echo-feeling 2.1.0

Multimodal sentiment analysis for e-commerce: text + emoji + sticker fusion with ML classifiers.

Echo Feeling

Echo Feeling is a multimodal sentiment analysis library for e-commerce platforms. It goes beyond text-only approaches by integrating emojis and stickers as expressive modalities, enabling more accurate and nuanced detection of customer sentiment.

---

Features

• Multimodal Fusion – combines text (BoW + TF-IDF + optional BERT), emoji, and sticker signals into a unified feature vector
• Multiple ML Classifiers – SVM, Random Forest, and Feedforward Neural Network with k-fold cross-validation
• Admin Panel Ready – REST API (Flask) + Node.js wrapper for seamless e-commerce integration
• Suspicious Review Detection – automatically flags potentially fraudulent or abusive feedback
• Product-level Dashboard – aggregated sentiment counts, percentages, and emoji trends per product

---

Architecture

six-phase pipeline
──────────────────
Phase 1  E-Commerce Website (data source + admin panel)
Phase 2  Data Collection & Preprocessing
           └─ text normalisation → emoji standardisation → sticker labelling
Phase 3  Feature Extraction
           └─ BoW / TF-IDF / BERT  |  emoji sentiment scores  |  sticker encoded weights
Phase 4  Multimodal Fusion
           └─ concatenate + scale → unified feature vector
Phase 5  Model Training & Evaluation
           └─ SVM (F1 0.85) · Random Forest (F1 0.83) · Neural Network (F1 0.89*)
Phase 6  Deployment as Node.js module (Flask API + JS wrapper)

* Results reported in the SAMANWAYA'26 paper.

---

Project Structure

final/
├── echo_feeling/               # Installable Python package
│   ├── __init__.py
│   ├── api.py                  # Public API: analyze(), analyze_batch(), product_dashboard()
│   └── engine.py               # SentimentEngine (loads models, runs inference)
│
├── training_phase/             # Training pipeline
│   ├── preprocessor.py         # Text normalisation, emoji standardisation, sticker labelling
│   ├── feature_extractor.py    # BoW, TF-IDF, BERT, emoji & sticker extractors, fusion
│   └── train.py                # CLI training script
│
├── deployment_phase/           # Deployment
│   ├── server.py               # Flask REST API server
│   └── sentiment_module.js     # Node.js wrapper module
│
├── models/                     # Saved model artefacts (generated by train.py)
├── data/                       # Place your datasets here
│   ├── reviews.csv
│   └── stickers/
│       ├── positive/   *.png
│       ├── negative/   *.png
│       └── neutral/    *.png
│
├── requirements.txt
└── setup.py

---

Installation

pip install echo-feeling
# or from source:
pip install -e ".[server]"

---

Quick Start

Python

from echo_feeling.api import analyze, product_dashboard

# Single review (text + emoji + sticker)
result = analyze("Absolutely love this! 😍👍", sticker="positive")
print(result)
# {'label': 'positive', 'confidence': 0.94, 'scores': {...}, 'emoji_score': 0.875}

# Product-level dashboard (for admin panel)
reviews = [
    "Great quality, fast delivery! 😊",
    "Terrible product, waste of money 😡",
    "It's okay, nothing special",
]
summary = product_dashboard(reviews)
print(summary["counts"])        # {'positive': 1, 'negative': 1, 'neutral': 1}
print(summary["percentages"])   # {'positive': 33.3, 'negative': 33.3, 'neutral': 33.3}

Training

# 1. Prepare your CSV with columns: review, label, sticker
# 2. Organise sticker PNG images into data/stickers/positive|negative|neutral/

python training_phase/train.py \
  --data_path data/reviews.csv \
  --sticker_root data/stickers \
  --output_dir models

CSV format:

review,label,sticker
"Great product! 😊",positive,positive
"Worst purchase ever 😡",negative,negative
"It's okay",neutral,neutral

REST API Server

python deployment_phase/server.py --port 5000

# Analyse a review
curl -X POST http://localhost:5000/analyze \
     -H "Content-Type: application/json" \
     -d '{"review": "Amazing! 😍", "sticker": "positive"}'

# Get product dashboard
curl http://localhost:5000/product/PROD123

Node.js Integration

const sentiment = require('./deployment_phase/sentiment_module');

// Analyse a review
const result = await sentiment.analyze("Excellent product! 😊", "positive");
console.log(result.label); // "positive"

// Add review to a product
await sentiment.addReview("PROD123", "Amazing quality!", "positive");

// Get admin dashboard for a product
const summary = await sentiment.getProductSummary("PROD123");
console.log(summary.counts);

// Delete a suspicious review (admin moderation)
await sentiment.deleteReview(reviewId);

// Use as Express middleware
const express = require('express');
const app = express();
app.use(sentiment.sentimentMiddleware());
app.post('/review', async (req, res) => {
  const result = await req.sentiment.analyze(req.body.text);
  res.json(result);
});

---

API Reference

analyze(text, sticker="neutral", model_dir=None) → dict

Field	Type	Description
label	str	positive / negative / neutral / suspicious
confidence	float	Model confidence [0, 1]
scores	dict	Per-class probability
emoji_score	float	Aggregate emoji sentiment score [-1, 1]

analyze_batch(texts, stickers=None) → list[dict]

product_dashboard(reviews, stickers=None) → dict

Field	Type	Description
total_reviews	int	Total review count
counts	dict	Count per sentiment label
percentages	dict	Percentage per sentiment label
avg_confidence	float	Mean model confidence
avg_emoji_score	float	Mean emoji sentiment
flagged_suspicious	list[int]	Indices of suspicious reviews

---

Model Performance

Model	F1-Score	Precision	Recall
Neural Network	0.89	0.90	0.88
SVM	0.85	0.86	0.84
Random Forest	0.83	0.84	0.82

Evaluated with 5-fold stratified cross-validation on multimodal fused features.

---

License

MIT License – see LICENSE for details.

---

Authors

Created by Swathi Lekshmi SS, Romin Varghese, R Arun, Ivin Issac
College of Engineering Karunagappally
Guided by Hridhya J (Asst. Professor, CSE)

Published at SAMANWAYA'26 – AI-Powered Review Mining for Subjective Text Understanding