fight Online Hate Speech
Y. L. Chung, E. Kuzmenko, S. S. Tekiroglu, M. Guerini
TL;DR
Online hate content is growing, while deletion and suspension can raise concerns about censorship and overblocking, and counter-narrative research lacks large suitable datasets. The paper constructs CONAN, a multilingual expert-based dataset of hate speech/counter-narrative pairs with metadata and augmented data. The dataset contains 4078 pairs, and experiments assess pair quality and show that augmented data improves automatic response selection.
Problem
Online hate content is difficult to address, while counter-narrative research lacks large datasets of appropriate responses for supervised approaches.
Method
The authors nichesource expert-written hate speech/counter-narrative pairs through three NGOs and add demographic, hate-type, response-type, translation, and paraphrasing annotations.
Results
CONAN contains 4078 pairs across English, French, and Italian; augmented data increased agreed “very relevant” responses from 18% to 27% in a tf-idf selection model.
Takeaways & Limitations
The dataset provides an expert-based multilingual resource for studying and developing counter-narratives against online hate speech.
Takeaways & Limitations
Future work must expand collection beyond Islam to targets such as migrants or LGBT+ and develop a counter-narrative generation tool.
Abstract
from arXiv · showhide
Although there is an unprecedented effort to provide adequate responses in terms of laws and policies to hate content on social media platforms, dealing with hatred online is still a tough problem. Tackling hate speech in the standard way of content deletion or user suspension may be charged with censorship and overblocking. One alternate strategy, that has received little attention so far by the research community, is to actually oppose hate content with counter-narratives (i.e. informed textual responses). In this paper, we describe the creation of the first large-scale, multilingual, expert-based dataset of hate speech/counter-narrative pairs. This dataset has been built with the effort of more than 100 operators from three different NGOs that applied their training and expertise to the task. Together with the collected data we also provide additional annotations about expert demographics, hate and response type, and data augmentation through translation and paraphrasing. Finally, we provide initial experiments to assess the quality of our data.
1 Introduction
Online hate speech spreads rapidly, remains difficult to define and monitor across cultures and languages, and can deepen prejudice. CONAN addresses this challenge by collecting expert-written counter-narratives as an alternative to deletion or suspension.
- Defining hate speech is difficult because its meaning varies across cultures and languages.
- Online hate speech can spread quickly, deepen prejudice, and expose bystanders to false messages.
- Social media platforms and governments have implemented laws and policies, but evolving hate content remains difficult to identify.
- Counter-narratives provide non-negative, fact-bound responses that preserve freedom of speech and can counter stereotypes and misleading information.
- CONAN uses trained NGO operators to build a large-scale, multilingual, publicly available dataset of hate speech/counter-narrative pairs.
2 Related Work
Prior research provides many hate-speech datasets and detection methods, but counter-narrative research remains limited. The authors therefore identify a lack of suitable counter-narrative corpora and build CONAN to address it.
- Hate datasets: Publicly available hate-speech datasets commonly provide binary hateful-versus-non-hateful annotations across several languages and platforms.
- Hate datasets: Existing datasets may be ephemeral because copyright restrictions often limit distribution to tweet IDs.
- Hate detection: Hate-speech detection studies use supervised classifiers based on lexical resources, n-grams, knowledge bases, and deep neural networks.
- Hate countering: Only a limited number of studies investigate counter-narratives, including collected comments, simulation models, and discourse effects.
- Hate countering: No suitable counter-narrative corpora were available for the authors’ purposes, partly because naturally occurring responses often fail required standards.
3 CONAN Dataset
CONAN is designed as a durable, multilingual, expert-based dataset of hate speech/counter-narrative pairs, collected through NGOs while protecting operator identities. It also expands and annotates the data to support cross-lingual and counter-narrative research.
- Dataset characteristics: CONAN provides copy-free hate speech/counter-narrative pairs in English, French, and Italian for cross-lingual studies.French and Italian pairs were also translated into English to create parallel data.
- Dataset characteristics: The dataset was collected through nichesourcing from three NGOs using trained operators, with simulated social-media activity to protect their identities.Operators were distinct from NGO trainers and were gathered in NGO premises for the collection sessions.
- Annotation and expansion: Additional annotations cover operator demographics, hate-speech sub-topics, and counter-narrative types, including the newly added COUNTER-QUESTIONS category.Each counter-narrative can receive more than one type label, while translated pairs provide parallel data across languages.
- Collection and augmentation: Collection used prototypical Islamophobic texts and standardized language-specific forms instructing operators to write fact-bounded, non-offensive responses.Operators were encouraged to follow their intuitions and produce diverse reasonable responses rather than a few perfect examples.
- Collection and augmentation: More than 500 hours of collection produced 4,078 original pairs, later expanded to over 15,000 through hate-message paraphrasing and translation.The original collection included 1,288 English, 1,719 French, and 1,071 Italian pairs, with at least 111 operators participating.
- Dataset analysis: Annotator agreement was moderate for counter-narrative types (Cohen’s Kappa 0.55) but very high for hate-speech sub-topics (Cohen’s Kappa 0.92).Multiple counter-narrative labels occurred in more than 50% of cases, whereas most hate messages had one sub-topic label.
4 Evaluation
The evaluation tests whether paraphrasing preserves hate speech/counter-narrative pair quality, improves retrieval, and whether operator demographics affect counter-narrative preference. Paraphrased pairs remained close to original pairs in perceived coherence, augmented data improved retrieval relevance, and gender matching was associated with greater appreciation.
- Augmentation reliability: 85% of ORIGINAL pairs, 74% of PARAPHRASE pairs, and 4% of UNRELATED pairs were judged clearly tied (p < .001).The difference was statistically significant by χ2 test.
- Augmentation reliability: Augmented pairs were assessed as almost as good as original pairs.
- Augmentation for counter-narrative selection: Adding augmented data increased ‘very relevant’ retrieval responses by 9% absolute, from 18% to 27% (p < .01).The comparison used two tf-idf response retrieval models on 100 natural hate tweets.
- Impact of Demographics: Gender affected counter-narrative preference even when operators followed identical guidelines and received the same argument instructions.
5 Conclusion
CONAN is a large-scale, multilingual, expert-based dataset of hate speech/counter-narrative pairs for English, French, and Italian. It contains 4078 pairs with metadata and translation- and paraphrase-based expansion, while future work will broaden hate-target coverage and support generation tools.
- Dataset: CONAN contains 4078 expert-based hate speech/counter-narrative pairs across English, French, and Italian.
- Annotations and augmentation: The dataset includes expert demographics, hate speech sub-topics, counter-narrative types, translations, and paraphrases.
- Future work: Future work will collect more data for Islam, add hate targets such as migrants and LGBT+, and develop a counter-narrative generation tool.