Spaces:
Sleeping
Sleeping
File size: 11,619 Bytes
5b6e847 1bbe15b 5b6e847 1bbe15b 5b6e847 1bbe15b 5b6e847 1bbe15b 5b6e847 1bbe15b 5b6e847 1bbe15b 5b6e847 1bbe15b 5b6e847 1bbe15b 5b6e847 1bbe15b 5b6e847 1bbe15b 5b6e847 1bbe15b 5b6e847 1bbe15b 5b6e847 1bbe15b 5b6e847 1bbe15b 5b6e847 1bbe15b 5b6e847 1bbe15b 5b6e847 1bbe15b 5b6e847 1bbe15b 5b6e847 1bbe15b 5b6e847 1bbe15b 5b6e847 1bbe15b 5b6e847 1bbe15b 5b6e847 1bbe15b 5b6e847 1bbe15b 5b6e847 1bbe15b 5b6e847 1bbe15b 5b6e847 1bbe15b 5b6e847 1bbe15b f433f81 5b6e847 1bbe15b 5b6e847 1bbe15b 5b6e847 1bbe15b 5b6e847 1bbe15b 5b6e847 1bbe15b 5b6e847 1bbe15b 5b6e847 1bbe15b 5b6e847 1bbe15b 5b6e847 1bbe15b 5b6e847 1bbe15b 116c121 5b6e847 116c121 5b6e847 1bbe15b 5b6e847 1bbe15b 5b6e847 1bbe15b 5b6e847 1bbe15b 5b6e847 1bbe15b 5b6e847 1bbe15b 5b6e847 1bbe15b 5b6e847 1bbe15b 5b6e847 1bbe15b 00a7178 5b6e847 00a7178 5b6e847 1bbe15b 5b6e847 1bbe15b 5b6e847 1bbe15b 5b6e847 1bbe15b 5b6e847 1bbe15b 5b6e847 1bbe15b 5b6e847 1bbe15b 5b6e847 1bbe15b 5b6e847 1bbe15b 5b6e847 1bbe15b 5b6e847 1bbe15b 5b6e847 1bbe15b 5b6e847 | 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 | """
Verification Module for VDHF
Handles claim verification using semantic similarity and NLI entailment.
"""
from typing import List, Tuple, Optional, Dict, Any
from dataclasses import dataclass
import numpy as np
try:
from transformers import pipeline, AutoTokenizer, AutoModelForSequenceClassification
import torch
except ImportError:
pipeline = None
torch = None
from config.settings import (
SIMILARITY_THRESHOLD,
NLI_MODEL,
EMBEDDING_MODEL
)
from core.claim_extractor import Claim
from ingestion.embeddings import EmbeddingModel, compute_cosine_similarity
from retrieval.retriever import RetrievedEvidence
@dataclass
class VerificationResult:
"""Result of claim verification."""
claim: Claim
is_supported: bool
similarity_score: float
entailment_label: str # ENTAILED, NEUTRAL, CONTRADICTED
entailment_score: float
best_evidence: str
evidence_source: str
def __str__(self) -> str:
status = "SUPPORTED" if self.is_supported else "UNSUPPORTED"
return (
f"[{status}] Claim: {self.claim.text[:50]}...\n"
f" Similarity: {self.similarity_score:.3f}, "
f"Entailment: {self.entailment_label} ({self.entailment_score:.3f})"
)
class SemanticSimilarityChecker:
"""
Semantic Similarity Module
Measures semantic closeness between claims and evidence using embeddings.
"""
def __init__(self, embedding_model: Optional[EmbeddingModel] = None):
self.embedding_model = embedding_model or EmbeddingModel()
def compute_similarity(self, claim: str, evidence: str) -> float:
"""
Compute semantic similarity between claim and evidence.
Args:
claim: Claim text
evidence: Evidence text
Returns:
Cosine similarity score (0 to 1)
"""
claim_embedding = self.embedding_model.embed_single(claim)
evidence_embedding = self.embedding_model.embed_single(evidence)
return compute_cosine_similarity(claim_embedding, evidence_embedding)
def find_best_evidence(
self,
claim: str,
evidence_list: List[RetrievedEvidence]
) -> Tuple[RetrievedEvidence, float]:
"""
Find the most similar evidence for a claim.
Args:
claim: Claim text
evidence_list: List of evidence candidates
Returns:
Tuple of (best evidence, similarity score)
"""
if not evidence_list:
return None, 0.0
best_evidence = None
best_score = 0.0
claim_embedding = self.embedding_model.embed_single(claim)
for evidence in evidence_list:
evidence_embedding = self.embedding_model.embed_single(evidence.content)
score = compute_cosine_similarity(claim_embedding, evidence_embedding)
if score > best_score:
best_score = score
best_evidence = evidence
return best_evidence, best_score
class EntailmentChecker:
"""
NLI Entailment Module
Checks whether evidence logically entails a claim.
"""
def __init__(self, model_name: str = NLI_MODEL):
self.model_name = model_name
self._pipeline = None
if pipeline is not None:
try:
self._pipeline = pipeline(
"text-classification",
model=model_name,
top_k=None
)
except Exception as e:
print(f"Warning: Could not load NLI model: {e}")
def check_entailment(
self,
premise: str,
hypothesis: str
) -> Tuple[str, float]:
"""
Check if premise entails hypothesis.
Args:
premise: Evidence text (premise)
hypothesis: Claim text (hypothesis)
Returns:
Tuple of (label, score) where label is ENTAILED/NEUTRAL/CONTRADICTED
"""
if self._pipeline is None:
# Fallback to simple heuristic
return self._heuristic_entailment(premise, hypothesis)
try:
# Format input for NLI model
input_text = f"{premise} [SEP] {hypothesis}"
# Get predictions
results = self._pipeline(input_text)
# Parse results
label_mapping = {
'ENTAILMENT': 'ENTAILED',
'CONTRADICTION': 'CONTRADICTED',
'NEUTRAL': 'NEUTRAL',
'entailment': 'ENTAILED',
'contradiction': 'CONTRADICTED',
'neutral': 'NEUTRAL',
}
best_label = 'NEUTRAL'
best_score = 0.0
# Handle both formats: list of dicts or list of list of dicts (top_k=None)
items = results
if isinstance(results, list) and results and isinstance(results[0], list):
items = results[0]
for result in items:
label = result['label'].upper()
score = result['score']
mapped_label = label_mapping.get(label, label)
if mapped_label in ['ENTAILED', 'NEUTRAL', 'CONTRADICTED']:
if score > best_score:
best_score = score
best_label = mapped_label
return best_label, best_score
except Exception as e:
print(f"NLI error: {e}")
return self._heuristic_entailment(premise, hypothesis)
def _heuristic_entailment(
self,
premise: str,
hypothesis: str
) -> Tuple[str, float]:
"""
Simple heuristic for entailment when NLI model unavailable.
Based on word overlap and key phrase matching.
"""
premise_words = set(premise.lower().split())
hypothesis_words = set(hypothesis.lower().split())
# Remove common stop words
stop_words = {'the', 'a', 'an', 'is', 'are', 'was', 'were', 'be', 'been',
'being', 'have', 'has', 'had', 'do', 'does', 'did', 'will',
'would', 'could', 'should', 'may', 'might', 'can', 'to',
'of', 'in', 'for', 'on', 'with', 'at', 'by', 'from', 'as',
'and', 'or', 'but', 'if', 'then', 'that', 'this', 'it'}
premise_words = premise_words - stop_words
hypothesis_words = hypothesis_words - stop_words
if not hypothesis_words:
return 'NEUTRAL', 0.5
overlap = len(premise_words & hypothesis_words)
overlap_ratio = overlap / len(hypothesis_words)
if overlap_ratio >= 0.5:
return 'ENTAILED', overlap_ratio
elif overlap_ratio >= 0.2:
return 'NEUTRAL', overlap_ratio
else:
return 'NEUTRAL', overlap_ratio
class ClaimVerifier:
"""
Claim Verification Module
Combines semantic similarity and NLI entailment to verify claims.
Verification Rule:
A claim is SUPPORTED if:
- Semantic similarity >= threshold (theta_sim)
- Entailment label == ENTAILED
"""
def __init__(
self,
similarity_threshold: float = SIMILARITY_THRESHOLD,
embedding_model: Optional[EmbeddingModel] = None
):
self.similarity_threshold = similarity_threshold
self.similarity_checker = SemanticSimilarityChecker(embedding_model)
self.entailment_checker = EntailmentChecker()
def verify_claim(
self,
claim: Claim,
evidence_list: List[RetrievedEvidence]
) -> VerificationResult:
"""
Verify a single claim against evidence.
Args:
claim: Claim to verify
evidence_list: Available evidence
Returns:
VerificationResult object
"""
# Find best matching evidence
best_evidence, similarity_score = self.similarity_checker.find_best_evidence(
claim.text,
evidence_list
)
if best_evidence is None:
return VerificationResult(
claim=claim,
is_supported=False,
similarity_score=0.0,
entailment_label='NEUTRAL',
entailment_score=0.0,
best_evidence="",
evidence_source=""
)
# Check entailment
entailment_label, entailment_score = self.entailment_checker.check_entailment(
premise=best_evidence.content,
hypothesis=claim.text
)
# Apply verification rule:
# Supported if EITHER high similarity OR entailment confirms it
is_supported = (
(similarity_score >= self.similarity_threshold and entailment_label in ('ENTAILED', 'NEUTRAL')) or
(similarity_score >= 0.5 and entailment_label == 'ENTAILED') or
(similarity_score >= 0.85)
)
# Update claim object
claim.is_verified = is_supported
claim.similarity_score = similarity_score
claim.entailment_label = entailment_label
claim.supporting_evidence = best_evidence.content
return VerificationResult(
claim=claim,
is_supported=is_supported,
similarity_score=similarity_score,
entailment_label=entailment_label,
entailment_score=entailment_score,
best_evidence=best_evidence.content,
evidence_source=best_evidence.metadata.get('source', 'Unknown')
)
def verify_all_claims(
self,
claims: List[Claim],
evidence_list: List[RetrievedEvidence]
) -> List[VerificationResult]:
"""
Verify all claims against evidence.
Args:
claims: List of claims to verify
evidence_list: Available evidence
Returns:
List of VerificationResult objects
"""
results = []
for claim in claims:
result = self.verify_claim(claim, evidence_list)
results.append(result)
return results
def get_verification_summary(
self,
results: List[VerificationResult]
) -> Dict[str, Any]:
"""
Get summary statistics of verification results.
Args:
results: List of VerificationResult objects
Returns:
Summary dictionary
"""
total = len(results)
supported = sum(1 for r in results if r.is_supported)
unsupported = total - supported
avg_similarity = np.mean([r.similarity_score for r in results]) if results else 0.0
return {
'total_claims': total,
'supported_claims': supported,
'unsupported_claims': unsupported,
'support_ratio': supported / total if total > 0 else 0.0,
'average_similarity': avg_similarity
}
def verify_response(
claims: List[Claim],
evidence_list: List[RetrievedEvidence],
similarity_threshold: float = SIMILARITY_THRESHOLD
) -> Tuple[List[VerificationResult], Dict[str, Any]]:
"""
Convenience function to verify all claims in a response.
Args:
claims: Extracted claims
evidence_list: Retrieved evidence
similarity_threshold: Verification threshold
Returns:
Tuple of (verification results, summary)
"""
verifier = ClaimVerifier(similarity_threshold=similarity_threshold)
results = verifier.verify_all_claims(claims, evidence_list)
summary = verifier.get_verification_summary(results)
return results, summary
|