dhyann2815/roberta-empathy-classifier

RoBERTa Empathy Classifier

This is a binary text classification model trained to detect empathy in text. It takes a piece of text and classifies it as either 1 (empathetic) or 0 (not empathetic). The model is built on top of roberta-base using the simpletransformers library.

Model Objective

The primary goal of this model is to analyze text (like chat transcripts, forum posts, or user feedback) and identify whether the author is demonstrating empathy.

Dataset & Preprocessing

The model was trained on the AcnEmpathize_dataset.csv.

• Cleaned up raw HTML entities, URLs, and extra whitespaces from the text.
• Filtered out extremely short inputs (less than 3 words).
• Split the data into a 70% training set and 30% testing set (stratified based on the target labels).
• Applied class weighting ([1.0, 3.0]) to aggressively counteract the heavy dataset imbalance.

Training Configuration

The model was fine-tuned using simpletransformers with the following hyperparameters to optimize minority class detection:

• Base Model: FacebookAI/roberta-base
• Epochs: 5
• Learning Rate: 3e-5
• Max Sequence Length: 256
• Class Weights: [1.0, 3.0]
• Train & Eval Batch Size: 16
• Seed: 42

Evaluation Metrics

Based on our 30% test split, the model achieved the following results after hyperparameter tuning:

• Accuracy: 75.90%
• Evaluation Loss: 0.6145
• F1 Score (Macro): 0.6393
• F1 Score (Positive Class): 0.4314
• AUROC: 0.7396
• AUPRC: 0.4716

Confusion Matrix Summary:

• True Negatives (TN): 2446
• False Negatives (FN): 558
• True Positives (TP): 335
• False Positives (FP): 325

(Note: The model successfully learned to identify the minority empathetic class, dramatically improving True Positives (from 0 to 335) and AUROC while maintaining strong overall accuracy.)

Usage

You can load and use this model directly with the simpletransformers library:

from simpletransformers.classification import ClassificationModel
import numpy as np

# Initialize the model
model = ClassificationModel(
    "roberta", 
    "dhyann2815/roberta-empathy-classifier",
    use_cuda=False # Set to True if you have a GPU
)

# Define prediction function with confidence score
def predict_empathy(text):
    predictions, raw_outputs = model.predict([text])
    logits = raw_outputs[0]
    probabilities = np.exp(logits) / np.sum(np.exp(logits))
    confidence = np.max(probabilities) * 100 
    
    status = "empathetic" if predictions[0] == 1 else "not empathetic"
    return status, round(confidence, 2)

# Run a prediction
status, confidence = predict_empathy("I'm so sorry you're going through this, that sounds really difficult.")
print(f"Prediction: {status} (Confidence: {confidence}%)")