Spaces:
Running
Running
File size: 21,088 Bytes
5dc3509 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 |
import logging
import numpy as np
from sklearn.metrics.pairwise import cosine_similarity
from datetime import datetime, timedelta
import re
# Import the centralized NLP model handler
from utils.models import get_nlp_model
logger = logging.getLogger("misinformation_detector")
def extract_entities(text):
"""Extract named entities from text"""
if not text:
return []
try:
# Use centralized NLP model
nlp_model = get_nlp_model()
doc = nlp_model(text)
entities = [
{
"text": ent.text,
"label": ent.label_,
"start": ent.start_char,
"end": ent.end_char
}
for ent in doc.ents
]
return entities
except Exception as e:
logger.error(f"Error extracting entities: {str(e)}")
return []
def get_vector_representation(text):
"""Get vector representation of text using spaCy"""
if not text:
return None
try:
# Use centralized NLP model
nlp_model = get_nlp_model()
doc = nlp_model(text)
# Return document vector if available
if doc.has_vector:
return doc.vector
# Fallback: average of token vectors
vectors = [token.vector for token in doc if token.has_vector]
if vectors:
return np.mean(vectors, axis=0)
return None
except Exception as e:
logger.error(f"Error getting vector representation: {str(e)}")
return None
def calculate_similarity(text1, text2):
"""Calculate semantic similarity between two texts"""
if not text1 or not text2:
return 0.0
try:
vec1 = get_vector_representation(text1)
vec2 = get_vector_representation(text2)
if vec1 is None or vec2 is None:
return 0.0
# Reshape vectors for cosine_similarity
vec1 = vec1.reshape(1, -1)
vec2 = vec2.reshape(1, -1)
# Calculate cosine similarity
similarity = cosine_similarity(vec1, vec2)[0][0]
return float(similarity)
except Exception as e:
logger.error(f"Error calculating similarity: {str(e)}")
return 0.0
def extract_date_from_evidence(evidence_text):
"""Extract date from evidence text"""
if not evidence_text:
return None
try:
# Look for date patterns in text
date_patterns = [
r'Date: (\d{4}-\d{2}-\d{2})', # ISO format
r'published.*?(\d{1,2}[/-]\d{1,2}[/-]\d{2,4})', # published on MM/DD/YYYY
r'(\d{1,2}\s+(?:January|February|March|April|May|June|July|August|September|October|November|December)\s+\d{4})', # DD Month YYYY
r'(?:January|February|March|April|May|June|July|August|September|October|November|December)\s+\d{1,2},?\s+\d{4}' # Month DD, YYYY
]
for pattern in date_patterns:
match = re.search(pattern, evidence_text)
if match:
date_str = match.group(1)
# Parse date string based on format
try:
if '-' in date_str:
return datetime.strptime(date_str, '%Y-%m-%d')
elif '/' in date_str or '-' in date_str:
formats = ['%m/%d/%Y', '%d/%m/%Y', '%m-%d-%Y', '%d-%m-%Y']
for fmt in formats:
try:
return datetime.strptime(date_str, fmt)
except ValueError:
continue
else:
# Try different month formats
formats = ['%d %B %Y', '%B %d, %Y', '%B %d %Y']
for fmt in formats:
try:
return datetime.strptime(date_str, fmt)
except ValueError:
continue
except Exception:
pass
return None
except Exception as e:
logger.error(f"Error extracting date from evidence: {str(e)}")
return None
def is_temporally_relevant(evidence_text, claim_text, max_days_old=30):
"""Check if evidence is temporally relevant to the claim"""
# Check if claim seems to require recent evidence
temporal_terms = ["today", "now", "current", "currently", "recent", "recently", "latest", "just", "this week", "this month", "this year"]
requires_recent = any(term in claim_text.lower() for term in temporal_terms)
# If claim doesn't specify temporality, consider evidence relevant
if not requires_recent:
return True
# Extract date from evidence
date = extract_date_from_evidence(evidence_text)
if not date:
return True # If we can't determine date, assume it's relevant
# Check if evidence is recent enough
cutoff = datetime.now() - timedelta(days=max_days_old)
return date >= cutoff
def has_authority_signal(evidence_text):
"""Check if evidence contains authority signals"""
authority_signals = {
"scientific_consensus": ["consensus", "scientists agree", "research shows", "studies confirm", "experts agree"],
"fact_check": ["fact check", "rated false", "rated true", "debunked", "confirmed", "verification"],
"high_authority": ["nasa", "world health organization", "who", "cdc", "national academy",
"oxford", "harvard", "stanford", "mit", "cambridge", "yale",
"princeton", "government", "official", "authorities", "minister",
"ministry", "department", "administration", "university", "professor"]
}
evidence_lower = evidence_text.lower()
authority_type = None
authority_score = 1.0
for signal_type, phrases in authority_signals.items():
if any(phrase in evidence_lower for phrase in phrases):
if signal_type == "scientific_consensus":
authority_score = 1.8
authority_type = "scientific_consensus"
elif signal_type == "fact_check":
authority_score = 1.5
authority_type = "fact_check"
elif signal_type == "high_authority":
authority_score = 1.3
authority_type = "high_authority"
break
return authority_score, authority_type
def analyze_evidence_relevance(claim, evidence_list, source_credibility=None):
"""
Analyze evidence relevance to claim using semantic similarity with improved handling
for claims requiring strong evidence
Args:
claim (str): The claim being verified
evidence_list (list): List of evidence items
source_credibility (dict): Dictionary mapping source domains to credibility scores
Returns:
list: Sorted list of evidence items with relevance scores
"""
if not evidence_list:
return []
# Ensure evidence_list is a list of strings
if not isinstance(evidence_list, list):
evidence_list = [str(evidence_list)]
# Filter out None or empty items
evidence_list = [item for item in evidence_list if item]
# Check if claim contains strong assertions that would require specific evidence
strong_assertion_markers = [
"solved", "cured", "discovered", "breakthrough", "revolutionary",
"first ever", "confirmed", "definitive", "conclusive", "proven",
"groundbreaking", "unprecedented", "remarkable", "extends lifespan",
"extends life", "definitively", "successfully"
]
# Determine if claim contains strong assertions
claim_has_strong_assertions = any(marker in claim.lower() for marker in strong_assertion_markers)
# Log detection result
if claim_has_strong_assertions:
logger.info(f"Evidence analysis: Detected claim with strong assertions requiring specific evidence")
# Extract named entities from claim
claim_entities = extract_entities(claim)
claim_entity_texts = [entity["text"].lower() for entity in claim_entities]
# Process each evidence item
analyzed_evidence = []
# Track domains found in evidence to identify source diversity
found_domains = set()
for evidence in evidence_list:
if not isinstance(evidence, str):
continue
# Calculate semantic similarity
similarity = calculate_similarity(claim, evidence)
# Check for entity overlap
evidence_entities = extract_entities(evidence)
evidence_entity_texts = [entity["text"].lower() for entity in evidence_entities]
# Calculate entity overlap
common_entities = set(claim_entity_texts).intersection(set(evidence_entity_texts))
entity_overlap = len(common_entities) / max(1, len(claim_entity_texts))
# Check temporal relevance
temporal_relevance = 1.0
if is_temporally_relevant(evidence, claim):
temporal_relevance = 1.2
else:
# Penalty for temporally irrelevant evidence
temporal_relevance = 0.7
# Check for authority signals
authority_score, authority_type = has_authority_signal(evidence)
# Extract source from evidence if available
source_boost = 1.0
domain = None
if source_credibility:
# Try to extract domain from URL in evidence
domain_match = re.search(r'URL: https?://(?:www\.)?([^/]+)', evidence)
if domain_match:
domain = domain_match.group(1)
# Check if domain or its parent domain is in credibility list
for cred_domain, cred_score in source_credibility.items():
if cred_domain in domain:
try:
source_boost = float(cred_score)
break
except (ValueError, TypeError):
pass
# Track this domain for source diversity
if domain:
found_domains.add(domain)
# For claims with strong assertions: check if evidence specifically addresses assertions
claim_specificity_match = 1.0
evidence_specificity_match = 1.0
if claim_has_strong_assertions:
# Check if evidence provides specific confirmation or contradiction
direct_contradiction_terms = [
"not yet", "has not", "have not", "cannot", "can't", "doesn't", "don't",
"unlikely", "challenging", "remains a challenge", "in the future",
"experimental", "in development", "proposed", "theoretical",
"preliminary", "hypothesized", "potential", "promising but"
]
# Check for contradictions to strong assertions
if any(term in evidence.lower() for term in direct_contradiction_terms):
# This evidence likely contradicts the strong assertion
evidence_specificity_match = 2.0 # Boost relevance of contradicting evidence
logger.debug(f"Found contradiction to strong assertion in evidence")
# For claims with strong assertions, check if evidence specifically confirms
direct_confirmation_terms = [
"successfully demonstrated", "breakthrough", "solved", "cured",
"confirmed", "definitive evidence", "conclusive results", "proven",
"revolutionary results", "milestone achievement", "groundbreaking results"
]
# If evidence confirms the strong assertion, adjust relevance
if any(term in evidence.lower() for term in direct_confirmation_terms):
# Apply higher scoring for evidence that specifically confirms
evidence_specificity_match = 1.8
logger.debug(f"Found confirmation of strong assertion in evidence")
# For claims with strong assertions, check for high-quality sources
high_quality_source_markers = [
"journal", "doi.org", "research", "university", "institute",
"laboratory", "professor", "study", "publication", "published in"
]
is_high_quality = any(term in evidence.lower() for term in high_quality_source_markers)
quality_boost = 1.4 if is_high_quality else 1.0
# Apply the quality boost
source_boost *= quality_boost
# Calculate final relevance score with improvements for all claim types
if claim_has_strong_assertions:
relevance_score = (
(similarity * 0.35) + # Semantic similarity
(entity_overlap * 0.25) + # Entity overlap
(0.25) # Base value to ensure all evidence has some relevance
) * temporal_relevance * authority_score * source_boost * claim_specificity_match * evidence_specificity_match
else:
# Original formula for regular claims
relevance_score = (
(similarity * 0.4) + # Semantic similarity
(entity_overlap * 0.3) + # Entity overlap
(0.3) # Base value to ensure all evidence has some relevance
) * temporal_relevance * authority_score * source_boost
# Add metadata and relevance score
analyzed_evidence.append({
"text": evidence,
"relevance_score": relevance_score,
"similarity": similarity,
"entity_overlap": entity_overlap,
"temporal_relevance": temporal_relevance,
"authority_score": authority_score,
"authority_type": authority_type,
"source_boost": source_boost,
"domain": domain
})
# Sort by relevance score (descending)
analyzed_evidence.sort(key=lambda x: x["relevance_score"], reverse=True)
# Ensure we have diverse sources in top results for all claims
if len(found_domains) > 1:
# Try to promote evidence from reliable sources if we haven't selected any yet
reliable_sources_seen = False
# Check if top 3 results contain any reliable sources
for item in analyzed_evidence[:3]:
domain = item.get("domain", "")
if domain and source_credibility and any(cred_domain in domain for cred_domain in source_credibility):
reliable_sources_seen = True
break
# If no reliable sources in top results, promote one if available
if not reliable_sources_seen:
for i, item in enumerate(analyzed_evidence[3:]):
domain = item.get("domain", "")
if domain and source_credibility and any(cred_domain in domain for cred_domain in source_credibility):
# Swap this item into the top 3
analyzed_evidence.insert(2, analyzed_evidence.pop(i+3))
break
return analyzed_evidence
def select_diverse_evidence(analyzed_evidence, max_items=5):
"""
Select diverse evidence items based on relevance, source diversity and claim characteristics
Args:
analyzed_evidence (list): List of evidence items with relevance scores
max_items (int): Maximum number of evidence items to return
Returns:
list: Selected diverse evidence items
"""
if not analyzed_evidence:
return []
# Check if top evidence suggests claim has strong assertions
strong_assertion_markers = [
"solved", "cured", "discovered", "breakthrough", "revolutionary",
"first ever", "confirmed", "definitive", "conclusive", "proven",
"groundbreaking", "unprecedented", "extends lifespan", "definitively"
]
# Determine if this is a claim with strong assertions by checking evidence text
has_strong_assertions = False
for item in analyzed_evidence[:3]: # Check just the top items for efficiency
if "text" in item:
item_text = item["text"].lower()
if any(f"claim {marker}" in item_text or f"claim has {marker}" in item_text
for marker in strong_assertion_markers):
has_strong_assertions = True
break
# Also check for contradiction markers in evidence which can indicate a strong assertion
contradiction_markers = [
"not yet solved", "hasn't been proven", "no evidence that",
"remains unsolved", "has not been confirmed", "remains theoretical"
]
if not has_strong_assertions:
for item in analyzed_evidence[:3]:
if "text" in item:
item_text = item["text"].lower()
if any(marker in item_text for marker in contradiction_markers):
has_strong_assertions = True
break
# Ensure we don't select more than available
max_items = min(max_items, len(analyzed_evidence))
# Initialize selected items with the most relevant item
selected = [analyzed_evidence[0]]
remaining = analyzed_evidence[1:]
# Track sources to ensure diversity
selected_sources = set()
for item in selected:
# Try to extract source from evidence
source_match = re.search(r'Source: ([^,]+)', item["text"])
if source_match:
selected_sources.add(source_match.group(1))
# For all claims, track if we have high-quality sources yet
has_quality_source = False
quality_source_markers = ["journal", "doi.org", "research", "university",
"institute", "laboratory", "professor", "study"]
# Check if our top item is already from a quality source
if any(marker in selected[0]["text"].lower() for marker in quality_source_markers):
has_quality_source = True
# Select remaining items balancing relevance and diversity
while len(selected) < max_items and remaining:
best_item = None
best_score = -1
for i, item in enumerate(remaining):
# Base score is the item's relevance
score = item["relevance_score"]
# Extract source if available
source = None
source_match = re.search(r'Source: ([^,]+)', item["text"])
if source_match:
source = source_match.group(1)
# Apply diversity bonus if source is new
if source and source not in selected_sources:
score *= 1.2 # Diversity bonus
# For claims with strong assertions, apply bonus for contradicting evidence
if has_strong_assertions:
# Check for contradiction markers in the text
if any(marker in item["text"].lower() for marker in contradiction_markers):
score *= 1.3 # Bonus for evidence that may contradict strong assertions
# For any claim, apply bonus for high-quality sources if we don't have one yet
if not has_quality_source:
is_item_quality = any(marker in item["text"].lower() for marker in quality_source_markers)
if is_item_quality:
score *= 1.5 # Significant bonus for quality sources
if score > best_score:
best_score = score
best_item = (i, item)
if best_item:
idx, item = best_item
selected.append(item)
remaining.pop(idx)
# Add source to selected sources
source_match = re.search(r'Source: ([^,]+)', item["text"])
if source_match:
selected_sources.add(source_match.group(1))
# Check if we found a quality source
if not has_quality_source:
if any(marker in item["text"].lower() for marker in quality_source_markers):
has_quality_source = True
else:
break
# For any claim with strong assertions, ensure we have at least one quality source if available
if has_strong_assertions and not has_quality_source and remaining:
for i, item in enumerate(remaining):
if any(marker in item["text"].lower() for marker in quality_source_markers):
# Replace the least relevant selected item with this quality one
selected.sort(key=lambda x: x["relevance_score"])
selected[0] = item
break
# Return only the text portion
return [item["text"] for item in selected] |