MOCHEG: Multi-modal Multi-hop Fact-Checking with Explanations
Multimodal fact-checking requiring reasoning over both text and images. Annotators verify claims using multimodal evidence, identify relevant spans in text evidence, and provide explanations for their verdicts. Supports multi-hop reasoning across modalities.
Archivo de configuraciónconfig.yaml
# MOCHEG: Multi-modal Multi-hop Fact-Checking with Explanations
# Based on Yao et al., SIGIR 2023
# Paper: https://dl.acm.org/doi/10.1145/3539618.3591766
#
# Multimodal fact-checking requires verifying claims using evidence from
# multiple modalities (text and images). Annotators must reason across
# both modalities and across multiple evidence pieces (multi-hop) to
# arrive at a verdict.
#
# Annotation Guidelines:
# 1. Read the claim carefully
# 2. Examine ALL evidence: both the text evidence and the image evidence
# 3. Highlight the specific spans in the text evidence that support your verdict
# 4. Assign a verdict: Supported, Refuted, or Not Enough Evidence
# 5. Write a brief explanation justifying your verdict
# 6. Consider whether the image corroborates, contradicts, or is irrelevant
# to the textual evidence
#
# Verdict Definitions:
# - Supported: The claim is fully supported by the combined evidence
# - Refuted: The evidence directly contradicts the claim
# - Not Enough Evidence: The evidence is insufficient to confirm or deny
annotation_task_name: "MOCHEG: Multimodal Fact-Checking"
task_dir: "."
data_files:
- sample-data.json
item_properties:
id_key: "id"
text_key: "text"
output_annotation_dir: "annotation_output/"
output_annotation_format: "json"
annotation_schemes:
# Step 1: Verdict
- annotation_type: radio
name: verdict
description: "Based on the text and image evidence, what is the verdict for this claim?"
labels:
- "Supported"
- "Refuted"
- "Not Enough Evidence"
keyboard_shortcuts:
"Supported": "s"
"Refuted": "r"
"Not Enough Evidence": "n"
tooltips:
"Supported": "The claim is fully supported by the combined multimodal evidence"
"Refuted": "The evidence directly contradicts the claim"
"Not Enough Evidence": "The available evidence is insufficient to confirm or deny the claim"
# Step 2: Highlight relevant evidence spans
- annotation_type: span
name: evidence_spans
description: "Highlight the specific parts of the text evidence that are relevant to your verdict."
labels:
- "Supporting Evidence"
- "Contradicting Evidence"
- "Contextual Information"
label_colors:
"Supporting Evidence": "#22c55e"
"Contradicting Evidence": "#ef4444"
"Contextual Information": "#3b82f6"
tooltips:
"Supporting Evidence": "Text that directly supports the claim"
"Contradicting Evidence": "Text that directly contradicts the claim"
"Contextual Information": "Text that provides useful context but does not directly support or refute"
allow_overlapping: false
# Step 3: Image relevance
- annotation_type: radio
name: image_relevance
description: "How relevant is the image evidence to verifying this claim?"
labels:
- "Highly relevant - key to verdict"
- "Somewhat relevant - provides additional context"
- "Not relevant - does not help verify the claim"
tooltips:
"Highly relevant - key to verdict": "The image is essential for determining the verdict"
"Somewhat relevant - provides additional context": "The image provides useful but non-essential context"
"Not relevant - does not help verify the claim": "The image does not contribute to fact-checking"
# Step 4: Explanation
- annotation_type: text
name: explanation
description: "Explain your verdict. Describe how the text and image evidence together support your conclusion."
html_layout: |
<div style="margin-bottom: 15px; padding: 12px; background: #fef3c7; border-radius: 6px; border-left: 4px solid #f59e0b;">
<strong>Claim to Verify:</strong>
<div style="font-size: 17px; margin-top: 6px;">{{text}}</div>
</div>
<div style="margin-bottom: 10px; padding: 8px; background: #f0f4f8; border-radius: 4px;">
<strong>Source:</strong> {{source}}
</div>
<div style="display: flex; gap: 15px; margin-bottom: 15px;">
<div style="flex: 1; padding: 12px; background: #f9fafb; border-radius: 6px; border: 1px solid #e5e7eb;">
<strong>Text Evidence:</strong>
<div style="margin-top: 8px; line-height: 1.6;">{{evidence_text}}</div>
</div>
<div style="flex: 1; text-align: center;">
<strong>Image Evidence:</strong><br/>
<img src="{{image_url}}" style="max-width: 100%; max-height: 400px; margin-top: 8px; border: 1px solid #ddd; border-radius: 4px;" />
</div>
</div>
allow_all_users: true
instances_per_annotator: 50
annotation_per_instance: 3
allow_skip: true
skip_reason_required: false
Datos de ejemplosample-data.json
[
{
"id": "mocheg_001",
"text": "The Eiffel Tower was painted red for the 2024 Paris Olympics opening ceremony.",
"evidence_text": "The Eiffel Tower has been repainted approximately every seven years since its construction. The tower's current color is a specially mixed shade called 'Eiffel Tower Brown.' For the 2024 Olympics, the city of Paris undertook extensive renovations of the tower area but the structure's color remained unchanged. The opening ceremony featured spectacular light shows projected onto landmarks across the city.",
"image_url": "https://example.com/mocheg/eiffel_tower_001.jpg",
"source": "Social Media Claim"
},
{
"id": "mocheg_002",
"text": "NASA's Perseverance rover discovered liquid water on the surface of Mars in 2023.",
"evidence_text": "NASA's Perseverance rover has been exploring Jezero Crater on Mars since February 2021. The rover has collected rock samples and analyzed soil composition. In 2023, the rover's instruments detected mineral deposits consistent with ancient water activity. Scientists have confirmed evidence of briny water beneath the surface using orbital radar data, but no liquid water has been observed on the current Martian surface.",
"image_url": "https://example.com/mocheg/mars_rover_002.jpg",
"source": "News Article"
}
]
// ... and 8 more itemsObtener este diseño
Clone or download from the repository
Inicio rápido:
git clone https://github.com/davidjurgens/potato-showcase.git cd potato-showcase/multimodal/mocheg-multimodal-factcheck potato start config.yaml
Detalles
Tipos de anotación
Dominio
Casos de uso
Etiquetas
¿Encontró un problema o desea mejorar este diseño?
Abrir un issueDiseños relacionados
Check-COVID: Fact-Checking COVID-19 News Claims
Fact-checking COVID-19 news claims. Annotators verify claims against evidence, identify supporting/refuting spans, and provide verdicts with explanations. Based on the Check-COVID dataset targeting misinformation during the pandemic.
Biomedical Entity Linking (MedMentions)
Entity mention detection and UMLS concept linking for biomedical text based on MedMentions. Annotators identify biomedical entity mentions in PubMed abstracts and link them to UMLS Concept Unique Identifiers (CUIs), supporting large-scale biomedical knowledge base construction and clinical NLP.
Clickbait Spoiling
Classification and extraction of spoilers for clickbait posts, including spoiler type identification and span-level spoiler detection. Based on SemEval-2023 Task 5 (Hagen et al.).