💬 NLP

PEACE 2.0: Grounded Explanations and Counter-Speech for Combating Hate Expressions

PEACE 2.0 is a novel tool that enhances hate speech analysis and mitigation by utilizing a Retrieval-Augmented Generation pipeline to provide evidence-grounded explanations and automatically generate factual counter-speech for both explicit and implicit hateful messages.

Original authors: Greta Damo, Stéphane Petiot, Elena Cabrio, Serena Villata

Published 2026-02-20

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Greta Damo, Stéphane Petiot, Elena Cabrio, Serena Villata

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the internet as a giant, bustling town square. Recently, this square has become overrun with people shouting insults, spreading rumors, and trying to hurt others with their words. This is hate speech.

For a long time, the "security guards" of this town square (the computer programs we use) have been really good at spotting the obvious bullies. If someone yells, "I hate you because of your skin color," the guard immediately points and says, "Stop!"

But here's the problem: The bullies have gotten smarter. They aren't just shouting; they are whispering, using code words, or making jokes that only some people understand. This is implicit hate speech. It's like a bully saying, "Some people just don't belong here," without ever naming who they mean. It's much harder to catch.

Furthermore, even when the guards catch a bully, they usually just throw them out (delete the post). But what if we could teach the bully a lesson, or help the bystanders understand the truth? That's where Counter-Speech comes in—answering hate with facts and kindness instead of just silence.

Enter PEACE 2.0. Think of this tool as a super-powered, high-tech "Peacekeeper" for our town square. It doesn't just catch the bullies; it explains why they are wrong and helps write a perfect, fact-based reply to stop the argument.

Here is how PEACE 2.0 works, broken down into three simple superpowers:

1. The "Fact-Checker" Backpack (RAG)

Imagine you are a student in a debate. If you try to argue without your textbooks, you might make things up or sound unsure. But if you have a backpack full of the world's most trusted books (like the UN's human rights laws), you can win any argument with facts.

PEACE 2.0 has this backpack. It uses a system called RAG (Retrieval-Augmented Generation).

Without the backpack: The computer tries to guess a reply based on what it "thinks" is true.
With the backpack: The computer instantly searches its library of 30,000+ trusted documents, finds the exact facts needed, and uses them to build its reply.

This means when PEACE 2.0 responds to hate, it's not just guessing; it's citing the law and history.

2. The "Detective's Magnifying Glass" (Explanations)

Sometimes, a message looks innocent but is actually a trap. PEACE 2.0 acts like a detective with a magnifying glass.

It reads a message and decides: "Is this hate speech?"
Then, it doesn't just say "Yes." It writes a note explaining exactly why.
The Magic: Because of the "Fact-Checker Backpack," the detective doesn't just say, "This feels mean." It says, "This sentence is hate speech because it violates Article 12 of the Human Rights Declaration, which protects group X." It connects the dots between the rude comment and the actual rules of the town.

3. The "Diplomat's Pen" (Generating Responses)

This is the newest and coolest feature. Once PEACE 2.0 spots the hate and finds the facts, it writes a response for you.

It doesn't write an angry rant back.
It writes a Counter-Speech: a calm, persuasive, and fact-filled reply that shuts down the hate without starting a fight.
It can show you two versions: one written by a "smart guesser" (no facts) and one written by the "expert with the backpack" (facts). The expert version is always better, more convincing, and harder to argue with.

Why Does This Matter?

The researchers tested PEACE 2.0 against the old way of doing things.

The Result: When PEACE 2.0 used its "Fact-Checker Backpack," its explanations were much clearer, and its replies were much more persuasive.
The Analogy: It's the difference between a neighbor saying, "That's not true, I think so," versus a neighbor saying, "That's not true. According to the official census data from 2024, the population is actually..." The second one wins the argument every time.

The Bottom Line

PEACE 2.0 is a tool that helps us move from just hiding hate speech to understanding and countering it. It combines the brainpower of a detective, the knowledge of a librarian, and the diplomacy of a peacemaker.

Its goal is to make our online town square a place where we don't just delete the bad words, but where we can actually have better, more informed conversations that make everyone feel safer and included. It's not just about stopping the noise; it's about bringing the peace.

1. Problem Statement

The proliferation of hate speech (HS) on online platforms presents a critical societal challenge. While Natural Language Processing (NLP) has advanced in detecting hate speech, effectively responding to it remains an open problem.

The Gap: Existing tools often focus solely on detection or explicit hate speech. They lack the ability to generate constructive, fact-based responses (counter-speech) for implicit or subtle hate, which relies on coded or context-dependent language.
The Need: There is a need for systems that not only explain why a message is hateful but also generate evidence-grounded counter-speech (CS) that challenges hate with facts and alternative perspectives, rather than relying solely on content removal.

2. Methodology: PEACE 2.0 Architecture

PEACE 2.0 is an extended version of the original PEACE tool, designed as a unified web-based platform. Its core innovation is the integration of a Retrieval-Augmented Generation (RAG) pipeline to ground both explanations and counter-speech in authoritative knowledge.

Core Components

Data Exploration & Visualization:
- Supports both Hate Speech (HS) and Counter-Speech (CS) datasets.
- HS Datasets: Includes Implicit Hate Corpus (IHC), ISHate, TOXIGEN, DynaHate, and SBIC.
- CS Datasets: Includes CONAN, Multitarget-CONAN, Twitter/YouTube datasets, and human-expert/RAG-generated responses.
- Visuals: Uses Sankey diagrams (linking targets, categories, and topics), Word Clouds, and Frequency charts to analyze patterns across explicit/implicit levels and target groups.
Data Augmentation:
- Generates adversarial examples to augment implicit HS data by modifying surface-level elements (e.g., named entity replacement, scalar adverb adjustment, synonym substitution) while preserving the underlying hateful stance.
Hate Speech Detection & Explanation:
- Detection: Uses a fine-tuned BERT classifier (trained on ISHate) for binary classification.
- Explanation: Generates natural language explanations for predictions.
- RAG Integration: When enabled, the system retrieves relevant passages from an authoritative knowledge base (UN Digital Library, Eur-Lex, FRA), summarizes them, and integrates this context into the Large Language Model (LLM) prompt to produce evidence-grounded explanations.
Knowledge-Grounded Counter-Speech Generation:
- Pipeline:
  1. Retrieval: The input hateful message is encoded using the BGE-M3 sentence transformer. FAISS performs inner-product similarity search to retrieve the top-3 relevant evidence passages from a curated knowledge base (32,792 documents, ~3.17M tokenized paragraphs).
  2. Summarization: The retrieved passages are summarized by the selected LLM.
  3. Generation: The original message and the evidence summary are fed to the LLM to generate a respectful, persuasive counter-speech response.
- Models: Supports open-source LLMs: Mistral-7B, Llama-3.1-8B, and CommandR.
- Comparison: Users can toggle RAG on/off to compare grounded vs. non-grounded outputs.

3. Key Contributions

Unified Platform: The first online tool providing in-depth analysis of both explicit and implicit hate speech alongside knowledge-grounded counter-speech generation.
RAG for CS: Introduces a novel pipeline that conditions counter-speech generation on authoritative human rights sources, ensuring responses are factual and relevant.
Evidence-Grounded Explanations: Moves beyond "black box" predictions by justifying classification decisions with retrieved legal and human rights evidence.
Visual Analytics: Provides interactive tools to explore the characteristics of counter-speech strategies across different targets and sources (expert vs. AI).
Comparative Framework: Enables direct user comparison between RAG-generated and non-RAG outputs to demonstrate the impact of knowledge grounding.

4. Experimental Results

The authors evaluated the system on 100 hate speech examples (50 explicit, 50 implicit) from five datasets, generating 200 responses (100 explanations, 100 counter-speech) with and without RAG.

Evaluation Metrics

Human Evaluation: Rated by 3 annotators on Fluency (F), Informativeness (I), Persuasiveness (P), Soundness (SO), and Specificity (SP) using a 1-5 Likert scale.
Automatic Metrics: Included Semantic Similarity, Perplexity (fluency), Distinct-3 (diversity), Faithfulness (similarity to evidence), and NLI-based metrics (Entailment/Contradiction).

Key Findings

Superiority of RAG: RAG-generated outputs consistently outperformed non-RAG outputs across all metrics.
- Informativeness: Significant gains for implicit hate (Explanation: 4.64 vs. 2.72; CS: 4.80 vs. 2.86).
- Persuasiveness: RAG improved persuasiveness for both explicit and implicit cases (e.g., CS Explicit: 4.68 vs. 2.64).
Quality Maintenance: Fluency and Soundness remained comparable between RAG and non-RAG, indicating that adding retrieval did not compromise readability or coherence.
Automatic Metrics: RAG outputs showed higher semantic similarity to the input, lower perplexity (more fluent), and higher faithfulness to retrieved evidence. NLI metrics confirmed that RAG outputs appropriately addressed the hate without contradicting the supporting evidence.
Statistical Significance: Wilcoxon signed-rank tests confirmed that the improvements were statistically significant ( $p < 0.05$ ).

5. Significance

Trust and Transparency: By grounding explanations and responses in verifiable human rights documents, PEACE 2.0 increases the trustworthiness of automated moderation tools.
Mitigation Strategy: It offers a constructive alternative to censorship, providing a scalable method to challenge hate speech with facts, which is particularly crucial for implicit hate that is difficult to detect and remove.
E-Democracy Application: The tool is positioned as a resource for enhancing inclusiveness and fairness in digital spaces, supporting content moderators and researchers in understanding and countering nuanced hate.
Future Directions: The authors plan to integrate adaptive retrieval strategies, dynamic knowledge base updates, and more sophisticated quality assessment metrics for counter-speech.