distilbert-restaurant-sentiment

Model Card for RinInori/distilbert-restaurant-sentiment

Model Description

This is a fine-tuned DistilBERT model for sentiment analysis of restaurant reviews. It is designed to classify reviews as either positive or negative. The model was developed as part of a project demonstrating the process of fine-tuning a DistilBERT model using the European Restaurant Reviews dataset, including data loading, model evaluation, handling class imbalance, and preventing overfitting.

Model type: DistilBERTForSequenceClassification

Language(s) (NLP): English

License: Free

Finetuned from model: distilbert-base-uncased

Model Sources [optional]

Repository: https://github.com/hennypurwadi/FineTuning_SentimentAnalysis_Restaurant_Review/tree/main

Uses

Direct Use

This model can be directly used for classifying the sentiment of restaurant reviews as either positive or negative. It is particularly suited for analyzing text from food/restaurant reviews, given its training data.

Anyone can use this model with:

from transformers import DistilBertTokenizer, DistilBertForSequenceClassification

model = DistilBertForSequenceClassification.from_pretrained("RinInori/distilbert-restaurant-sentiment")
tokenizer = DistilBertTokenizer.from_pretrained("RinInori/distilbert-restaurant-sentiment" )

Downstream Use [optional]

This model can be integrated into larger applications requiring sentiment analysis of customer feedback in the restaurant industry, such as customer service analytics platforms, review aggregation services, or market research tools.

Out-of-Scope Use

This model is specifically fine-tuned for restaurant review sentiment analysis. Using it for sentiment analysis in other domains (e.g., product reviews, movie reviews, social media posts) without further fine-tuning may lead to inaccurate results. It is also not intended for generating text or performing other NLP tasks beyond sentiment classification.

Bias, Risks, and Limitations

This model was trained on European restaurant reviews. Therefore, it may exhibit biases present in the training data, such as specific linguistic nuances, cultural references, or common sentiments expressed in European contexts. Its performance might degrade when applied to reviews from vastly different cultural or linguistic backgrounds. The model classifies sentiment into two categories (positive/negative) and does not capture nuanced or mixed sentiments. Overfitting was addressed during training using techniques like early stopping and dropout regularization, but it's always a potential risk with fine-tuned models.

Recommendations

Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.

How to Get Started with the Model

Use the code below to get started with the model.

from transformers import DistilBertTokenizer, DistilBertForSequenceClassification

model = DistilBertForSequenceClassification.from_pretrained("RinInori/distilbert-restaurant-sentiment")
tokenizer = DistilBertTokenizer.from_pretrained("RinInori/distilbert-restaurant-sentiment")

# Example usage:
text = "The food was absolutely delicious and the service was outstanding!"
inputs = tokenizer(text, return_tensors="pt")
outputs = model(**inputs)
predictions = torch.nn.functional.softmax(outputs.logits, dim=-1)

# Get predicted class and confidence
predicted_class = torch.argmax(predictions, dim=-1).item()
sentiment = "Positive" if predicted_class == 1 else "Negative"
confidence = predictions[0][predicted_class].item()

print(f"Review: {text}")
print(f"Sentiment: {sentiment}")
print(f"Confidence: {confidence:.2%}")

Training Details

Training Data

The model was fine-tuned on the European Restaurant Reviews dataset from Kaggle. This dataset contains customer reviews for various restaurants, divided into positive and negative sentiments.

Dataset URL: https://www.kaggle.com/datasets/gorororororo23/european-restaurant-reviews

Training Procedure

Preprocessing [optional]

Preprocessing steps included tokenization using DistilBertTokenizer. The dataset was split into training and validation sets. Class imbalance was addressed using class weighting, and overfitting was prevented using early stopping and dropout regularization.

Training Hyperparameters

Training regime: The model was trained using the Hugging Face Trainer API. Key techniques included class weighting to handle imbalance, early stopping to prevent overfitting, and dropout regularization.

Evaluation

Testing Data, Factors & Metrics

Testing Data: The model's performance was evaluated on a held-out test set from the European Restaurant Reviews dataset.

Factors: Sentiment classification (Positive/Negative).

Metrics: The evaluation included standard classification metrics such as accuracy, precision, recall, and F1-score. Confusion matrices were also used to visualize performance.

Results

Fine-Tuned the model with dataset made the model performance become better.

Summary

Model Examination [optional]

More Information Needed

Hardware Type: GPU, A100 Cloud Provider: Google Colab (implied by notebook)

Technical Specifications [optional]

Model Architecture and Objective

The model uses the DistilBERT architecture, which is a distilled version of BERT, making it smaller and faster while retaining much of BERT's language capabilities. The objective was to fine-tune this architecture for binary sentiment classification (positive/negative) on restaurant reviews.

Compute Infrastructure

Training was performed in a Google Colab environment with GPU

Software

Python, PyTorch, Hugging Face Transformers library, pandas, numpy, scikit-learn, matplotlib, seaborn, nltk.

Citation

RinInori. (2024). distilbert-restaurant-sentiment. Hugging Face Hub. Retrieved from https://huggingface.co/RinInori/distilbert-restaurant-sentiment