CrossCheck Extension

A browser extension for automatic fact-checking using hybrid retrieval and natural language inference.

---

Motivation

Misinformation spreads rapidly on the internet.

This project explores how retrieval techniques and natural language inference can be combined to provide lightweight fact-checking directly inside the browser.

The goal is not only to build a usable browser extension, but also to investigate how retrieval quality affects downstream fact-checking performance.

---

Features

• Highlight text directly on webpages.
• Retrieve supporting evidence from Wikipedia.
• Hybrid retrieval using lexical and semantic search.
• Cross-Encoder evidence reranking.
• Sentence-level evidence selection.
• NLI-based verdict prediction.
• Display verdicts inside the browser popup.
• Provide evidence sentences and confidence scores.
• Cache previous claims to reduce repeated inference.
• Optional LLM inference for further analysis and explanation.

---

System Architecture

User Selection
      ↓
Browser Extension
      ↓
FastAPI Backend
      ↓
Topic Extraction
      ↓
Hybrid Retrieval
      ↓
Cross-Encoder Ranking
      ↓
Sentence Selection
      ↓
DeBERTa NLI
      ↓
Verdict + Evidence

---

Project Structure

crosscheck-extension/
│
├── backend/
│   ├── app/
│   │   ├── api/
│   │   ├── core/
│   │   ├── pipelines/
│   │   ├── services/
│   │   ├── schemas/
│   │   └── main.py
│   │
│   ├── tests/
│   │   ├── retrieval_test.py
│   │   └── nli_test.py
│   │
│   └── requirements.txt
│
├── extension/
│   ├── src/
│   │   ├── popup/
│   │   ├── background/
│   │   ├── content/
│   │   ├── services/
│   │   ├── types/
│   │   └── utils/
│   │
│   └── package.json
│
├── ml/
│   ├── evaluation/
│   ├── experiments/
│   └── kaggle/
│
└── README.md

---

Fact-Checking Pipeline

1. Topic Extraction
2. Entity Linking
3. Hybrid Retrieval
4. Cross-Encoder Ranking
5. Sentence Selection
6. NLI Prediction
7. Confidence Calibration
8. Claim Caching

---

Experiments

Several retrieval and inference strategies were investigated during development.

Retrieval Experiments

• BM25 retrieval
• Dense retrieval
• Entity-based retrieval
• Hybrid retrieval
• Graph-based retrieval

Evaluation Tasks

• Recall@k evaluation
• Mean Reciprocal Rank (MRR)
• End-to-end fact-check accuracy
• Evidence quality analysis

Experimental Environment

Experiments and benchmarks are executed independently inside the ml/ directory and Kaggle notebooks.

---

Evaluation

Model Evaluation

The NLI component was evaluated independently on the FEVER v1.0 dataset.

Label	Precision	Recall	F1
SUPPORTS	0.81	0.88	0.84
REFUTES	0.70	0.69	0.69
NOT_ENOUGH_INFO	0.81	0.75	0.78

Overall Metric	Score
Accuracy	0.77
Macro F1	0.77
Weighted F1	0.77

Dataset: FEVER v1.0
Evaluation samples: 19,998

Retrieval Performance

Metric	Score
Recall@1	0.69
Recall@3	0.84
Recall@5	0.86
MRR	0.75

End-to-End Performance

Metric	Score
Accuracy	0.41
F1 Score	0.395

---

Installation

Backend

`git clone https://github.com/24127382/CrossCheck-Extension`

cd backend

pip install -r requirements.txt

uvicorn app.main:app --reload

Extension

cd extension

npm install

npm run build

Load the generated extension inside the browser developer mode.

---

Limitations

• Retrieval quality strongly affects final predictions.
• Wikipedia may not contain evidence for recent events.
• Multi-hop reasoning is not supported.
• Confidence calibration remains imperfect.
• The system is currently limited to English claims.

---

Future Work

• Larger retrieval corpora.
• Multi-hop retrieval.
• Knowledge graph integration.
• LLM-based evidence summarization.
• Human feedback loop.
• Deployment and scalability improvements.
• Better confidence calibration.

---