Organizations: Indian Institute of Information Technology Dharwad · Work done at FLaMe.nlp Lab, IIIT Delhi. · Manipal Institute of Technology, Manipal Academy of Higher Education Manipal, Karnataka, India · Indraprastha Institute of Information Technology Delhi
Subtle and indirect hate speech remains an underexplored challenge in online safety research, particularly when harmful intent is embedded within misleading or manipulative narratives. Existing hate speech datasets primarily capture overt toxicity, underrepresenting the nuanced ways misinformation can incite or normalize hate. To address this gap, we present HateMirage, a novel dataset of Faux Hate comments designed to advance reasoning and explainability research on hate emerging from fake or distorted narratives. The dataset was constructed by identifying widely debunked misinformation claims from fact-checking sources and tracing related YouTube discussions, resulting in 4,530 user comments. Each comment is annotated along three interpretable dimensions: Target (who is affected), Intent (the underlying motivation or goal behind the comment), and Implication (its potential social impact). Unlike prior explainability datasets such as HateXplain and HARE, which offer token-level or single-dimensional reasoning, HateMirage introduces a multi-dimensional explanation framework that captures the interplay between misinformation, harm, and social consequence. We benchmark multiple open-source language models on HateMirage using ROUGE-L F1 and Sentence-BERT similarity to assess explanation coherence. Results suggest that explanation quality may depend more on pretraining diversity and reasoning-oriented data rather than on model scale alone. By coupling misinformation reasoning with harm attribution, HateMirage establishes a new benchmark for interpretable hate detection and responsible AI research.
Figures & tables
Typical Hateful Comment
– I dislike the skin colour of people from < country > .
– < country > is dishonest toward everyone; we should boycott it.
– < group/religion > people are ruining our society; they should leave.
Faux Hate Comments
– < country > is spreading COVID on purpose to harm the world.
– It’s not love, it’s a plan; < group/religion > use relationships to change people.
Table 1: Illustrative examples comparing Traditional Hate comments with Faux Hate comments. Placeholders such as COUNTRY and GROUP/RELIGION denote target entities abstractly.
Dataset
Hate
Fake
Explanation
Interdisciplinary
HateXplain Mathew et al. (2021)
✓
✗
✗
✗
HatEval Basile et al. (2019)
✓
✗
✗
✗
Implicit Hate ElSherief et al. (2021)
✓
✗
✓
✗
HARE Yang et al. (2023)
✓
✗
✓
✗
CoAID Cui and Lee (2020)
✗
✓
✗
✗
FNVE Chen et al. (2025)
✗
✓
✓
✗
Table 2: Comparison of existing datasets across four key dimensions: hate speech labeling, fake/misinformation labeling, structured/multi-dimensional explanatory annotations, and interdisciplinary scope. The interdisciplinary attribute denotes the integration of two distinct domains within a single dataset, allowing their interaction to be studied rather than treated separately. ✓ indicates the presence of a feature, while ✗ indicates its absence. Unlike prior resources, HateMirage incorporates all four dimensions, supporting richer analysis and interpretability.
Comment
Fake Claim Base
Target
Intent
Implication
The coronavirus is just a scam planned by [Political Party] and [Country] to fool us all.
"[Country] is Using Covid as a Biological Weapon"
[Political Party], [Country]
To link [Political Party] and [Country] to a conspiracy theory about COVID-19.
Could fuel public anger against [Political Party] and [Country].
[Religious Group] and [Political Party], with [Person] in the mix, spread COVID to ruin [City] and harm [Country].
"[Religious Group] and [Political Party] collaborated to spread COVID-19."
[Political Party], [Person], [Religious Group]
To accuse [Political Party], [Person], and [Religious Group] of intentionally worsening the pandemic.
Could stir up communal tensions and increase distrust in political and religious groups.
[Religious Group] men are trapping innocent [Community] women in fake love to convert them. This is how [Country] is being taken over slowly.
"Love Jihad: [Religious Group] men are systematically luring women from other communities to convert them."
[Religious Group]
To portray [Religious Group] as deceptive and predatory, reinforcing the fabricated ’Love Jihad’ narrative.
Could deepen mistrust between communities, normalize religious profiling, and justify hostility toward [Religious Group].
Table 3: Representative examples of Faux Hate comments illustrating how explanations are derived from comments rooted in fake or manipulative narratives. Each example decomposes the comment into its underlying fake claim base, the entities targeted, inferred author intent, and the broader social or emotional implications.
Field
Avg Tokens
Median Tokens
Min Tokens
Max Tokens
Avg Sent
Max Sent Len
Multi-Sent(%)
Total Tokens
Train (3,624 samples)
Comments
30.56
29
3
800
2.82
240
76.49
110741
Target
1.48
1
1
27
1.00
12
0.30
5338
Intent
14.66
14
5
39
1.02
34
1.41
53101
Implication
15.16
15
6
48
1.01
38
0.86
54878
Test (906 samples)
Comments
30.28
27
4
316
2.84
78
73.18
27438
Target
1.51
1
1
44
1.01
24
0.44
1371
Table 4: Quantitative summary of the HateMirage dataset, reporting key statistics for the complete dataset and its train and test splits. Metrics include text length, sentence structure, and multi-sentence prevalence, providing an overview of the linguistic complexity and scale of the dataset.
Model
Target
Intent
Implication
SBERT
R-L (F1)
SBERT
R-L (F1)
SBERT
R-L (F1)
LLaMA-3.2-1B-Instruct
Zero-shot
50.19
30.91
57.99
26.25
47.04
12.79
LLaMA-3.2-3B-Instruct
53.43
34.62
57.67
26.81
47.76
16.36
LLaMA-3.1-8B-Instruct
48.29
21.39
56.68
27.95
51.24
15.62
Qwen-2.5-1.5B
53.57
35.93
57.86
25.56
43.41
14.45
Qwen-2.5-3B-Instruct
52.91
36.13
59.96
19.06
53.09
13.75
Table 5: Comparison of LLM performance across Zero-Shot and RAG-based settings. Highest results are highlighted in bold , and second-best results are marked with † . (Metrics reported in %)
Comment
Ground Truth
Model Prediction
Target
During the time of Corona, you also actively spread information about the <organization> . Have some shame for you were metaphorically “hit with a shoe on your face.” Please report the news with a bit of discretion.
Media / journalist
<organization>
Hey Brother, <organization> is spreading the coronavirus here because you’re not paying attention to it.
<organization>
Brother
Intent
<@user> Were the people from the group in Markaz there to catch corona?
To question the motives of the <organization> gathering and imply reckless behavior.
The commenter is questioning the intentions of people in Markaz, suggesting they might be there to spread misinformation about the COVID-19 pandemic.
I still advocate calling it the “kung-flu”. And label anyone suspected of being infected as “Flu Kang”.
To use racially insensitive language to mock and stigmatize people, especially Asians, in relation to COVID-19.
To spread hateful and fake narratives about the flu.
Implication
<@user> Why.. Did they have corona too? Oh, so you all have been spreading this disease since then… We’ve been blaming the <organization> for no reason… This has been going on traditionally.
It implies that the criticism directed toward <organization> was unwarranted and highlights bias or unfair treatment.
The comment could be interpreted as promoting general negativity or unrest, possibly encouraging harmful behavior.
If <organization> is responsible, then Namaste Trump is more responsible. <person> had given a warning on February 11 regarding Corona. He was mocked then. Who will you hold responsible for Namaste Trump? Trump or our PM? #shameOnRupani
This criticism aims to highlight perceived double standards and questions leadership decisions in handling the COVID-19 situation.
The comment could be interpreted as drawing a general comparison between different parties, possibly implying blame or criticism.
Table 6: Error analysis showing model predictions for Target, Intent, and Implication. Placeholders such as <organization> and <user> are used to anonymize entities and generalize examples.
SBERT
R-L (F1)
Target
0.7479
0.6293
Intent
0.7020
0.3458
Implication
0.5780
0.2013
Table 7: Human evaluation results comparing GPT-4 generated explanations with human-written explanations for 500 randomly sampled records.
Figure 1: Comparison of average ratings from two annotators for GPT-4-generated explanations across Target, Intent, and Implication.
Online hate speech and misinformation frequently overlap, yet NLP research has mainly treated them in isolation. While LLMs represent a scalable solution for assisting humans in the generation of counterspeech for both threats, zero-shot models frequently generate repetitive and vague responses, underscoring the need for high-quality examples to steer model generation. However, existing counterspeech datasets against the overlap of hate and misinformation are scarce and limited to single-turn English dialogues, while real-life interactions span across multiple turns and languages. To bridge this gap, we introduce the first large-scale, expert-curated, multilingual dataset of dialogues tackling the intersection of hate and misinformation. To ensure factual grounding, the dialogues are also anchored in verified external knowledge (i.e., fact-checking articles and NGO reports) and include document- and chunk-level span annotations, making it directly applicable for RAG systems. Covering five languages and targeting hate directed at seven marginalized groups, this novel resource enables the training and evaluation of more persuasive, factually grounded counterspeech models.
Helena Bonaldi, Genoveffa Martone, Marco Guerini
Fondazione Bruno Kessler, Italy · Università Cattolica del Sacro Cuore, Italy
Hate, derogatory, and offensive speech remains a persistent challenge in online platforms and public discourse. While automated detection systems are widely used, most focus on censorship or removal, raising concerns for transparency and freedom of expression, and limiting opportunities to explain why content is harmful. To address these issues, explanatory approaches have emerged as a promising solution, aiming to make hate speech detection more transparent, accountable, and informative. In this paper, we present a hybrid approach that combines Large Language Models (LLMs) with three newly created and curated vocabularies to detect and explain hate speech in English, French, and Greek. Our system captures both inherently derogatory expressions tied to identity characteristics and direct group-targeted content through two complementary pipelines: one that detects and disambiguates problematic terms using the curated vocabularies, and one that leverages LLMs as context-aware evaluators of group-targeting content. The outputs are fused into grounded explanations that clarify why content is flagged. Human evaluation shows that our hybrid approach is accurate, with high-quality explanations, outperforming LLM-only baselines.
Jason Liartis, Eirini Kaldeli, Lambrini Gyftokosta +2
National Technical University of Athens · Datoptron · Independent Researcher +3
Hate speech online targets individuals or groups based on identity attributes and spreads rapidly, posing serious social risks. Memes, which combine images and text, have emerged as a nuanced vehicle for disseminating hate speech, often relying on cultural knowledge for interpretation. However, existing multimodal hate speech datasets suffer from coarse-grained labeling and a lack of integration with surrounding discourse, leading to imprecise and incomplete assessments. To bridge this gap, we propose an agentic annotation framework that coordinates seven specialized agents to generate hierarchical labels and rationales. Based on this framework, we construct M^3 (Multi-platform, Multi-lingual, and Multimodal Meme), a dataset of 2,455 memes collected from X, 4chan, and Weibo, featuring fine-grained hate labels and human-verified rationales. Benchmarking state-of-the-art Multimodal Large Language Models reveals that these models struggle to effectively utilize surrounding post context, which often fails to improve or even degrades detection performance. Our finding highlights the challenges these models face in reasoning over memes embedded in real-world discourse and underscores the need for a context-aware multimodal architecture. Our dataset and code are available at https://github.com/mira-ai-lab/M3.
Rui Xing, Qi Chai, Jie Ma +5
MOE KLINNS Lab, Xi’an Jiaotong University · The Hong Kong University of Science and Technology (Guangzhou) · School of Cyber Science and Engineering, Xi’an Jiaotong University +1