classifier.py

# classifier.py
import torch
from model_loader import classifier_model  # Updated import
from paraphraser import paraphrase_comment
from metrics import compute_semantic_similarity, compute_emotion_shift, compute_empathy_score, compute_bleu_score, compute_rouge_score, compute_entailment_score

def classify_toxic_comment(comment):
    """
    Classify a comment as toxic or non-toxic using the fine-tuned XLM-RoBERTa model.
    If toxic, paraphrase the comment, re-evaluate, and compute additional Stage 3 metrics.
    Returns the prediction label, confidence, color, toxicity score, bias score, paraphrased comment (if applicable), and its metrics.
    """
    if not comment.strip():
        return "Error: Please enter a comment.", None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None

    # Access the model and tokenizer
    model = classifier_model.model
    tokenizer = classifier_model.tokenizer

    # Tokenize the input comment
    inputs = tokenizer(comment, return_tensors="pt", truncation=True, padding=True, max_length=512)

    # Run inference
    with torch.no_grad():
        outputs = model(**inputs)
        logits = outputs.logits

    # Get the predicted class (0 = non-toxic, 1 = toxic)
    predicted_class = torch.argmax(logits, dim=1).item()
    label = "Toxic" if predicted_class == 1 else "Non-Toxic"
    confidence = torch.softmax(logits, dim=1)[0][predicted_class].item()
    label_color = "red" if label == "Toxic" else "green"

    # Compute Toxicity Score (approximated as the probability of the toxic class)
    toxicity_score = torch.softmax(logits, dim=1)[0][1].item()
    toxicity_score = round(toxicity_score, 2)

    # Simulate Bias Score (placeholder)
    bias_score = 0.01 if label == "Non-Toxic" else 0.15
    bias_score = round(bias_score, 2)

    # If the comment is toxic, paraphrase it and compute additional metrics
    paraphrased_comment = None
    paraphrased_prediction = None
    paraphrased_confidence = None
    paraphrased_color = None
    paraphrased_toxicity_score = None
    paraphrased_bias_score = None
    semantic_similarity = None
    original_emotion = None
    paraphrased_emotion = None
    emotion_shift_positive = None
    empathy_score = None
    bleu_score = None
    rouge_scores = None
    entailment_score = None

    if label == "Toxic":
        # Paraphrase the comment
        paraphrased_comment = paraphrase_comment(comment)

        # Re-evaluate the paraphrased comment
        paraphrased_inputs = tokenizer(paraphrased_comment, return_tensors="pt", truncation=True, padding=True, max_length=512)
        with torch.no_grad():
            paraphrased_outputs = model(**paraphrased_inputs)
            paraphrased_logits = paraphrased_outputs.logits

        paraphrased_predicted_class = torch.argmax(paraphrased_logits, dim=1).item()
        paraphrased_label = "Toxic" if paraphrased_predicted_class == 1 else "Non-Toxic"
        paraphrased_confidence = torch.softmax(paraphrased_logits, dim=1)[0][paraphrased_predicted_class].item()
        paraphrased_color = "red" if paraphrased_label == "Toxic" else "green"
        paraphrased_toxicity_score = torch.softmax(paraphrased_logits, dim=1)[0][1].item()
        paraphrased_toxicity_score = round(paraphrased_toxicity_score, 2)
        paraphrased_bias_score = 0.01 if paraphrased_label == "Non-Toxic" else 0.15  # Placeholder
        paraphrased_bias_score = round(paraphrased_bias_score, 2)

        # Compute additional Stage 3 metrics
        semantic_similarity = compute_semantic_similarity(comment, paraphrased_comment)
        original_emotion, paraphrased_emotion, emotion_shift_positive = compute_emotion_shift(comment, paraphrased_comment)
        empathy_score = compute_empathy_score(paraphrased_comment)
        bleu_score = compute_bleu_score(comment, paraphrased_comment)
        rouge_scores = compute_rouge_score(comment, paraphrased_comment)
        entailment_score = compute_entailment_score(comment, paraphrased_comment)

    return (
        f"Prediction: {label}", confidence, label_color, toxicity_score, bias_score,
        paraphrased_comment, f"Prediction: {paraphrased_label}" if paraphrased_comment else None,
        paraphrased_confidence, paraphrased_color, paraphrased_toxicity_score, paraphrased_bias_score,
        semantic_similarity, f"Original: {original_emotion}, Paraphrased: {paraphrased_emotion}, Positive Shift: {emotion_shift_positive}" if original_emotion else None,
        empathy_score, bleu_score, rouge_scores, entailment_score
    )