ullahhafeez321/bert-sentiment-classifier

BERT Sentiment Classifier

A fine-tuned BERT Base Uncased model for English movie review sentiment classification.

This model classifies movie reviews into two sentiment categories:

• POSITIVE (1): The review expresses a positive opinion.
• NEGATIVE (0): The review expresses a negative opinion.

The model was fine-tuned using the IMDb Movie Reviews dataset and evaluated on a separate test set containing 2,000 examples.

Model Details

Property	Description
Model architecture	BERT Base Uncased
Base model	google-bert/bert-base-uncased
Task	Binary Sentiment Classification
Number of classes	2
Positive label	1
Negative label	0
Maximum sequence length	192 tokens
Framework	PyTorch
Library	Hugging Face Transformers

Dataset

The model was fine-tuned using the IMDb Movie Reviews Dataset.

The notebook uses the following dataset split:

Dataset	Number of samples
Training	8,000
Validation	1,000
Testing	2,000
Total	11,000

The dataset contains English movie reviews with binary sentiment labels:

• 0: Negative
• 1: Positive

The training set is used to fine-tune the model, the validation set is used to monitor training and evaluate checkpoints, and the test set is used for final performance evaluation.

Training Configuration

The model was fine-tuned for three epochs using the following configuration:

Parameter	Value
Base model	BERT Base Uncased
Epochs	3
Batch size	16
Learning rate	2e-5
Weight decay	0.01
Warmup steps	100
Maximum sequence length	192
Random seed	42
Best checkpoint metric	Validation accuracy
Training framework	PyTorch

A two-class classification head was added to the pretrained BERT encoder and fine-tuned for sentiment prediction.

Training Results

The following table shows the training progress across all three epochs.

Epoch	Training Loss	Validation Loss	Accuracy	Precision	Recall	F1-score
1	0.376433	0.259453	0.896000	0.919321	0.867735	0.892784
2	0.166479	0.358983	0.884000	0.894845	0.869739	0.882114
3	0.134408	0.445110	0.894000	0.887574	0.901804	0.894632

Training loss decreased throughout training. However, validation loss increased after the first epoch. This may indicate overfitting, even though the validation accuracy remained relatively stable.

Final Validation Results

The final validation evaluation produced the following results:

Metric	Result
Validation Loss	0.259079
Accuracy	89.70%
Precision	91.95%
Recall	86.97%
F1-score	89.39%

Final Test Results

The final evaluation was performed on a separate test set containing 2,000 movie reviews, with 1,000 Negative and 1,000 Positive examples.

Metric	Result
Test Loss	0.259910
Accuracy	88.80%
Precision	91.54%
Recall	85.50%
F1-score	88.42%

The model correctly classified 1,776 out of 2,000 test examples.

Classification Report

Class	Precision	Recall	F1-score	Support
Negative	0.8640	0.9210	0.8916	1,000
Positive	0.9154	0.8550	0.8842	1,000
Macro Average	0.8897	0.8880	0.8879	2,000
Weighted Average	0.8897	0.8880	0.8879	2,000

Confusion Matrix

The confusion matrix shows the actual and predicted sentiment labels.

Actual / Predicted	Predicted Negative	Predicted Positive
Actual Negative	921	79
Actual Positive	145	855

Interpretation:

• 921 Negative reviews were correctly classified.
• 855 Positive reviews were correctly classified.
• 79 Negative reviews were incorrectly classified as Positive.
• 145 Positive reviews were incorrectly classified as Negative.

The model achieved higher recall for Negative reviews (92.10%) than Positive reviews (85.50%).

Model Architecture

This project uses an encoder-only transformer architecture.

BERT reads the input text and creates contextual representations of its tokens. The classification head uses these representations to predict the sentiment of the complete review.

The model follows this general workflow:

1. Receive an English movie review.
2. Tokenize the input text.
3. Pass the tokens through the pretrained BERT encoder.
4. Generate classification logits.
5. Apply softmax to obtain class probabilities.
6. Return the predicted sentiment and confidence score.

Usage

You can load this model directly using Hugging Face Transformers.

Install Dependencies

pip install transformers torch

Load the Model

import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification

MODEL_NAME = "ullahhafeez321/bert-sentiment-classifier"

tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)

model = AutoModelForSequenceClassification.from_pretrained(MODEL_NAME)

model.eval()

Make a Prediction

def predict_sentiment(text):

    inputs = tokenizer(
        text,
        return_tensors="pt",
        truncation=True,
        max_length=192
    )

    with torch.inference_mode():
        outputs = model(**inputs)

        probabilities = torch.softmax(
            outputs.logits,
            dim=-1
        )[0]

    predicted_id = int(probabilities.argmax())

    sentiment = model.config.id2label[predicted_id]

    confidence = float(probabilities[predicted_id]) * 100

    return {
        "text": text,
        "sentiment": sentiment,
        "confidence": round(confidence, 2)
    }


result = predict_sentiment(
    "This movie was excellent. I really enjoyed it!"
)

print(result)

Example Predictions

Input Review	Expected Sentiment
This movie was absolutely wonderful!	POSITIVE
This movie was terrible and boring.	NEGATIVE
I loved every moment of this film.	POSITIVE
I hated this movie. It was a waste of time.	NEGATIVE

These examples are illustrative and are not guaranteed predictions.

Applications

This model can be integrated into:

• Sentiment analysis websites
• Movie review classification applications
• Text classification projects
• Python applications
• Streamlit applications
• Educational NLP projects

Limitations

• The model was fine-tuned on English IMDb movie reviews.
• It is primarily intended for movie-review sentiment classification.
• It may not perform equally well on other domains, such as product reviews, news, or social media.
• Sarcasm, irony, ambiguous language, and mixed sentiment can be difficult to classify.
• Confidence scores are model probabilities, not guarantees of correctness.
• Validation loss increased after the first epoch, suggesting possible overfitting.

Technologies Used

• Python
• PyTorch
• Hugging Face Transformers
• Hugging Face Datasets
• Scikit-learn
• IMDb Movie Reviews Dataset
• Google BERT Base Uncased

Training Notebook

The training notebook includes:

1. Library installation and environment setup.
2. IMDb dataset loading and splitting.
3. Pretrained BERT tokenizer and model initialization.
4. Dataset tokenization.
5. Model fine-tuning.
6. Validation and test evaluation.
7. Classification report and confusion matrix.
8. Custom sentiment prediction.
9. Saving and uploading the model to Hugging Face.

Project Integration

This model is also used as the sentiment classification component in my NLP Analysis Studio application, alongside a separate decoder-only Qwen model for Named Entity Recognition.

• Live Application: https://sentimentclassifynamedentitymodels.streamlit.app/
• GitHub Repository: https://github.com/ullahhafeez321/SentimentClassify_NamedEntityModels
• BERT Model: https://huggingface.co/ullahhafeez321/bert-sentiment-classifier
• Qwen NER Model: https://huggingface.co/ullahhafeez321/Qwen-NER-Decoder

License

This project is released under the MIT License.