Harieeeeviji/sentiment-analysis

MindCare AI: Emotion & Sentiment Analysis Models (Trained from Scratch)

This repository hosts serialized machine learning models trained strictly from scratch without pretrained transformer weights, developed as part of the MindCare AI framework.

📊 Benchmark Results

1. 3-Class Sentiment (Positive, Negative, Neutral)

• Dataset: GoEmotions (54,258 authentic samples)
• Best Model: Logistic Regression / Calibrated SVM (LinearSVC)
• Held-out Test Accuracy: 68.80%
• Weighted F1 Score: 0.6909

2. 10-Class Fine-Grained Emotion Recognition

• Classes: Happiness, Sadness, Anger, Fear, Surprise, Disgust, Neutral, Excitement, Frustration, Gratitude
• Best Model: Custom Bidirectional LSTM with Self-Attention Context Pooling
• Held-out Test Accuracy: 56.99%
• Weighted F1 Score: 0.5610

📦 Model Files Included

• sentiment_lr.joblib - Calibrated Logistic Regression for 3-class sentiment
• sentiment_svm.joblib - Calibrated Linear Support Vector Machine
• sentiment_nb.joblib - Multinomial Naive Bayes
• sentiment_rf.joblib - Random Forest Classifier
• sentiment_vectorizer.joblib - TF-IDF n-gram vectorizer (ngram_range=(1,2), max_features=12,000)
• sentiment_mlp.pt - PyTorch Multi-Layer Perceptron
• sentiment_lstm.pt - PyTorch Bidirectional LSTM with Attention
• emotion_lr.joblib / emotion_svm.joblib / emotion_lstm.pt - 10-class emotion models
• vocab.json - Tokenizer vocabulary
• sentiment_metadata.json / emotion_metadata.json - Validation & test performance metrics

🚀 Quick Usage (Python)

import joblib

# Load TF-IDF vectorizer and trained Logistic Regression model
vectorizer = joblib.load("sentiment_vectorizer.joblib")
model = joblib.load("sentiment_lr.joblib")

text = "I am grateful for all the support and kindness."
X = vectorizer.transform([text])
prediction = model.predict(X)[0]
probs = model.predict_proba(X)[0]

print(f"Sentiment: {prediction}")
print(f"Probabilities: {dict(zip(model.classes_, probs))}")

⚖️ Ethical Boundary & Non-Medical Disclaimer

This model identifies linguistic patterns and sentiment correlations in text. It is not a psychiatric diagnostic instrument or medical device.