radwa-f/DeBERTA-Imdb-SentimentAnalysis

DeBERTa IMDB Sentiment Analysis

Fine-tuned DeBERTa-v3-base model for binary sentiment classification on IMDB movie reviews, achieving 96% accuracy and demonstrating the superiority of transfer learning over traditional approaches.

Model Description

This model classifies movie reviews as positive or negative sentiment. It was fine-tuned as part of a comparative study demonstrating that modern transformer-based transfer learning significantly outperforms traditional deep learning approaches.

Performance Comparison:

• DeBERTa (this model): 96% accuracy
• LSTM baseline: 89% accuracy
• Improvement: +7 percentage points

Quick Start

from transformers import pipeline

classifier = pipeline(
    "sentiment-analysis",
    model="radwa-f/DeBERTA-Imdb-SentimentAnalysis"
)

result = classifier("This movie was absolutely brilliant!")
print(result)
# [{'label': 'POSITIVE', 'score': 0.9987}]

result = classifier("Worst movie I've ever seen. Complete waste of time.")
print(result)
# [{'label': 'NEGATIVE', 'score': 0.9991}]

Detailed Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

tokenizer = AutoTokenizer.from_pretrained("radwa-f/DeBERTA-Imdb-SentimentAnalysis")
model = AutoModelForSequenceClassification.from_pretrained("radwa-f/DeBERTA-Imdb-SentimentAnalysis")

text = "The cinematography was stunning and the story was captivating!"
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=512)

with torch.no_grad():
    outputs = model(**inputs)
    predictions = torch.nn.functional.softmax(outputs.logits, dim=-1)
    
predicted_class = predictions.argmax().item()
confidence = predictions[0][predicted_class].item()

labels = ["NEGATIVE", "POSITIVE"]
print(f"Prediction: {labels[predicted_class]} (confidence: {confidence:.2%})")

Training Details

Dataset

• Source: IMDB Movie Reviews Dataset
• Size: 50,000 reviews (25,000 train, 25,000 test)
• Classes: Binary (Positive/Negative)
• Balance: 50/50 split between positive and negative reviews
• Average Length: ~250 words per review

Base Model

• Architecture: DeBERTa-v3-base (microsoft/deberta-v3-base)
• Parameters: ~184M parameters
• Tokenizer: DeBERTa tokenizer with 128K vocabulary

Training Procedure

Hyperparameters:

• Optimizer: AdamW
• Learning Rate: 2e-5
• Weight Decay: 0.01
• Batch Size: 16
• Epochs: 3
• Max Sequence Length: 512 tokens
• Warmup Steps: 500
• Scheduler: Linear with warmup

Hardware:

• GPU: Google Colab A100 GPU
• Training Time: 33 min

Framework:

• PyTorch 2.0+
• Hugging Face Transformers 4.30+
• CUDA 11.8

Training Data Processing

• Text cleaning and normalization
• HTML tag removal
• Tokenization with DeBERTa tokenizer
• Padding/truncation to 512 tokens
• 80/20 train/validation split from training set

Evaluation

Test Set Performance

Metric	Score
Accuracy	96.0%
F1-Score	0.96
Precision	0.96
Recall	0.96

Comparison Study

This model was developed as part of a comparative analysis studying transfer learning vs traditional approaches:

Model	Architecture	Accuracy	Training Time
DeBERTa (this)	Transformer + Transfer Learning	96%	~2 hours
LSTM Baseline	Recurrent Neural Network	89%	~4 hours

Key Finding: Transfer learning with pre-trained transformers significantly outperforms traditional deep learning approaches trained from scratch.

Confusion Matrix Analysis

• True Positives: High precision on positive reviews
• True Negatives: High precision on negative reviews
• Misclassifications: Mostly occur with sarcastic or nuanced reviews

Intended Uses

Primary Use Cases

• Movie review sentiment analysis
• Educational demonstrations of transfer learning
• Baseline model for sentiment classification research
• Product review analysis (similar domains)

Downstream Applications

• Automated review aggregation
• Content recommendation systems
• Market research and opinion mining
• Customer feedback analysis

Limitations

Known Limitations:

• Trained specifically on movie reviews; may not generalize perfectly to other domains
• Struggles with sarcasm and highly nuanced sentiment
• Maximum input length of 512 tokens (longer reviews are truncated)
• English language only
• May reflect biases present in IMDB review dataset
• Performance may degrade on reviews from different time periods or cultures

Not Suitable For:

• Real-time streaming applications (inference time ~100ms per review)
• Non-English text
• Highly domain-specific jargon outside entertainment
• Multi-class sentiment (only binary: positive/negative)

Bias and Ethical Considerations

Potential Biases:

• IMDB dataset may over-represent certain demographics and film genres
• Model may perform differently on independent/international films
• Temporal bias: trained on historical reviews, may not capture evolving language
• May inherit biases from pre-training corpus

Responsible Use:

• Should not be used as sole basis for critical decisions
• Human review recommended for ambiguous cases
• Be aware of domain adaptation limitations
• Consider fairness implications when deploying

Model Creators

Radwa Fattouhi

• Final-year Software Engineering Student
• École Nationale des Sciences Appliquées (ENSA), El Jadida, Morocco
• Email: [email protected]
• LinkedIn: radwa-fattouhi
• GitHub: radwa-f

Amine Boktaya

• Final-year Software Engineering Student
• École Nationale des Sciences Appliquées (ENSA), El Jadida, Morocco
• Email: [email protected]
• LinkedIn: amine-boktaya
• GitHub: BoktayaAmine

Citation

If you use this model in your research, please cite:

@misc{fattouhi2025imdb,
  author = {Fattouhi, Radwa and Boktaya, Amine},
  title = {DeBERTa IMDB Sentiment Analysis: Transfer Learning vs Traditional Approaches},
  year = {2025},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/radwa-f/DeBERTA-Imdb-SentimentAnalysis}}
}

Related Work

Published Research:

• AgriAlertX: Climate-driven disaster prevention for agriculture
• Journal: SoftwareX (Elsevier)
• DOI: 10.1016/j.softx.2025.102350

Related Projects:

• Riot Detection System (DeBERTa for social media classification)
• Tweet Detoxification (BART for style transfer)

Acknowledgments

• Base model: Microsoft Research for DeBERTa-v3
• Dataset: IMDB for the movie review dataset
• Framework: Hugging Face Transformers team

---

Model card last updated: December 2025