🧠 Sentiment Analyzer

Aspect-level sentiment analysis with multilingual support and real-time insights

Understanding what people feel — about every aspect — in any language.

---

📖 Overview

SentimentVision is a full-stack AI application that performs Aspect-Based Sentiment Analysis (ABSA) on product and service reviews. It:

• 🎯 Extracts specific aspects from a review (e.g., Display, Battery Life, Camera)
• 💬 Detects sentiment for each aspect individually — Positive, Negative, or Neutral
• 🌐 Translates input automatically — supports Telugu, Hindi, and 200+ languages via NLLB-200
• 🔍 Handles typos gracefully using fuzzy matching (batter → Battery)
• ⚡ Runs fully offline — no API keys, no internet, no cloud inference
• 🎨 Serves a dark-themed Streamlit UI inspired by Modal.com

---

🏗️ Architecture

Sentiment_Analyzer/
│
├── app.py                      # Streamlit UI — hero header, textarea, output cards
├── model_pipeline.py           # Core ML pipeline — ASTE + translation + fuzzy logic
├── styles.css                  # Dark theme CSS — glowing input, neon accents, badge styles
├── run_app.bat                 # One-click launcher (double-click to run)
├── requirements.txt            # Python dependencies
├── README.md                   # Project documentation
│
├── .streamlit/
│   └── config.toml             # Native dark theme config (background, text color, accent)
│
└── Models/                     # AI model weights (see download instructions below)
    ├── aste_model.zip          # DeBERTa-v3 ASTE weights — extract to ../aste_model/
    └── Translation.zip         # NLLB-200 translation model — extract to ../Translation/

Model files live outside the app folder (not committed to git):

Projects/Minor/
├── aste_model/
│   ├── deberta_aste_weights.pt     # Custom-trained DeBERTa-v3 weights (736 MB)
│   ├── config.json                 # DeBERTa-v3-base model config
│   ├── tokenizer.json              # SentencePiece tokenizer
│   └── tokenizer_config.json
│
└── Translation/
    └── nllb_200_model/
        ├── model.safetensors       # NLLB-200 weights (2.46 GB)
        ├── config.json
        ├── tokenizer.json
        └── tokenizer_config.json

---

⚡ Quick Start

Prerequisites

• Python 3.12+
• ~4 GB free RAM
• ~3.5 GB disk space (for models)

1. Clone the Repository

git clone https://github.com/your-username/Sentiment_Analyzer.git
cd Sentiment_Analyzer

2. Download the AI Models

Extract Models/aste_model.zip → ../aste_model/ Extract Models/Translation.zip → ../Translation/nllb_200_model/

Verify the final paths:

• ..\aste_model\deberta_aste_weights.pt
• ..\Translation\nllb_200_model\model.safetensors

3. Create a Virtual Environment

python -m venv .venv
.venv\Scripts\activate
pip install -r requirements.txt

4. Run the App

Option A — Double-click launcher:

run_app.bat

Option B — Terminal:

.venv\Scripts\streamlit.exe run app.py

---

🧠 AI Models

1. ASTE Model — Aspect-Sentiment-Triple Extraction

Property	Details
Architecture	DeBERTa-v3-base (Microsoft)
Task	BIO sequence labeling (Aspect + Opinion span extraction)
Training Data	Custom product review dataset
Weights	deberta_aste_weights.pt (736 MB)
Inference	CPU — ~1–3 seconds per review

BIO Label Mapping:

Label	Meaning
O	Outside (not an aspect/opinion word)
B-Aspect	Beginning of an aspect span
I-Aspect	Inside an aspect span
B-Opinion	Beginning of an opinion span
I-Opinion	Inside an opinion span

2. Translation Model — NLLB-200 (Meta AI)

Property	Details
Architecture	M2M100 (encoder-decoder)
Languages	200+ including Telugu, Hindi, Tamil, Arabic, etc.
Weights	model.safetensors (2.46 GB)
Inference	CPU — ~30–60 seconds per review (lazy-loaded)
Loading	On-demand — only loads when non-English input is detected

---

🔍 How It Works

User Input (any language)
        │
        ▼
  Language Detection (langdetect)
        │
   Non-English? ──── Yes ────▶ NLLB-200 Translation ──▶ English Text
        │                                                       │
       No ◀─────────────────────────────────────────────────────┘
        │
        ▼
 DeBERTa-v3 Tokenization
        │
        ▼
 BIO Sequence Labeling (Aspect & Opinion spans)
        │
        ▼
 Span Consolidation (compound aspects: "Battery" + "Life" → "Battery Life")
        │
        ▼
 Fuzzy Typo Correction ("batter" → "Battery", cutoff = 0.82)
        │
        ▼
 Sentiment Scoring (clause isolation + negation detection)
        │
        ▼
 Output: [Aspect | Sentiment | Confidence Score]

---

🎯 Supported Aspect Categories

Category	Example Aspects
📱 Display	Screen, Display, Picture Quality
🔋 Battery	Battery, Battery Life
📷 Camera	Camera, Front Camera, Rear Camera
⚡ Performance	Processor, Speed, RAM, Performance
🔊 Audio	Speaker, Sound, Volume
🏗️ Build	Build Quality, Design, Body
💰 Price	Price, Value, Cost

---

🌐 Multilingual Support

SentimentVision automatically detects and translates 200+ languages before analysis:

Language	Script	Example Input
Telugu	తెలుగు	ఫోన్ డిస్‌ప్లే చాలా బాగుంది కానీ బ్యాటరీ అంత బాగా లేదు
Hindi	हिन्दी	फ़ोन की डिस्प्ले अच्छी है पर बैटरी ख़राब है
Tamil	தமிழ்	திரை மிகவும் நல்லது ஆனால் பேட்டரி மோசம்
English	Latin	The display is great but battery life is poor

---

🎨 Tech Stack

Layer	Technology
NLP / AI	DeBERTa-v3-base, PyTorch 2.x
Translation	NLLB-200 (Meta AI), Transformers 5.x
Language Detection	langdetect
Fuzzy Matching	Python difflib
Frontend	Streamlit, Custom CSS (Modal.com dark theme)
Tokenization	HuggingFace Transformers, SentencePiece
Environment	Python 3.12, uv / pip

---

📊 Example Output

Input: "Phone display is good but the battery is not good"

Aspect	Sentiment	Score
Display	✅ Positive	+1.00
Battery	❌ Negative	−1.00

---

📄 License

This project is licensed under the MIT License. See LICENSE for details.

---

Because average sentiment ratings never tell the full story. 🎯