Amey9766/BERT-hospitality-review-classifier

BERT Hospitality Review Classifier

Model Description

BERT Hospitality Review Classifier is a fine-tuned bert-base-uncased model designed to classify hotel reviews into positive or negative sentiment.
The model is trained on a large-scale hospitality-specific dataset to capture nuanced language patterns commonly found in hotel guest feedback.

• Base model: bert-base-uncased
• Task: Binary Sentiment Classification
• Domain: Hospitality / Hotel Reviews
• Language: English
• Labels: negative, positive

---

Dataset

The model is trained and evaluated using the kaenova/hotel-sentiment dataset from Hugging Face.

• Total samples: 742,612 reviews
• Training set: 668,350
• Validation set: 37,131
• Test set: 37,131
• Classes: Binary (Negative, Positive)
• Label distribution: Approximately balanced

Dataset link: https://huggingface.co/datasets/kaenova/hotel-sentiment

---

Training Details

• Architecture: BERT with sequence classification head
• Max sequence length: 256 tokens
• Optimizer: AdamW
• Loss function: Cross-entropy
• Hardware: NVIDIA T4 GPU (Google Colab)
• Framework: Hugging Face Transformers

The classification head was newly initialized and trained during fine-tuning, which is expected behavior when adapting BERT to a downstream task.

---

Evaluation Results

The model was evaluated on a held-out test set of 37,131 reviews.

Overall Metrics

Metric	Score
Accuracy	0.9690
Precision	0.9772
Recall	0.9668
F1-score	0.9720
ROC-AUC	0.9923

---

Confusion Matrix

• True Positives: 19,946
• True Negatives: 16,034
• False Positives: 466
• False Negatives: 685

---

ROC Curve

The ROC curve shows strong class separability, with an AUC of 0.9923, indicating excellent discriminative performance.

---

Error Analysis (Qualitative)

Common False Positives

• Reviews with strong positive wording but subtle complaints
• Location-focused praise masking dissatisfaction in services

Common False Negatives

• Reviews with mild criticism framed in neutral language
• Operational complaints without strong emotional markers

This behavior aligns with real-world hospitality feedback patterns.

---

Intended Use

This model is suitable for:

• Hotel review sentiment analysis
• Guest feedback monitoring systems
• Hospitality analytics dashboards
• Customer experience automation
• AI-powered hotel assistants

---

Limitations

• Binary sentiment only (no neutral or aspect-based sentiment)
• Trained on English-language reviews
• Performance may degrade on highly sarcastic or ambiguous text

---

Ethical Considerations

• The model reflects sentiment trends in publicly available hotel reviews
• Predictions should support, not replace, human decision-making
• Not intended for high-stakes automated judgments

---

Author

Amey Tillu Hospitality Data Analyst & AI/ML Hobbyist

How to Use

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

tokenizer = AutoTokenizer.from_pretrained("Amey9766/BERT-hospitality-review-classifier")
model = AutoModelForSequenceClassification.from_pretrained("Amey9766/BERT-hospitality-review-classifier")

text = "The hotel was clean and staff were very helpful."
inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True)

with torch.no_grad():
    outputs = model(**inputs)
    prediction = torch.argmax(outputs.logits, dim=1).item()

label_map = {0: "negative", 1: "positive"}
print(label_map[prediction])