| --- |
| language: en |
| license: apache-2.0 |
| library_name: transformers |
| tags: |
| - sentiment-analysis |
| - roberta |
| - amazon-reviews |
| - e-commerce |
| datasets: |
| - amazon_fine_food_reviews |
| metrics: |
| - accuracy |
| - f1 |
| pipeline_tag: text-classification |
| --- |
| |
| # Model Card: Amazon Sentiment RoBERTa Base |
|
|
| ## Model Description |
| This model is a fine-tuned version of **RoBERTa-base** specifically optimized for sentiment analysis of customer reviews. It was trained on a balanced subset of the Amazon Fine Food Reviews dataset to classify text into three distinct categories: **Negative**, **Neutral**, and **Positive**. |
|
|
| - **Model Type:** Transformer-based Text Classification |
| - **Language:** English |
| - **Base Model:** `roberta-base` |
|
|
| ## Intended Use |
| - **Primary Use Case:** Real-time sentiment tracking for e-commerce platforms. |
| - **Scope:** Analyzing short to medium-length customer feedback and product reviews. |
| - **Out-of-Scope:** Not recommended for legal documents, medical advice, or languages other than English. |
|
|
| ## Training Data & Methodology |
| ### Dataset |
| - **Source:** Amazon Fine Food Reviews (Kaggle). |
| - **Preprocessing:** - Removal of duplicates and HTML tags. |
| - POS-tag-based Lemmatization for linguistic normalization. |
| - Undersampling to 15,000 samples (5,000 per class) to handle class imbalance. |
| - **Labels:** - `0`: Negative (1-2 stars) |
| - `1`: Neutral (3 stars) |
| - `2`: Positive (4-5 stars) |
|
|
| ### Hyperparameters |
| - **Learning Rate:** 2e-5 |
| - **Batch Size:** 16 |
| - **Epochs:** 2 |
| - **Weight Decay:** 0.01 |
| - **Max Sequence Length:** 128 tokens |
|
|
| ## Performance Metrics |
| The model was evaluated on a held-out test set (20% of the balanced data): |
|
|
| | Metric | Value | |
| | :--- | :--- | |
| | **Accuracy** | 78.0% | |
| | **Weighted F1-Score** | 0.78 | |
| | **Precision (Positive)** | 0.83 | |
| | **Recall (Positive)** | 0.89 | |
|
|
| ### Key Strengths |
| - **Contextual Understanding:** Successfully handles complex structures, such as negation and sarcasm (e.g., "Don't listen to the haters, this is great!"). |
| - **Robustness:** Significantly outperforms traditional TF-IDF and DistilBERT baselines in identifying ambiguous "Neutral" reviews. |
|
|
| ## Limitations & Bias |
| - **Neutral Class:** Still remains the most frequent source of misclassification due to the inherent subjectivity of 3-star ratings. |
| - **Domain Specificity:** Performance may vary when applied to domains outside of food and beverages (e.g., electronics or fashion). |
| - **Sarcasm:** While improved, extremely subtle sarcasm may still lead to errors. |
|
|
| ## How to Use |
| ```python |
| from transformers import pipeline |
| |
| # Load the model directly from the Hub |
| model_path = "mlklt3/amazon-sentiment-roberta-base" |
| sentiment_pipeline = pipeline("sentiment-analysis", model=model_path) |
| |
| # Example usage |
| text = "The product was okay, but I expected much better flavor for this price." |
| result = sentiment_pipeline(text) |
| print(result) |
| ``` |
| ## Citation |
| If you use this model in your research or project, please credit the Amazon Fine Food Reviews dataset and the Hugging Face Transformers library. |
