mediaunmasked/analyzers/sentiment_analyzer.py

import logging
from typing import Dict, Any, List, Optional
from textblob import TextBlob
from transformers import pipeline
import numpy as np

logger = logging.getLogger(__name__)

class SentimentAnalyzer:
    def __init__(self, use_ai: bool = True, model_registry: Optional[Any] = None):
        """
        Initialize sentiment analyzer with both traditional and LLM-based approaches.
        
        Args:
            use_ai: Boolean indicating whether to use AI-powered analysis (True) or traditional analysis (False)
            model_registry: Optional shared model registry for better performance
        """
        self.use_ai = use_ai
        self.llm_available = False
        self.model_registry = model_registry
        self.toxicity_available = False
        
        # Traditional manipulation patterns
        self.manipulative_patterns = [
            "experts say",
            "sources claim",
            "many believe",
            "some say",
            "everyone knows",
            "clearly",
            "obviously",
            "without doubt",
            "certainly"
        ]
        
        if use_ai:
            try:
                if model_registry and model_registry.is_available:
                    # Use shared models
                    self.sentiment_pipeline = model_registry.sentiment
                    self.zero_shot = model_registry.zero_shot
                    self.toxicity_pipeline = getattr(model_registry, 'toxicity', None)
                    self.toxicity_available = self.toxicity_pipeline is not None
                    self.llm_available = True
                    logger.info("Using shared model pipelines for sentiment analysis")
                    if self.toxicity_available:
                        logger.info("Toxicity analysis enabled")
                    else:
                        logger.info("Toxicity analysis not available")
                else:
                    # Initialize own pipelines
                    self.sentiment_pipeline = pipeline(
                        "text-classification",
                        model="SamLowe/roberta-base-go_emotions",
                        device=-1,
                        batch_size=16
                    )
                    self.zero_shot = pipeline(
                        "zero-shot-classification",
                        model="facebook/bart-large-mnli",
                        device=-1,
                        batch_size=8
                    )
                    try:
                        self.toxicity_pipeline = pipeline(
                            "text-classification",
                            model="unitary/toxic-bert",
                            device=-1,
                            batch_size=16
                        )
                        self.toxicity_available = True
                        logger.info("Toxicity analysis enabled")
                    except Exception as tox_error:
                        logger.warning(f"Toxicity pipeline initialization failed: {str(tox_error)}")
                        self.toxicity_available = False
                    
                    self.llm_available = True
                    logger.info("Initialized dedicated model pipelines for sentiment analysis")
            except Exception as e:
                logger.warning(f"Failed to initialize LLM pipelines: {str(e)}")
                self.llm_available = False
        else:
            logger.info("Initializing sentiment analyzer in traditional mode")

    def _analyze_with_llm(self, text: str) -> Dict[str, Any]:
        """Perform sentiment analysis using LLM models."""
        try:
            logger.info("\n" + "="*50)
            logger.info("SENTIMENT ANALYSIS STARTED")
            logger.info("="*50)
            
            # Clean the text of formatting markers
            logger.info("Cleaning and preparing text...")
            cleaned_text = text.replace('$!/$', '').replace('##', '').replace('#', '')
            cleaned_text = '\n'.join(line for line in cleaned_text.split('\n') 
                                   if not line.startswith('[') and not line.startswith('More on'))
            
            logger.info(f"Text prepared - Length: {len(cleaned_text)} characters")
            
            # Split text into chunks of 512 tokens (approximate)
            chunks = [cleaned_text[i:i+2000] for i in range(0, len(cleaned_text), 2000)]
            logger.info(f"Split text into {len(chunks)} chunks for processing")
            
            # Initialize aggregation variables
            sentiment_scores = []
            toxicity_scores = []
            manipulation_scores = []
            flagged_phrases = []
            
            manipulation_categories = [
                "emotional manipulation",
                "fear mongering",
                "propaganda",
                "factual reporting",
                "balanced perspective"
            ]
            
            # Process each chunk
            for i, chunk in enumerate(chunks, 1):
                logger.info(f"\n{'-'*30}")
                logger.info(f"Processing chunk {i}/{len(chunks)}")
                logger.info(f"Chunk length: {len(chunk)} characters")
                
                try:
                    # Get emotion scores
                    logger.info("Analyzing emotions...")
                    emotions = self.sentiment_pipeline(chunk)
                    logger.debug(f"Raw emotion response: {emotions}")
                    
                    # Handle different response formats
                    if isinstance(emotions, list):
                        for emotion in emotions:
                            if isinstance(emotion, dict) and 'label' in emotion and 'score' in emotion:
                                sentiment_scores.append(emotion)
                                logger.info(f"Detected emotion: {emotion['label']} (score: {emotion['score']:.3f})")
                    elif isinstance(emotions, dict) and 'label' in emotions and 'score' in emotions:
                        sentiment_scores.append(emotions)
                        logger.info(f"Detected emotion: {emotions['label']} (score: {emotions['score']:.3f})")
                    
                    # Get toxicity scores if available
                    if self.toxicity_available:
                        logger.info("Analyzing toxicity...")
                        try:
                            toxicity = self.toxicity_pipeline(chunk)
                            if isinstance(toxicity, list):
                                toxicity_scores.extend(toxicity)
                            else:
                                toxicity_scores.append(toxicity)
                            logger.info(f"Toxicity analysis complete for chunk {i}")
                            logger.debug(f"Toxicity scores: {toxicity_scores[-1]}")
                        except Exception as tox_error:
                            logger.warning(f"Toxicity analysis failed for chunk {i}: {str(tox_error)}")
                    
                    # Get manipulation scores
                    logger.info("Analyzing manipulation patterns...")
                    manipulation = self.zero_shot(
                        chunk,
                        manipulation_categories,
                        multi_label=True
                    )
                    
                    if isinstance(manipulation, dict) and 'labels' in manipulation and 'scores' in manipulation:
                        chunk_scores = {
                            label: score 
                            for label, score in zip(manipulation['labels'], manipulation['scores'])
                        }
                        manipulation_scores.append(chunk_scores)
                        logger.info("Manipulation scores for chunk:")
                        for label, score in chunk_scores.items():
                            logger.info(f"  - {label}: {score:.3f}")
                    
                    # Analyze sentences for manipulation
                    logger.info("Analyzing individual sentences for manipulation...")
                    sentences = chunk.split('.')
                    for sentence in sentences:
                        if len(sentence.strip()) > 10:
                            sent_result = self.zero_shot(
                                sentence.strip(),
                                manipulation_categories,
                                multi_label=False
                            )
                            if (sent_result['labels'][0] in ["emotional manipulation", "fear mongering", "propaganda"] 
                                and sent_result['scores'][0] > 0.7):
                                logger.info(f"Found manipulative content (score: {sent_result['scores'][0]:.3f}): {sentence.strip()}")
                                flagged_phrases.append({
                                    'text': sentence.strip(),
                                    'type': sent_result['labels'][0],
                                    'score': sent_result['scores'][0]
                                })
                
                except Exception as chunk_error:
                    logger.error(f"Error processing chunk {i}: {str(chunk_error)}")
                    continue
            
            logger.info("\nAggregating final scores...")
            
            # Aggregate scores with error handling
            def aggregate_scores(scores_list, score_type: str):
                try:
                    if not scores_list:
                        logger.warning(f"No {score_type} scores to aggregate")
                        return {}
                        
                    all_scores = {}
                    for scores in scores_list:
                        if isinstance(scores, dict):
                            if 'label' in scores and 'score' in scores:
                                label = scores['label']
                                score = scores['score']
                            else:
                                # Handle direct label-score mapping
                                for label, score in scores.items():
                                    if label not in all_scores:
                                        all_scores[label] = []
                                    if isinstance(score, (int, float)):
                                        all_scores[label].append(score)
                                continue
                        else:
                            logger.warning(f"Unexpected score format in {score_type}: {scores}")
                            continue
                            
                        if isinstance(label, (str, bytes)):
                            if label not in all_scores:
                                all_scores[label] = []
                            if isinstance(score, (int, float)):
                                all_scores[label].append(score)
                                
                    return {k: float(np.mean(v)) for k, v in all_scores.items() if v}
                except Exception as agg_error:
                    logger.error(f"Error aggregating {score_type} scores: {str(agg_error)}")
                    return {}
            
            emotion_scores = aggregate_scores(sentiment_scores, "emotion")
            toxicity_scores = aggregate_scores(toxicity_scores, "toxicity") if self.toxicity_available else {}
            
            logger.info("\nFinal emotion scores:")
            for emotion, score in emotion_scores.items():
                logger.info(f"  - {emotion}: {score:.3f}")
            
            if toxicity_scores:
                logger.info("\nFinal toxicity scores:")
                for category, score in toxicity_scores.items():
                    logger.info(f"  - {category}: {score:.3f}")
            
            # Aggregate manipulation scores
            manipulation_agg = {
                category: float(np.mean([
                    scores.get(category, 0) 
                    for scores in manipulation_scores
                ])) 
                for category in manipulation_categories
                if manipulation_scores
            }
            
            logger.info("\nFinal manipulation scores:")
            for category, score in manipulation_agg.items():
                logger.info(f"  - {category}: {score:.3f}")
            
            # Calculate manipulation score based on multiple factors
            manipulation_indicators = {
                'emotional manipulation': 0.4,
                'fear mongering': 0.3,
                'propaganda': 0.3
            }
            
            if self.toxicity_available:
                manipulation_indicators.update({
                    'toxic': 0.2,
                    'severe_toxic': 0.3,
                    'threat': 0.2
                })
            
            # Combine toxicity and manipulation scores
            combined_scores = {**toxicity_scores, **manipulation_agg}
            
            # Calculate manipulation score with fallback
            if combined_scores:
                manipulation_score = min(100, sum(
                    combined_scores.get(k, 0) * weight 
                    for k, weight in manipulation_indicators.items()
                ) * 100)
            else:
                # Fallback to traditional analysis if no scores available
                manipulation_score = len(self._detect_manipulative_phrases(text)) * 10
            
            logger.info(f"\nFinal manipulation score: {manipulation_score:.1f}")
            
            # Determine overall sentiment
            positive_emotions = ['admiration', 'joy', 'amusement', 'approval']
            negative_emotions = ['disgust', 'anger', 'disappointment', 'fear']
            neutral_emotions = ['neutral', 'confusion', 'realization']
            
            pos_score = sum(emotion_scores.get(emotion, 0) for emotion in positive_emotions)
            neg_score = sum(emotion_scores.get(emotion, 0) for emotion in negative_emotions)
            neu_score = sum(emotion_scores.get(emotion, 0) for emotion in neutral_emotions)
            
            logger.info(f"\nSentiment component scores:")
            logger.info(f"  - Positive: {pos_score:.3f}")
            logger.info(f"  - Negative: {neg_score:.3f}")
            logger.info(f"  - Neutral: {neu_score:.3f}")
            
            # Determine sentiment based on highest score
            max_score = max(pos_score, neg_score, neu_score)
            if max_score == pos_score and pos_score > 0.3:
                sentiment = "Positive"
            elif max_score == neg_score and neg_score > 0.3:
                sentiment = "Negative"
            else:
                sentiment = "Neutral"
            
            logger.info(f"\nFinal sentiment determination: {sentiment}")
            
            # Sort and limit flagged phrases by manipulation score
            sorted_phrases = sorted(flagged_phrases, key=lambda x: x['score'], reverse=True)
            unique_phrases = []
            seen = set()
            for phrase in sorted_phrases:
                clean_text = phrase['text'].strip()
                if clean_text not in seen:
                    unique_phrases.append(clean_text)
                    seen.add(clean_text)
                if len(unique_phrases) >= 5:
                    break
            
            logger.info(f"\nFlagged {len(unique_phrases)} unique manipulative phrases")
            
            logger.info("\nSentiment analysis completed successfully")
            
            return {
                "sentiment": sentiment,
                "manipulation_score": round(manipulation_score, 1),
                "flagged_phrases": unique_phrases,
                "detailed_scores": {
                    "emotions": emotion_scores,
                    "manipulation": manipulation_agg,
                    "toxicity": toxicity_scores
                }
            }
            
        except Exception as e:
            logger.error(f"LLM analysis failed: {str(e)}", exc_info=True)
            return None

    def analyze(self, text: str) -> Dict[str, Any]:
        """
        Analyze sentiment using LLM with fallback to traditional methods.
        
        Args:
            text: The text to analyze
            
        Returns:
            Dict containing sentiment analysis results
        """
        try:
            # Try LLM analysis if enabled and available
            if self.use_ai and self.llm_available:
                llm_result = self._analyze_with_llm(text)
                if llm_result:
                    return llm_result
            
            # Use traditional analysis
            logger.info("Using traditional sentiment analysis")
            blob = TextBlob(text)
            sentiment_score = blob.sentiment.polarity
            
            manipulative_phrases = self._detect_manipulative_phrases(text)
            manipulation_score = len(manipulative_phrases) * 10
            
            if sentiment_score > 0.2:
                sentiment = "Positive"
            elif sentiment_score < -0.2:
                sentiment = "Negative"
            else:
                sentiment = "Neutral"
            
            return {
                "sentiment": sentiment,
                "manipulation_score": min(manipulation_score, 100),
                "flagged_phrases": manipulative_phrases[:5]  # Limit to top 5 phrases
            }
            
        except Exception as e:
            logger.error(f"Error in sentiment analysis: {str(e)}")
            return {
                "sentiment": "Error",
                "manipulation_score": 0,
                "flagged_phrases": []
            }

    def _detect_manipulative_phrases(self, text: str) -> List[str]:
        """Detect potentially manipulative phrases."""
        found_phrases = []
        text_lower = text.lower()
        
        for pattern in self.manipulative_patterns:
            if pattern in text_lower:
                start = text_lower.find(pattern)
                context = text[max(0, start-20):min(len(text), start+len(pattern)+20)]
                found_phrases.append(context.strip())
        
        return found_phrases