File size: 17,711 Bytes
876b12f
a2624a3
5c3b4a6
876b12f
55cdb25
 
876b12f
 
 
 
a2624a3
5c3b4a6
 
876b12f
5c3b4a6
 
a2624a3
5c3b4a6
 
 
a2624a3
5c3b4a6
 
 
a2624a3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5c3b4a6
 
a2624a3
5c3b4a6
 
 
 
 
 
876b12f
 
 
 
 
 
 
 
5c3b4a6
876b12f
 
a2624a3
 
876b12f
a2624a3
 
 
 
 
876b12f
a2624a3
 
876b12f
 
 
5c3b4a6
 
 
1360e33
 
 
 
 
 
5c3b4a6
 
 
 
 
 
 
1360e33
 
a2624a3
 
 
1360e33
a2624a3
 
 
 
 
 
5c3b4a6
 
 
 
 
 
 
 
 
 
1360e33
5c3b4a6
 
 
 
 
 
 
 
 
 
1360e33
5c3b4a6
a2624a3
 
1360e33
5c3b4a6
a2624a3
 
 
1360e33
a2624a3
 
 
 
 
 
5c3b4a6
a2624a3
 
 
 
 
1360e33
a2624a3
 
 
 
 
 
 
 
 
 
 
 
 
 
1360e33
 
 
a2624a3
 
 
 
 
 
1360e33
 
 
 
 
 
 
 
 
 
a2624a3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1360e33
a2624a3
 
 
5c3b4a6
a2624a3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1360e33
a2624a3
 
 
 
 
 
 
 
1360e33
a2624a3
 
 
 
5c3b4a6
 
 
 
 
 
 
a2624a3
 
 
 
 
 
 
 
 
5c3b4a6
1360e33
 
5c3b4a6
 
a2624a3
5c3b4a6
 
 
 
 
 
 
 
 
1360e33
5c3b4a6
a2624a3
 
 
 
5c3b4a6
55cdb25
5c3b4a6
 
 
 
 
 
 
 
55cdb25
5c3b4a6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
876b12f
5c3b4a6
 
 
 
 
 
876b12f
 
5c3b4a6
876b12f
 
 
 
 
 
 
 
 
5c3b4a6
876b12f
 
5c3b4a6
 
 
 
876b12f
5c3b4a6
876b12f
 
5c3b4a6
 
 
 
 
 
 
 
a2624a3
5c3b4a6
 
a2624a3
 
 
 
 
 
 
876b12f
a2624a3
 
 
 
 
 
876b12f
5c3b4a6
 
 
 
876b12f
5c3b4a6
 
 
876b12f
 
5c3b4a6
876b12f
 
5c3b4a6
876b12f
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
import logging
from typing import Dict, Any, List, Optional
from transformers import pipeline, AutoTokenizer
import numpy as np
import nltk
from nltk.tokenize import sent_tokenize

logger = logging.getLogger(__name__)

class HeadlineAnalyzer:
    def __init__(self, use_ai: bool = True, model_registry: Optional[Any] = None):
        """
        Initialize the analyzers for headline analysis.
        
        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
        
        if use_ai:
            try:
                if model_registry and model_registry.is_available:
                    # Use shared models
                    self.nli_pipeline = model_registry.nli
                    self.zero_shot = model_registry.zero_shot
                    self.tokenizer = AutoTokenizer.from_pretrained("roberta-large-mnli")
                    self.llm_available = True
                    logger.info("Using shared model pipelines for headline analysis")
                else:
                    # Initialize own pipelines
                    self.nli_pipeline = pipeline(
                        "text-classification", 
                        model="roberta-large-mnli",
                        batch_size=16
                    )
                    self.zero_shot = pipeline(
                        "zero-shot-classification",
                        model="facebook/bart-large-mnli",
                        device=-1,
                        batch_size=8
                    )
                    self.tokenizer = AutoTokenizer.from_pretrained("roberta-large-mnli")
                    self.llm_available = True
                    logger.info("Initialized dedicated model pipelines for headline analysis")
                
                self.max_length = 512
                
            except Exception as e:
                logger.warning(f"Failed to initialize LLM pipelines: {str(e)}")
                self.llm_available = False
        else:
            logger.info("Initializing headline analyzer in traditional mode")

    def _split_content(self, headline: str, content: str) -> List[str]:
        """Split content into sections that fit within token limit."""
        content_words = content.split()
        sections = []
        current_section = []
        
        # Account for headline and [SEP] token in the max length
        headline_tokens = len(self.tokenizer.encode(headline))
        sep_tokens = len(self.tokenizer.encode("[SEP]")) - 2
        max_content_tokens = self.max_length - headline_tokens - sep_tokens
        
        # Process words into sections with 4000 character chunks
        current_text = ""
        for word in content_words:
            if len(current_text) + len(word) + 1 <= 4000:
                current_text += " " + word
            else:
                sections.append(current_text.strip())
                current_text = word
        
        if current_text:
            sections.append(current_text.strip())
        
        return sections

    def _analyze_section(self, headline: str, section: str) -> Dict[str, Any]:
        """Analyze a single section for headline accuracy and sensationalism."""
        try:
            logger.info("\n" + "-"*30)
            logger.info("ANALYZING SECTION")
            logger.info("-"*30)
            logger.info(f"Headline: {headline}")
            logger.info(f"Section length: {len(section)} characters")
            
            # Download NLTK data if needed
            try:
                nltk.data.find('tokenizers/punkt')
            except LookupError:
                nltk.download('punkt')
            
            sentences = sent_tokenize(section)
            logger.info(f"Found {len(sentences)} sentences in section")
            
            if not sentences:
                logger.warning("No sentences found in section")
                return {
                    "accuracy_score": 50.0,
                    "flagged_phrases": [],
                    "detailed_scores": {
                        "nli": {"ENTAILMENT": 0.0, "CONTRADICTION": 0.0, "NEUTRAL": 1.0},
                        "sensationalism": {"factual reporting": 0.5, "accurate headline": 0.5}
                    }
                }
            
            # Categories for sensationalism check
            sensationalism_categories = [
                "clickbait",
                "sensationalized",
                "misleading",
                "factual reporting",
                "accurate headline"
            ]
            
            logger.info("Checking headline for sensationalism...")
            sensationalism_result = self.zero_shot(
                headline,
                sensationalism_categories,
                multi_label=True
            )
            
            sensationalism_scores = {
                label: score 
                for label, score in zip(sensationalism_result['labels'], sensationalism_result['scores'])
            }
            logger.info(f"Sensationalism scores: {sensationalism_scores}")
            
            # Filter relevant sentences (longer than 20 chars)
            relevant_sentences = [s.strip() for s in sentences if len(s.strip()) > 20]
            logger.info(f"Found {len(relevant_sentences)} relevant sentences after filtering")
            
            if not relevant_sentences:
                logger.warning("No relevant sentences found in section")
                return {
                    "accuracy_score": 50.0,
                    "flagged_phrases": [],
                    "detailed_scores": {
                        "nli": {"ENTAILMENT": 0.0, "CONTRADICTION": 0.0, "NEUTRAL": 1.0},
                        "sensationalism": sensationalism_scores
                    }
                }
            
            # Process sentences in batches for contradiction/support
            nli_scores = []
            flagged_phrases = []
            batch_size = 8
            
            logger.info("Processing sentences for contradictions...")
            for i in range(0, len(relevant_sentences), batch_size):
                batch = relevant_sentences[i:i+batch_size]
                batch_inputs = [f"{headline} [SEP] {sentence}" for sentence in batch]
                
                try:
                    # Get NLI scores for batch
                    batch_results = self.nli_pipeline(batch_inputs, top_k=None)
                    if not isinstance(batch_results, list):
                        batch_results = [batch_results]
                    
                    for sentence, result in zip(batch, batch_results):
                        scores = {item['label']: item['score'] for item in result}
                        nli_scores.append(scores)
                        
                        # Flag contradictory content with lower threshold
                        if scores.get('CONTRADICTION', 0) > 0.3:  # Lowered threshold
                            logger.info(f"Found contradictory sentence (score: {scores['CONTRADICTION']:.2f}): {sentence}")
                            flagged_phrases.append({
                                'text': sentence,
                                'type': 'Contradiction',
                                'score': scores['CONTRADICTION'],
                                'highlight': f"[CONTRADICTION] (Score: {round(scores['CONTRADICTION'] * 100, 1)}%) \"{sentence}\""
                            })
                        
                        # Flag highly sensationalized content
                        if sensationalism_scores.get('sensationalized', 0) > 0.6 or sensationalism_scores.get('clickbait', 0) > 0.6:
                            logger.info(f"Found sensationalized content: {sentence}")
                            flagged_phrases.append({
                                'text': sentence,
                                'type': 'Sensationalized',
                                'score': max(sensationalism_scores.get('sensationalized', 0), sensationalism_scores.get('clickbait', 0)),
                                'highlight': f"[SENSATIONALIZED] \"{sentence}\""
                            })
                            
                except Exception as batch_error:
                    logger.warning(f"Batch processing error: {str(batch_error)}")
                    continue
            
            # Calculate aggregate scores with validation
            if not nli_scores:
                logger.warning("No NLI scores available")
                avg_scores = {"ENTAILMENT": 0.0, "CONTRADICTION": 0.0, "NEUTRAL": 1.0}
            else:
                try:
                    avg_scores = {
                        label: float(np.mean([
                            score.get(label, 0.0) 
                            for score in nli_scores
                        ])) 
                        for label in ['ENTAILMENT', 'CONTRADICTION', 'NEUTRAL']
                    }
                    logger.info(f"Average NLI scores: {avg_scores}")
                except Exception as agg_error:
                    logger.error(f"Error aggregating NLI scores: {str(agg_error)}")
                    avg_scores = {"ENTAILMENT": 0.0, "CONTRADICTION": 0.0, "NEUTRAL": 1.0}
            
            # Calculate headline accuracy score with validation
            try:
                accuracy_components = {
                    'entailment': avg_scores.get('ENTAILMENT', 0.0) * 0.4,
                    'non_contradiction': (1 - avg_scores.get('CONTRADICTION', 0.0)) * 0.3,
                    'non_sensational': (
                        sensationalism_scores.get('factual reporting', 0.0) +
                        sensationalism_scores.get('accurate headline', 0.0)
                    ) * 0.15,
                    'non_clickbait': (
                        1 - sensationalism_scores.get('clickbait', 0.0) -
                        sensationalism_scores.get('sensationalized', 0.0)
                    ) * 0.15
                }
                
                logger.info(f"Accuracy components: {accuracy_components}")
                accuracy_score = sum(accuracy_components.values()) * 100
                
                # Validate final score
                if np.isnan(accuracy_score) or not np.isfinite(accuracy_score):
                    logger.warning("Invalid accuracy score calculated, using default")
                    accuracy_score = 50.0
                else:
                    accuracy_score = float(accuracy_score)
                    logger.info(f"Final accuracy score: {accuracy_score:.1f}")
                    
            except Exception as score_error:
                logger.error(f"Error calculating accuracy score: {str(score_error)}")
                accuracy_score = 50.0
            
            # Sort and limit flagged phrases
            sorted_phrases = sorted(
                flagged_phrases,
                key=lambda x: x['score'],
                reverse=True
            )
            unique_phrases = []
            seen = set()
            
            for phrase in sorted_phrases:
                if phrase['text'] not in seen:
                    unique_phrases.append(phrase)
                    seen.add(phrase['text'])
                if len(unique_phrases) >= 5:
                    break
            
            logger.info(f"Final number of flagged phrases: {len(unique_phrases)}")
            
            return {
                "accuracy_score": accuracy_score,
                "flagged_phrases": unique_phrases,
                "detailed_scores": {
                    "nli": avg_scores,
                    "sensationalism": sensationalism_scores
                }
            }
            
        except Exception as e:
            logger.error(f"Section analysis failed: {str(e)}")
            return {
                "accuracy_score": 50.0,
                "flagged_phrases": [],
                "detailed_scores": {
                    "nli": {"ENTAILMENT": 0.0, "CONTRADICTION": 0.0, "NEUTRAL": 1.0},
                    "sensationalism": {}
                }
            }

    def _analyze_traditional(self, headline: str, content: str) -> Dict[str, Any]:
        """Traditional headline analysis method."""
        try:
            # Download NLTK data if needed
            try:
                nltk.data.find('tokenizers/punkt')
            except LookupError:
                nltk.download('punkt')

            # Basic metrics
            headline_words = set(headline.lower().split())
            content_words = set(content.lower().split())
            
            # Calculate word overlap
            overlap_words = headline_words.intersection(content_words)
            overlap_score = len(overlap_words) / len(headline_words) if headline_words else 0
            
            # Check for clickbait patterns
            clickbait_patterns = [
                "you won't believe",
                "shocking",
                "mind blowing",
                "amazing",
                "incredible",
                "unbelievable",
                "must see",
                "click here",
                "find out",
                "what happens next"
            ]
            
            clickbait_count = sum(1 for pattern in clickbait_patterns if pattern in headline.lower())
            clickbait_penalty = clickbait_count * 10  # 10% penalty per clickbait phrase
            
            # Calculate final score (0-100)
            base_score = overlap_score * 100
            final_score = max(0, min(100, base_score - clickbait_penalty))
            
            # Find potentially misleading phrases
            flagged_phrases = []
            sentences = sent_tokenize(content)
            
            for sentence in sentences:
                # Flag sentences that directly contradict headline words
                sentence_words = set(sentence.lower().split())
                if len(headline_words.intersection(sentence_words)) > 2:
                    flagged_phrases.append(sentence.strip())
                
                # Flag sentences with clickbait patterns
                if any(pattern in sentence.lower() for pattern in clickbait_patterns):
                    flagged_phrases.append(sentence.strip())
            
            return {
                "headline_vs_content_score": round(final_score, 1),
                "flagged_phrases": list(set(flagged_phrases))[:5]  # Limit to top 5 unique phrases
            }
            
        except Exception as e:
            logger.error(f"Traditional analysis failed: {str(e)}")
            return {
                "headline_vs_content_score": 0,
                "flagged_phrases": []
            }

    def analyze(self, headline: str, content: str) -> Dict[str, Any]:
        """Analyze how well the headline matches the content."""
        try:
            logger.info("\n" + "="*50)
            logger.info("HEADLINE ANALYSIS STARTED")
            logger.info("="*50)
            
            if not headline.strip() or not content.strip():
                logger.warning("Empty headline or content provided")
                return {
                    "headline_vs_content_score": 0,
                    "flagged_phrases": []
                }

            # Use LLM analysis if available and enabled
            if self.use_ai and self.llm_available:
                logger.info("Using LLM analysis for headline")
                # Split content if needed
                sections = self._split_content(headline, content)
                section_results = []
                
                # Analyze each section
                for section in sections:
                    result = self._analyze_section(headline, section)
                    section_results.append(result)
                
                # Aggregate results across sections
                accuracy_scores = [r['accuracy_score'] for r in section_results]
                final_score = np.mean(accuracy_scores)
                
                # Combine and deduplicate flagged phrases
                all_phrases = []
                for result in section_results:
                    if 'flagged_phrases' in result:
                        all_phrases.extend(result['flagged_phrases'])
                
                # Sort by score and get unique phrases
                sorted_phrases = sorted(all_phrases, key=lambda x: x['score'], reverse=True)
                unique_phrases = []
                seen = set()
                
                for phrase in sorted_phrases:
                    if phrase['text'] not in seen:
                        unique_phrases.append(phrase)
                        seen.add(phrase['text'])
                    if len(unique_phrases) >= 5:
                        break
                
                return {
                    "headline_vs_content_score": round(final_score, 1),
                    "flagged_phrases": unique_phrases
                }
            else:
                # Use traditional analysis
                logger.info("Using traditional headline analysis")
                return self._analyze_traditional(headline, content)
            
        except Exception as e:
            logger.error(f"Headline analysis failed: {str(e)}")
            return {
                "headline_vs_content_score": 0,
                "flagged_phrases": []
            }