metrics.py

# metrics.py
import nltk
from nltk.translate.bleu_score import sentence_bleu
from rouge_score import rouge_scorer
from model_loader import metrics_models

# Download required NLTK data
nltk.download('punkt')

def compute_semantic_similarity(original, paraphrased):
    """
    Compute semantic similarity between the original and paraphrased comment using Sentence-BERT.
    Returns a similarity score between 0 and 1.
    """
    try:
        sentence_bert = metrics_models.load_sentence_bert()
        embeddings = sentence_bert.encode([original, paraphrased])
        similarity = float(embeddings[0] @ embeddings[1].T)
        return round(similarity, 2)
    except Exception as e:
        print(f"Error computing semantic similarity: {str(e)}")
        return None

def compute_emotion_shift(original, paraphrased):
    """
    Compute the emotion shift between the original and paraphrased comment.
    Returns the original emotion, paraphrased emotion, and whether the shift is positive.
    """
    try:
        emotion_classifier = metrics_models.load_emotion_classifier()
        original_emotions = emotion_classifier(original)
        paraphrased_emotions = emotion_classifier(paraphrased)

        # Get the top emotion for each
        original_emotion = max(original_emotions[0], key=lambda x: x['score'])['label']
        paraphrased_emotion = max(paraphrased_emotions[0], key=lambda x: x['score'])['label']

        # Define negative and positive emotions
        negative_emotions = ['anger', 'sadness', 'fear']
        positive_emotions = ['joy', 'love', 'surprise']

        # Determine if the shift is positive
        emotion_shift_positive = (
            (original_emotion in negative_emotions and paraphrased_emotion in positive_emotions) or
            (original_emotion in negative_emotions and paraphrased_emotion not in negative_emotions)
        )

        return original_emotion, paraphrased_emotion, emotion_shift_positive
    except Exception as e:
        print(f"Error computing emotion shift: {str(e)}")
        return None, None, None

def compute_empathy_score(paraphrased):
    """
    Compute an empathy score for the paraphrased comment (placeholder).
    Returns a score between 0 and 1.
    """
    try:
        # Placeholder: Compute empathy based on word presence (e.g., "sorry", "understand")
        empathy_words = ["sorry", "understand", "care", "help", "support"]
        words = paraphrased.lower().split()
        empathy_count = sum(1 for word in words if word in empathy_words)
        score = empathy_count / len(words) if words else 0
        return round(score, 2)
    except Exception as e:
        print(f"Error computing empathy score: {str(e)}")
        return None

def compute_bleu_score(original, paraphrased):
    """
    Compute the BLEU score between the original and paraphrased comment.
    Returns a score between 0 and 1.
    """
    try:
        reference = [nltk.word_tokenize(original.lower())]
        candidate = nltk.word_tokenize(paraphrased.lower())
        score = sentence_bleu(reference, candidate, weights=(0.25, 0.25, 0.25, 0.25))
        return round(score, 2)
    except Exception as e:
        print(f"Error computing BLEU score: {str(e)}")
        return None

def compute_rouge_score(original, paraphrased):
    """
    Compute ROUGE scores (ROUGE-1, ROUGE-2, ROUGE-L) between the original and paraphrased comment.
    Returns a dictionary with ROUGE scores.
    """
    try:
        scorer = rouge_scorer.RougeScorer(['rouge1', 'rouge2', 'rougeL'], use_stemmer=True)
        scores = scorer.score(original, paraphrased)
        return {
            'rouge1': round(scores['rouge1'].fmeasure, 2),
            'rouge2': round(scores['rouge2'].fmeasure, 2),
            'rougeL': round(scores['rougeL'].fmeasure, 2)
        }
    except Exception as e:
        print(f"Error computing ROUGE scores: {str(e)}")
        return None

def compute_entailment_score(original, paraphrased):
    """
    Compute the entailment score to check factual consistency using an NLI model.
    Returns a score between 0 and 1.
    """
    try:
        nli_classifier = metrics_models.load_nli_classifier()
        result = nli_classifier(
            original,
            paraphrased,
            candidate_labels=["entailment", "contradiction", "neutral"]
        )
        entailment_score = next(score for label, score in zip(result['labels'], result['scores']) if label == "entailment")
        return round(entailment_score, 2)
    except Exception as e:
        print(f"Error computing entailment score: {str(e)}")
        return None