askveracity / modules /classification.py
ankanghosh's picture
Upload 21 files
6d11371 verified
import logging
import re
from utils.models import get_llm_model, get_nlp_model
from utils.performance import PerformanceTracker
logger = logging.getLogger("misinformation_detector")
performance_tracker = PerformanceTracker()
def classify_with_llm(query, evidence):
"""
Classification function that evaluates evidence against a claim
to determine support, contradiction, or insufficient evidence.
This function analyzes the provided evidence to evaluate if it supports,
contradicts, or is insufficient to verify the claim. It implements:
- Strict output formatting requirements
- Evidence source validation
- Confidence scoring based on confirmation strength
- Flexible regex pattern matching
- Detailed debug logging
- Fallback parsing for non-standard responses
Args:
query (str): The factual claim being verified
evidence (list): Evidence items to evaluate against the claim
Returns:
list: Classification results with labels and confidence scores
"""
logger.info(f"Classifying evidence for claim: {query}")
# Get the LLM model
llm_model = get_llm_model()
# Skip if no evidence
if not evidence:
logger.warning("No evidence provided for classification")
return []
# Normalize evidence to a list
if not isinstance(evidence, list):
if evidence:
try:
evidence = [evidence]
except Exception as e:
logger.error(f"Could not convert evidence to list: {e}")
return []
else:
return []
# Extract essential claim components for improved keyword detection
claim_components = extract_claim_keywords(query)
essential_keywords = claim_components.get("keywords", [])
essential_entities = claim_components.get("entities", [])
# Ensure processing is limited to top 10 evidence items to reduce token usage
evidence = evidence[:10]
# Validate evidence for verifiable sources
validated_evidence = []
for idx, chunk in enumerate(evidence):
# Basic evidence validation
if not isinstance(chunk, str) or not chunk.strip():
continue
# Check if evidence contains source information
has_valid_source = False
if "URL:" in chunk and ("http://" in chunk or "https://" in chunk):
has_valid_source = True
elif "Source:" in chunk and len(chunk.split("Source:")[1].strip()) > 3:
has_valid_source = True
# Add validation flag to evidence
validated_evidence.append({
"text": chunk,
"index": idx + 1,
"has_valid_source": has_valid_source
})
# If no valid evidence remains, return early
if not validated_evidence:
logger.warning("No valid evidence items to classify")
return []
try:
# Format evidence items with validation information
evidence_text = ""
for item in validated_evidence:
# Truncate long evidence
chunk_text = item["text"]
if len(chunk_text) > 1000:
chunk_text = chunk_text[:1000] + "..."
# Include validation status in the prompt
source_status = "WITH VALID SOURCE" if item["has_valid_source"] else "WARNING: NO CLEAR SOURCE"
evidence_text += f"EVIDENCE {item['index']}:\n{chunk_text}\n[{source_status}]\n\n"
# Create a structured prompt with explicit format instructions and validation requirements
prompt = f"""
CLAIM: {query}
EVIDENCE:
{evidence_text}
TASK: Evaluate if each evidence supports, contradicts, or is insufficient/irrelevant to the claim.
INSTRUCTIONS:
1. For each evidence, provide your analysis in EXACTLY this format:
EVIDENCE [number] ANALYSIS:
Classification: [Choose exactly one: support/contradict/insufficient]
Confidence: [number between 0-100]
Reason: [brief explanation]
2. Support = Evidence EXPLICITLY confirms ALL parts of the claim are true
3. Contradict = Evidence EXPLICITLY confirms the claim is false
4. Insufficient = Evidence is irrelevant, ambiguous, or doesn't provide enough information
CRITICAL VALIDATION RULES:
- Mark as "support" ONLY when evidence EXPLICITLY mentions ALL key entities AND actions from the claim
- Do not label evidence as "support" if it only discusses the same topic without confirming the specific claim
- Do not make inferential leaps - if the evidence doesn't explicitly state the claim, mark as "insufficient"
- Assign LOW confidence (0-50) when evidence doesn't explicitly mention all claim elements
- Assign ZERO confidence (0) to evidence without valid sources
- If evidence describes similar but different events, mark as "insufficient", not "support" or "contradict"
- If evidence describes the same topic as the claim but does not confirm or contradict the claim, mark as "insufficient", not "support" or "contradict"
- If evidence is in a different language or unrelated topic, mark as "insufficient" with 0 confidence
- Check that all entities (names, places, dates, numbers) in the claim are explicitly confirmed
FOCUS ON THE EXACT CLAIM ONLY.
ESSENTIAL KEYWORDS TO LOOK FOR: {', '.join(essential_keywords)}
ESSENTIAL ENTITIES TO VERIFY: {', '.join(essential_entities)}
IMPORTANT NOTE ABOUT VERB TENSES: When analyzing this claim, treat present tense verbs (like "unveils")
and perfect form verbs (like "has unveiled") as equivalent to their simple past tense forms
(like "unveiled"). The tense variation should not affect your classification decision.
"""
# Get response with temperature=0 for consistency
result = llm_model.invoke(prompt, temperature=0)
result_text = result.content.strip()
# Log the raw LLM response for debugging
logger.debug(f"Raw LLM classification response:\n{result_text}")
# Define a more flexible regex pattern matching the requested format
# This pattern accommodates variations in whitespace and formatting
analysis_pattern = r'EVIDENCE\s+(\d+)\s+ANALYSIS:[\s\n]*Classification:[\s\n]*(support|contradict|insufficient)[\s\n]*Confidence:[\s\n]*(\d+)[\s\n]*Reason:[\s\n]*(.*?)(?=[\s\n]*EVIDENCE\s+\d+\s+ANALYSIS:|[\s\n]*$)'
# Parse each evidence analysis
classification_results = []
# Try matching with our pattern
matches = list(re.finditer(analysis_pattern, result_text, re.IGNORECASE | re.DOTALL))
# Log match information for debugging
logger.debug(f"Found {len(matches)} structured evidence analyses in response")
# Process matches
for match in matches:
try:
evidence_idx = int(match.group(1)) - 1
classification = match.group(2).lower()
confidence = int(match.group(3)) / 100.0 # Convert to 0-1 scale
reason = match.group(4).strip()
# Check if this evidence item exists in our original list
if 0 <= evidence_idx < len(evidence):
# Get the original evidence text
evidence_text = evidence[evidence_idx]
# Check for valid source
source_valid = False
if "URL:" in evidence_text and ("http://" in evidence_text or "https://" in evidence_text):
source_valid = True
elif "Source:" in evidence_text:
source_valid = True
# Reduce confidence for evidence without valid sources
if not source_valid and confidence > 0.3:
confidence = 0.3
reason += " (Confidence reduced due to lack of verifiable source)"
# Create result entry
classification_results.append({
"label": classification,
"confidence": confidence,
"evidence": evidence_text,
"reason": reason
})
except (ValueError, IndexError) as e:
logger.error(f"Error parsing evidence analysis: {e}")
# If no structured matches were found, try using a simpler approach
if not classification_results:
logger.warning("No structured evidence analysis found, using fallback method")
# Log detailed information about the failure
logger.warning(f"Expected format not found in response. Response excerpt: {result_text[:200]}...")
# Simple fallback parsing based on keywords
for idx, ev in enumerate(evidence):
# Check for keywords in the LLM response
ev_mention = f"EVIDENCE {idx+1}"
if ev_mention in result_text:
# Find the section for this evidence
parts = result_text.split(ev_mention)
if len(parts) > 1:
analysis_text = parts[1].split("EVIDENCE")[0] if "EVIDENCE" in parts[1] else parts[1]
# Determine classification
label = "insufficient" # Default
confidence = 0.0 # Default - zero confidence for fallback parsing
# Check for support indicators
if "support" in analysis_text.lower() or "confirms" in analysis_text.lower():
label = "support"
confidence = 0.4 # Lower confidence for fallback support
# Check for contradict indicators
elif "contradict" in analysis_text.lower() or "false" in analysis_text.lower():
label = "contradict"
confidence = 0.4 # Lower confidence for fallback contradict
# Check for valid source to adjust confidence
source_valid = False
if "URL:" in ev and ("http://" in ev or "https://" in ev):
source_valid = True
elif "Source:" in ev:
source_valid = True
if not source_valid:
confidence = min(confidence, 0.3)
# Create basic result
classification_results.append({
"label": label,
"confidence": confidence,
"evidence": ev,
"reason": f"Determined via fallback parsing. {'Valid source found.' if source_valid else 'Warning: No clear source identified.'}"
})
logger.debug(f"Fallback parsing for evidence {idx+1}: {label} with confidence {confidence}")
logger.info(f"Classified {len(classification_results)} evidence items")
return classification_results
except Exception as e:
logger.error(f"Error in evidence classification: {str(e)}")
# Provide a basic fallback
fallback_results = []
for ev in evidence:
fallback_results.append({
"label": "insufficient",
"confidence": 0.5,
"evidence": ev,
"reason": "Classification failed with error, using fallback"
})
return fallback_results
def normalize_tense(claim):
"""
Normalize verb tenses in claims to ensure consistent classification.
This function standardizes verb forms by converting present simple tense
verbs (e.g., "unveils") and perfect forms (e.g., "has unveiled") to their
past tense equivalents (e.g., "unveiled"). This ensures that semantically
equivalent claims are processed consistently regardless of verb tense
variations.
Args:
claim (str): The original claim text to normalize
Returns:
str: The normalized claim with consistent tense handling
Note:
This function specifically targets present simple and perfect forms,
preserving the semantic differences of continuous forms (is unveiling)
and future tense (will unveil).
"""
# Define patterns to normalize common verb forms.
# Each tuple contains (regex_pattern, replacement_text)
tense_patterns = [
# Present simple to past tense conversions
(r'\bunveils\b', r'unveiled'),
(r'\blaunches\b', r'launched'),
(r'\breleases\b', r'released'),
(r'\bannounces\b', r'announced'),
(r'\binvites\b', r'invited'),
(r'\bretaliates\b', r'retaliated'),
(r'\bends\b', r'ended'),
(r'\bbegins\b', r'began'),
(r'\bstarts\b', r'started'),
(r'\bcompletes\b', r'completed'),
(r'\bfinishes\b', r'finished'),
(r'\bintroduces\b', r'introduced'),
(r'\bcreates\b', r'created'),
(r'\bdevelops\b', r'developed'),
(r'\bpublishes\b', r'published'),
(r'\bacquires\b', r'acquired'),
(r'\bbuys\b', r'bought'),
(r'\bsells\b', r'sold'),
# Perfect forms (has/have/had + past participle) to simple past
(r'\b(has|have|had)\s+unveiled\b', r'unveiled'),
(r'\b(has|have|had)\s+launched\b', r'launched'),
(r'\b(has|have|had)\s+released\b', r'released'),
(r'\b(has|have|had)\s+announced\b', r'announced'),
(r'\b(has|have|had)\s+invited\b', r'invited'),
(r'\b(has|have|had)\s+retaliated\b', r'retaliated'),
(r'\b(has|have|had)\s+ended\b', r'ended'),
(r'\b(has|have|had)\s+begun\b', r'began'),
(r'\b(has|have|had)\s+started\b', r'started'),
(r'\b(has|have|had)\s+introduced\b', r'introduced'),
(r'\b(has|have|had)\s+created\b', r'created'),
(r'\b(has|have|had)\s+developed\b', r'developed'),
(r'\b(has|have|had)\s+published\b', r'published'),
(r'\b(has|have|had)\s+acquired\b', r'acquired'),
(r'\b(has|have|had)\s+bought\b', r'bought'),
(r'\b(has|have|had)\s+sold\b', r'sold')
]
# Apply normalization patterns
normalized = claim
for pattern, replacement in tense_patterns:
normalized = re.sub(pattern, replacement, normalized, flags=re.IGNORECASE)
# Log if normalization occurred for debugging purposes
if normalized != claim:
logger.info(f"Normalized claim from: '{claim}' to: '{normalized}'")
return normalized
def aggregate_evidence(classification_results):
"""
Aggregate evidence classifications to determine overall verdict
using a weighted scoring system of evidence count and quality.
Args:
classification_results (list): List of evidence classification results
Returns:
tuple: (verdict, confidence) - The final verdict and confidence score
"""
logger.info(f"Aggregating evidence from {len(classification_results) if classification_results else 0} results")
if not classification_results:
logger.warning("No classification results to aggregate")
return "Uncertain", 0.0 # Default with zero confidence
# Only consider support and contradict evidence items
support_items = [item for item in classification_results if item.get("label") == "support"]
contradict_items = [item for item in classification_results if item.get("label") == "contradict"]
# Count number of support and contradict items
support_count = len(support_items)
contradict_count = len(contradict_items)
# Calculate confidence scores for support and contradict items
support_confidence_sum = sum(item.get("confidence", 0) for item in support_items)
contradict_confidence_sum = sum(item.get("confidence", 0) for item in contradict_items)
# Apply weights: 55% for count, 45% for quality (confidence)
# Normalize counts to avoid division by zero
max_count = max(1, max(support_count, contradict_count))
# Calculate weighted scores
count_support_score = (support_count / max_count) * 0.55
count_contradict_score = (contradict_count / max_count) * 0.55
# Normalize confidence scores to avoid division by zero
max_confidence_sum = max(1, max(support_confidence_sum, contradict_confidence_sum))
quality_support_score = (support_confidence_sum / max_confidence_sum) * 0.45
quality_contradict_score = (contradict_confidence_sum / max_confidence_sum) * 0.45
# Total scores
total_support = count_support_score + quality_support_score
total_contradict = count_contradict_score + quality_contradict_score
# Check if all evidence is irrelevant/insufficient
if support_count == 0 and contradict_count == 0:
logger.info("All evidence items are irrelevant/insufficient")
return "Uncertain", 0.0
# Determine verdict based on higher total score
if total_support > total_contradict:
verdict = "True (Based on Evidence)"
min_score = total_contradict
max_score = total_support
else:
verdict = "False (Based on Evidence)"
min_score = total_support
max_score = total_contradict
# Calculate final confidence using the formula:
# (1 - min_score/max_score) * 100%
if max_score > 0:
final_confidence = 1.0 - (min_score / max_score)
else:
final_confidence = 0.0
# Handle cases where confidence is very low
if final_confidence == 0.0:
return "Uncertain", 0.0
elif final_confidence < 0.1: # Less than 10%
# Keep the verdict but with very low confidence
logger.info(f"Very low confidence verdict: {verdict} with {final_confidence:.2f} confidence")
logger.info(f"Final verdict: {verdict}, confidence: {final_confidence:.2f}")
return verdict, final_confidence
def extract_claim_keywords(claim):
"""
Extract important keywords from claim using NLP processing
Args:
claim (str): The claim text
Returns:
dict: Dictionary containing keywords and other claim components
"""
try:
# Get NLP model
nlp = get_nlp_model()
# Process claim with NLP
doc = nlp(claim)
# Extract entities
entities = [ent.text for ent in doc.ents]
# Extract important keywords (non-stopword nouns, adjectives, and verbs longer than 3 chars)
keywords = []
for token in doc:
# Keep all important parts of speech, longer than 3 characters
if token.pos_ in ["NOUN", "PROPN", "ADJ", "VERB"] and not token.is_stop and len(token.text) > 3:
keywords.append(token.text.lower())
# Also include some important modifiers and quantifiers
elif token.pos_ in ["NUM", "ADV"] and not token.is_stop and len(token.text) > 1:
keywords.append(token.text.lower())
# Extract verbs separately
verbs = [token.lemma_.lower() for token in doc if token.pos_ == "VERB" and not token.is_stop]
# Also extract multi-word phrases that might be important
noun_phrases = []
for chunk in doc.noun_chunks:
if len(chunk.text) > 3 and not all(token.is_stop for token in chunk):
noun_phrases.append(chunk.text.lower())
# Add phrases to keywords if not already included
for phrase in noun_phrases:
if phrase not in keywords and phrase.lower() not in [k.lower() for k in keywords]:
keywords.append(phrase.lower())
# Return all components
return {
"entities": entities,
"keywords": keywords,
"verbs": verbs,
"noun_phrases": noun_phrases
}
except Exception as e:
logger.error(f"Error extracting claim keywords: {e}")
# Return basic fallback using simple word extraction
words = [word.lower() for word in claim.split() if len(word) > 3]
return {"keywords": words, "entities": [], "verbs": [], "noun_phrases": []}