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

1 Introduction

The growing volume of hate speech (HS) on social media represents a major societal challenge, requiring automated tools that go beyond surface-level detection. Hate speech can appear in explicit, implicit, or subtle forms, with the latter relying on coded, indirect, or context-dependent language that is particularly difficult to identify and interpret [Ocampo et al., 2023]. While advances in Natural Language Processing (NLP) and Generation (NLG) have improved detection of explicit content, understanding and addressing nuanced forms of hate remains an open problem. To address this challenge, the tool PEACE was introduced (Providing Explanations and Analysis for Combating Hate Expressions) [Damo et al., 2024], a web-based system designed to support exploration, detection, and explanation of explicit, and implicit HS. PEACE enables content moderators and researchers to analyze hateful messages, inspect model predictions, and obtain natural language explanations clarifying why a message is considered hateful. In this paper, we present PEACE 2.0, an extended version that moves beyond analysis and explanation to support actionable responses through automatic counterspeech (CS) generation. Counter-speech challenges hateful messages with factual information or alternative perspectives, offering a constructive alternative to content removal or censorship [Benesch, 2014]. However, generating effective CS is particularly challenging for implicit hate, where responses must be both informative and carefully framed. PEACE 2.0 introduces three key new functionalities to address these challenges: 1. Knowledge-grounded counter-speech generation. A Retrieval-Augmented Generation (RAG) pipeline, as in [Damo et al., 2025a], retrieves evidence from authoritative human rights sources and conditions counter-speech on this information. Users can explicitly compare responses generated with and without retrieval, highlighting the impact of knowledge grounding on factuality, relevance, and effectiveness. 2. Evidence-grounded explanations for hate speech classification. Using the same RAG mechanism, PEACE 2.0 generates explanations that justify the predictions of a finetuned BERT classifier, grounding label decisions in retrieved evidence to improve transparency and interpretability. 3. Visual analytics for counter-speech exploration. Interactive tools allow analysis of numerous counter-speech datasets supporting experimental evaluation of counter-speech performance on explicit versus implicit hate speech and offering insights into the role of retrieval-based grounding.

RelatedWork. Existing systems address HS detection, visualization, monitoring, or CS generation separately. Tools such as RECAST [Wright et al., 2021], MUDES [Ranasinghe and Zampieri, 2021], MUTED [Tillmann et al., 2023], IFAN [Mosca et al., 2023], CRYPTEXT [Le et al., 2023], TweetNLP [Camacho-collados et al., 2022], and McMillanMajor et al.'s framework [McMillan-Major et al., 2022] provide detection and analysis interfaces, while dashboards like the Indonesian HS monitoring system [Wijanarko et al., 2024] support large-scale tracking. In the CS domain, work is limited; CounterHelp [Babic et al., 2025] leverages large language models for context-sensitive responses. However, prior systems treat these components in isolation, focus mainly on explicit hate, and do not systematically compare knowledgegrounded and non-grounded generation.

Overall, PEACE 2.0 offers a unified, interactive platform for hate speech exploration, detection, explanation, and response generation, bridging analytical insights and practical mitigation. To our knowledge, it is the only online tool providing in-depth analysis of both explicit and implicit hate speech together with knowledge-grounded CS generation.

2 PEACE 2.0Main Functionalities

In this section, we report the main features of the original demo, and the new functionalities of PEACE 2.0 1

2.1 Data Exploration and Visualization

This module offers interactive visualizations for exploring HS and CS datasets. Users can switch between the two views.

Hate Speech. The hate speech view covers implicit hate datasets used in PEACE: the Implicit Hate Corpus (IHC) [ElSherief et al., 2021], Implicit and Subtle Hate (ISHate) [Ocampo et al., 2023], TOXIGEN [Hartvigsen et al., 2022], DynaHate (DYNA) [Vidgen et al., 2021], and Social Bias Inference Corpus (SBIC) [Sap et al., 2020]. Messages are organized by hatefulness, implicitness, and target group, with sanitized labels for consistency, that users can filter. Visualizations include: Sankey diagrams linking target groups and HS categories (explicit vs. implicit) with topic distributions derived via Latent Dirichlet Allocation (LDA); Word Clouds displaying the most frequent lexical items within selected attributes; and Target Frequency charts displaying attacked groups across datasets and implicitness levels.

Counter-speech. In PEACE 2.0, the same visual framework is extended to CS datasets, including CONAN [Chung et al., 2019], Multitarget-CONAN [Fanton et al., 2021], Twitter and YouTube datasets [Mathew et al., 2020; Albanyan and Blanco, 2022; Mathew et al., 2019], knowledge-grounded human expert datasets [Bonaldi et al., 2025; Chung et al., 2021], and RAG-generated responses from multiple LLMs and retrieval strategies [Damo et al., 2025a]. Labels follow the same sanitization scheme as PEACE. Sankey diagrams connect targets, counter-speech sources (expert, user, RAG, No-RAG), and LDA topics; Word Clouds highlight frequent terms within filtered messages, and Frequency charts display CS volume per target and source, and the distribution of different CS strategies across targets.

2.2 Data Augmentation

This module generates adversarial examples by modifying messages while preserving their implicit hateful meaning. The aim is to augment data for implicit HS by altering surface-level elements without changing the underlying stance. Following [Ocampo et al., 2023], the strategies are: named entity replacement; adjustment of scalar adverbs; addition of adverbial modifiers; adjective synonym substitution; replacement of domain-specific expressions with semantically similar variants; Easy Data Augmentation (random replacement, insertion, swap, deletion); and back-translation. Users can apply these methods to custom inputs.

2.3 Hate Speech Detection and Explanation

This module allows users to input custom messages for binary hate speech classification. The system outputs the predicted label with its confidence score. In addition, PEACE 2.0 generates concise, human-readable explanations conditioned on the message, predicted label, and probability, helping users understand the rationale behind the decision. PEACE 2.0 further introduces an optional RAG pipeline from [Damo et al., 2025a], to produce evidence-grounded explanations. When enabled, the system retrieves relevant passages from an authoritative knowledge base, summarizes them into factual context, and integrates this information into the LLM prompt. Users can choose whether to use RAG and select the underlying LLM. For transparency, the system also shows the retrieved paragraphs together with their similarity score.

1 The demo is available at https://3ia-demos.inria.fr/peace/ and the video by clicking here. We also provide a public API at https://gitlab.inria.fr/nocampo/peace.

2.4 Knowledge-Grounded CS Generation

Building on the RAG pipeline, a central innovation in PEACE 2.0 is evidence-grounded CS generation based on a curated human rights knowledge base. Users can input a hateful message and generate counter-speech responses with or without knowledge grounding, enabling direct comparison of evidence on outputs in terms of relevance, factuality, and persuasiveness. Multiple LLMs are supported, offering flexibility in style and quality. The CS generation pipeline consists of three steps. First, the input message is encoded using the BGEM3 sentence transformer, and FAISS performs inner-product similarity search over precomputed paragraph embeddings to retrieve the top-3 most relevant evidence passages, with deduplication applied. Second, the retrieved passages are concatenated and summarized into a concise summary using the selected LLM. Finally, the original message and the evidence summary are provided to the same LLM to generate a respectful and persuasive CS response suitable for social media. Users may also disable retrieval to compare RAG and non-RAG outputs. For transparency, the system also shows the retrieved paragraphs together with their similarity score.

Available models. For detection, the demo uses a BERT classifier fine-tuned on the ISHate training set. For explanation and CS generation, PEACE 2.0 supports opensource LLMs: Mistral (Mistral-7B-Instruct-v0.3), LLaMa (Llama-3.1-8B-Instruct), and CommandR (c4ai-command-r7b-12-2024). The knowledge base comprises 32,792 documents from the United Nations Digital Library, Eur-Lex, and the European Agency for Fundamental Rights (from 2000 to 2025), totaling 3,173,630 tokenized paragraphs.

3 Experiments and Results

3.1 Experimental Setting

We hypothesize that RAG-grounded explanations and CS will outperform non-RAG generations for both explicit and implicit HS, demonstrating the value of evidence-grounded generation. Specifically, we test: H1: RAG outputs are better overall and more informative for both explanations and CS. H2: RAG improves persuasiveness; H3: RAG improves outputs for both implicit and explicit cases. To evaluate this, from the implicit HS datasets (IHC, ISHate, TOXIGEN, DYNA, SBIC), we randomly sample 20 messages per dataset, evenly split between explicit and implicit, resulting in 100 HS examples. For each message, we generate explanations with and without RAG (100 RAG, 100 non-RAG). The same procedure is applied to CS, producing 200 responses. We assess all generations through both automatic and human evaluation.

Human Evaluation. We evaluate a subset of 100 explanations and 100 CS responses. For each task, 50 explicit and 50 implicit (25 RAG, 25 non-RAG) cases are sampled, resulting in 200 human-evaluated instances overall. Each example is assessed by three trained annotators. Evaluation criteria for explanations follow PEACE, while CS metrics are adapted from [Zheng et al., 2023; Bonaldi et al., 2024; Damo et al., 2025b]. Both explanations and CS are rated on: Fluency (F) (grammatical correctness), Informativeness (I) (relevant contextual or factual content), Persuasiveness (P) (ability to convince the hater and foster empathy in bystanders), Soundness (SO) (logical coherence), and Specificity (SP) (direct engagement with the HS and target). All dimensions are rated on a 1-5 Likert scale (5 = highest).

Automatic metrics. We compute automatic measures of linguistic quality, diversity, and faithfulness. These include Distinct-3 for lexical diversity; Semantic Similarity (Sentence-BERT) between HS and generated explanations or counter-speech; and Perplexity as a proxy for fluency. For RAG outputs, we also measure faithfulness via similarity between generations and retrieved evidence. Finally, NLI-based metrics (entailment and contradiction with roberta-large-nli) assess whether outputs appropriately address the original HS and remain consistent with the supporting evidence.

3.2 Results

Human evaluation. FromTable 1, RAG-generated outputs consistently outperform non-RAG outputs across all metrics. Supporting H1, RAG outputs are significantly more informative, particularly for implicit content (Explanation-Imp.: 4.64 vs. 2.72; Counter-speech-Imp.: 4.80 vs. 2.86), and of better quality (see highlighted results). In line with H2, RAG also improves persuasiveness for both explicit and implicit hate (e.g., Explanation-Exp.: 4.68 vs. 3.52; Counter-speechExp.: 4.68 vs. 2.64). Consistent with H3, gains are larger for both implicit and explicit cases, highlighting the benefit of retrieval in subtle contexts. Fluency and Soundness remain comparable across RAG and non-RAG outputs, indicating that these improvements do not compromise readability or coherence. Wilcoxon signed-rank tests confirm these differences are statistically significant (p < 0.05). Inter-annotator agreement, measured using Krippendorff's α, is substantial to perfect across dimensions (k = 0.57-1).

Automatic metrics. From Table 2, we see that RAGgenerated outputs consistently outperform No-RAG in both Explanation and CS tasks. They achieve higher semantic similarity, are faithful to retrieved evidence, have lower perplexity, indicating more informative and fluent outputs, while maintaining diversity (Distinct-3). NLI metrics further show that RAG outputs have higher Hate-Entailment and low Evidence-Contradiction, indicating that they appropriately address the content of the original HS without supporting it, and are aligned with retrieved content. Gains are for both implicit and explicit content, confirming that RAG improves informativeness, persuasiveness, and faithfulness across both tasks. Differences are statistically significant with Wilcoxon signed-rank tests.

4 Conclusion

By integrating retrieval, summarization, and generation within a single pipeline, the PEACE 2.0 tool bridges analytical understanding and practical mitigation of both implicit and explicit HS messages, producing evidence-backed CS that improves trustworthiness and effectiveness. The explanation and generation modules enhance the transparency of the abusive language classification and the CS generation, making PEACE 2.0 a suitable tool for e-democracy applications with the aim to enhance inclusiveness and fairness. Future work includes the integration of adaptive retrieval strategies, dynamic knowledge base updates, and the inclusion of evaluation metrics to assess CS quality.

Acknowledgments

This work has been partially supported by the French government, through the 3IA Cˆote d'Azur investments in the project managed by the National Research Agency (ANR) with the reference number ANR-23-IACL-0001.