Deceptive Review Detection

Author & Copyright © 2026 Kavindu Hansaka Jayasinghe
All rights reserved.

---

📌 Overview

This project implements a semantic deceptive review detection system that identifies internal contradictions within a single product review.

The solution follows a hybrid NLP + LLM approach, combining:

• Sentence-level semantic embeddings for efficient candidate selection
• Lightweight clustering and polarity analysis
• Granite 4.0 h Tiny LLM (GGUF, local inference, compulsory) for final contradiction reasoning

The system detects explicit, implicit, and expectation-level contradictions, fully aligned with Twist Digital’s problem definition.

---

🎯 Problem Alignment

This implementation directly addresses the Twist Digital technical assignment requirements:

• ✅ Detects contradictions within a single review
• ✅ Produces structured outputs (flag, confidence, explanation)
• ✅ Uses modern NLP / LLM-based techniques
• ✅ Goes beyond rule-based or classical NLP limitations
• ✅ Demonstrates clear architectural reasoning

---

🏗️ System Architecture

Input Review
   ↓
Sentence Segmentation
   ↓
Sentence Embeddings (MiniLM)
   ↓
Semantic Clustering (DBSCAN)
   ↓
Polarity Estimation (Positive vs Negative Anchors)
   ↓
Candidate Pair Filtering
   ↓
LLM Reasoning
   ↓
Contradiction Flag + Confidence + Explanation

Granite is always executed before producing a final decision, satisfying the compulsory LLM requirement.

---

🤖 Large Language Model

Granite LLM

• Base model:
  https://huggingface.co/ibm-granite/granite-4.0-h-tiny
• Execution format: GGUF (Q4_0)
• GGUF conversion: Created by the author (KavinduHansaka/granite-4.0-h-tiny-gguf)
• Runtime: Local inference using llama-cpp-python

LLM is used for semantic reasoning only, not text generation, ensuring deterministic and explainable behavior.

---

🧩 Design Rationale (Rubric-Oriented)

Why not classical NLP only?

Classical NLP techniques struggle with:

• Implicit contradictions
• Multi-sentence logical conflicts
• Expectation-level inconsistencies

LLM-based reasoning is required to address these reliably.

Why a hybrid approach?

• Embeddings → reduce computational cost
• LLM → authoritative semantic judgment

This balances accuracy, performance, and scalability.

---

⚙️ Features

• Detects explicit and implicit contradictions
• Handles expectation-level inconsistencies
• Produces human-readable explanations
• Confidence scoring (0–1)
• Fully local LLM execution
• REST API (FastAPI)
• Simple Tailwind-based UI
• Batch testing support

---

📂 Project Structure

deceptive-review-detection/
│
├── app.py
├── README.md
├── requirements.txt
├── test_reviews.py
│
├── src/
│   └── detector.py
│
├── models/
│   └── granite-4.0-h-tiny-Q4_0.gguf
│
├── frontend/
│   └── index.html
│
├── data/
    └── dataset.txt

---

🚀 How to Run

Install dependencies

pip install -r requirements.txt

Place Granite GGUF model

models/granite-4.0-h-tiny-Q4_0.gguf

Start the server

uvicorn app:app --reload

Open UI

http://127.0.0.1:8000

---

🧪 Testing

python test_reviews.py

Sample output:

❌ Contradiction: YES
Confidence: 0.72
Explanation:
Granite detected a logical inconsistency between durability and failure claims.

---

⚠️ Limitations

• Subtle sarcasm may remain ambiguous
• Confidence is heuristic-based
• Requires sufficient textual content

These limitations are inherent to natural language understanding tasks.

---

🛡️ License & Attribution

© 2026 Kavindu Hansaka Jayasinghe. All rights reserved.

• Codebase: Copyrighted by the author
• Granite base model: Apache-2.0 (IBM)
• GGUF conversion: Performed by the author
• GGUF redistribution subject to IBM Granite licensing

---