Sentiment-Analysis-with-ALBERT-Explainable-AI-XAI-

Sentiment Analysis with ALBERT + Explainable AI (XAI)

Fine-tuned ALBERT model for 3-class sentiment analysis (Negative · Neutral · Positive)
With comprehensive explainability using SHAP, LIME, and Captum (Integrated Gradients)
Evaluated on both Twitter data and IMDb movie reviews

---

Overview

This project demonstrates:

• Fine-tuning ALBERT on a large ternary sentiment dataset (~500k tweets)
• Achieving high performance on balanced 3-class sentiment classification
• Providing human-interpretable explanations using three leading XAI methods
• Cross-domain evaluation on IMDb (long-form text) to test generalization

---

Model Performance

Dataset	Accuracy	F1-Score (Macro)	Negative	Neutral	Positive
Test Set (Tweets)	92.4%	0.91	0.93	0.89	0.94
IMDb (Cross-domain)	88.7%	0.88	0.90	—	0.87

Note: IMDb has no true neutral class → mapped to binary (neg=0, pos=2)

---

Explainable AI Visualizations

Input: "I absolutely love this phone, but the battery life is short."

Captum – Integrated Gradients


Green = Positive contribution | Red = Negative | "but" strongly pulls toward neutral

SHAP Text Plot


SHAP correctly highlights "love", "absolutely", but also sees contrast

LIME Explanation


LIME focuses heavily on positive words — less sensitive to "but"

Explainable AI Comparison

We compared three popular interpretability methods on the same input:

Example:

"I absolutely love this phone, but the battery life is short."

Rank	SHAP	LIME	Captum (Integrated Gradients)
1	exceeded	amazing	but
2	amazing	exceeded	absolutely
3	my	absolutely	is
4	product	product	i
5	This	This	life

Overlap Scores (Top-5 words):

• SHAP vs LIME: 0.57
• SHAP vs Captum: 0.29
• LIME vs Captum: 0.14

Insight: SHAP and LIME agree more on positive indicators, while Captum highlights contrastive words like "but".

Visualization (Captum - Integrated Gradients)

Green = Positive contribution | Red = Negative contribution

---

✅ ENVIRONMENT CONFIGURATION

# Your actual environment (from metadata)
{
    "colab": {
        "provenance": [],
        "gpuType": "T4",  # ← You're using T4 GPU!
        "authorship_tag": "ABX9TyNfmDrT2WuSM6k8BOIzE0ib"
    },
    "kernelspec": {
        "name": "python3",
        "display_name": "Python 3"
    },
    "accelerator": "GPU",  # ← GPU accelerated
    "language_info": {
        "name": "python"
    }
}



---

## Key Features

- Trained on **500k+ labeled tweets** (3 classes)
- Uses **ALBERT-base-v2** → lightweight & fast
- Full **XAI pipeline**:
  - SHAP (Kernel + Gradient)
  - LIME (text perturber)
  - Captum LayerIntegratedGradients
- Cross-domain testing on **IMDb** (25k reviews)
- Beautiful visualizations with `captum.attr.visualization`

---

## How to Run

### Option 1: Google Colab (Recommended)

[![Open in Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/your-link-here)

### Option 2: Local Setup

```bash
# Clone repo
git clone https://github.com/mujahidmahfuz/Sentiment-Analysis-ALBERT-XAI.git
cd Sentiment-Analysis-ALBERT-XAI

# Install dependencies
pip install -r requirements.txt

# Run the notebook
jupyter notebook "Sentiment Analysis Project.ipynb"