File size: 12,441 Bytes
a561338 |
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 |
"""
Semantic analysis with MiniLM embeddings for paraphrasing detection.
Detects even heavily paraphrased content from LLMs like ChatGPT.
"""
import re
from typing import List, Dict
import logging
import numpy as np
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger("semantic_utils")
# Lazy load model to avoid multiple loads
_encoder_model = None
def get_encoder():
"""Lazy load SentenceTransformer MiniLM model"""
global _encoder_model
if _encoder_model is None:
logger.info("π Loading SentenceTransformer MiniLM-L6-v2 model...")
try:
from sentence_transformers import SentenceTransformer
_encoder_model = SentenceTransformer('all-MiniLM-L6-v2')
logger.info("β
Model loaded successfully")
except ImportError:
logger.error("β sentence-transformers not installed. Install with: pip install sentence-transformers")
raise
return _encoder_model
def split_into_sentences(text: str) -> List[str]:
"""Split text into sentences, filtering out junk"""
sentences = re.split(r'(?<=[.!?])\s+', text)
cleaned = []
for sent in sentences:
sent = sent.strip()
if len(sent) > 20 and len(sent.split()) >= 5:
cleaned.append(sent)
return cleaned
def extract_key_sentences(text: str, num_sentences: int = 5) -> List[str]:
"""
Extract the most important sentences from text based on:
- Position (first and last sentences often important)
- Length (15-30 words is optimal)
- Keyword frequency
"""
sentences = split_into_sentences(text)
if len(sentences) <= num_sentences:
return sentences
scored_sentences = []
for idx, sent in enumerate(sentences):
score = 0.0
word_count = len(sent.split())
# Position score
if idx < len(sentences) * 0.2 or idx > len(sentences) * 0.8:
score += 0.3
# Length score
if 15 <= word_count <= 30:
score += 0.3
elif 10 <= word_count <= 40:
score += 0.15
# Keyword diversity
words = set(sent.lower().split())
common_words = {'the', 'a', 'an', 'and', 'or', 'is', 'are', 'was', 'be', 'it', 'this', 'that'}
unique_words = len(words - common_words)
score += min(unique_words / 10, 0.4)
scored_sentences.append((score, sent))
scored_sentences.sort(reverse=True)
key_sents = [sent for _, sent in scored_sentences[:num_sentences]]
result = []
for sent in sentences:
if sent in key_sents:
result.append(sent)
return result
def generate_five_queries(text: str, max_words: int = 3000) -> List[str]:
"""
Generate 5 high-quality semantic search queries from document.
Query 1: Beginning (main topic)
Query 2: Early Middle
Query 3: Center (supporting evidence)
Query 4: Late Middle
Query 5: End (conclusions)
"""
logger.info(" π Generating 5 semantic queries from content...")
words = text.split()
if len(words) > max_words:
text = ' '.join(words[:max_words])
sentences = split_into_sentences(text)
if len(sentences) < 5:
# Fallback for very short text
logger.warning(" β οΈ Very short document, using basic queries")
return [
' '.join(words[:30]),
' '.join(words[max(0, len(words)//5):max(0, len(words)//5)+30]),
' '.join(words[max(0, 2*len(words)//5):max(0, 2*len(words)//5)+30]),
' '.join(words[max(0, 3*len(words)//5):max(0, 3*len(words)//5)+30]),
' '.join(words[max(0, 4*len(words)//5):max(0, 4*len(words)//5)+30])
]
queries = []
total_sentences = len(sentences)
# β
Query 1: BEGINNING - First 3-4 sentences
beginning_end = min(4, total_sentences // 5)
query1_sents = sentences[:beginning_end]
query1 = ' '.join(query1_sents)
queries.append(query1)
logger.debug(f" Query 1 (Beginning) length: {len(query1.split())} words")
# β
Query 2: EARLY MIDDLE - First third
early_start = beginning_end
early_end = min(early_start + 4, total_sentences // 3)
query2_sents = sentences[early_start:early_end]
query2 = ' '.join(query2_sents)
queries.append(query2)
logger.debug(f" Query 2 (Early Middle) length: {len(query2.split())} words")
# β
Query 3: CENTER - Middle section (3-4 sentences)
mid_start = max(early_end, total_sentences // 3)
mid_end = min(mid_start + 4, 2 * total_sentences // 3)
query3_sents = sentences[mid_start:mid_end]
query3 = ' '.join(query3_sents)
queries.append(query3)
logger.debug(f" Query 3 (Center) length: {len(query3.split())} words")
# β
Query 4: LATE MIDDLE - Second two-thirds
late_start = mid_end
late_end = min(late_start + 4, 2 * total_sentences // 3 + (total_sentences // 3))
query4_sents = sentences[late_start:late_end]
query4 = ' '.join(query4_sents)
queries.append(query4)
logger.debug(f" Query 4 (Late Middle) length: {len(query4.split())} words")
# β
Query 5: END - Last 3-4 sentences
end_start = max(late_end, total_sentences - 4)
query5_sents = sentences[end_start:]
query5 = ' '.join(query5_sents)
queries.append(query5)
logger.debug(f" Query 5 (End) length: {len(query5.split())} words")
# β
Clean and validate queries
final_queries = []
for q in queries:
q = q.strip()
if len(q.split()) >= 15: # Minimum 15 words for good search
final_queries.append(q)
logger.info(f" β
Generated {len(final_queries)} queries:")
for i, q in enumerate(final_queries, 1):
word_count = len(q.split())
preview = q[:80] + "..." if len(q) > 80 else q
logger.info(f" Query {i} ({word_count} words): {preview}")
return final_queries
def find_semantic_matches(
doc_text: str,
source_text: str,
threshold: float = 0.50
) -> List[Dict]:
"""
Find semantically similar passages using MiniLM embeddings.
Detects paraphrased content from LLMs.
threshold: 0.65+ = high confidence, 0.50+ = catch paraphrasing, 0.35+ = catch weak matches
"""
try:
encoder = get_encoder()
except ImportError:
logger.warning("β οΈ SentenceTransformer not available, falling back to string matching")
return _fallback_semantic_matches(doc_text, source_text, threshold)
# Split into sentences
doc_sentences = split_into_sentences(doc_text)
source_sentences = split_into_sentences(source_text)
if not doc_sentences or not source_sentences:
return []
logger.debug(f" Encoding {len(doc_sentences)} doc sentences + {len(source_sentences)} source sentences...")
# Encode all sentences
try:
doc_embeddings = encoder.encode(doc_sentences, convert_to_numpy=True, show_progress_bar=False)
source_embeddings = encoder.encode(source_sentences, convert_to_numpy=True, show_progress_bar=False)
except Exception as e:
logger.error(f" Encoding error: {e}, falling back to string matching")
return _fallback_semantic_matches(doc_text, source_text, threshold)
matches = []
matched_source_indices = set()
# Compare using cosine similarity
for doc_idx, doc_emb in enumerate(doc_embeddings):
best_similarity = 0.0
best_source_idx = -1
for source_idx, source_emb in enumerate(source_embeddings):
if source_idx in matched_source_indices:
continue
# Cosine similarity
similarity = np.dot(doc_emb, source_emb) / (
np.linalg.norm(doc_emb) * np.linalg.norm(source_emb) + 1e-8
)
if similarity > best_similarity:
best_similarity = similarity
best_source_idx = source_idx
# Record if above threshold
if best_similarity >= threshold and best_source_idx >= 0:
if best_source_idx not in matched_source_indices:
matched_source_indices.add(best_source_idx)
matches.append({
'doc_text': doc_sentences[doc_idx],
'source_text': source_sentences[best_source_idx],
'similarity': float(best_similarity),
'doc_index': doc_idx,
'source_index': best_source_idx
})
logger.debug(f" Found {len(matches)} semantic matches (threshold: {threshold})")
return matches
def _fallback_semantic_matches(doc_text: str, source_text: str, threshold: float) -> List[Dict]:
"""Fallback to string matching if embeddings not available"""
from difflib import SequenceMatcher
doc_sentences = split_into_sentences(doc_text)
source_sentences = split_into_sentences(source_text)
if not doc_sentences or not source_sentences:
return []
matches = []
matched_source_indices = set()
for doc_idx, doc_sent in enumerate(doc_sentences):
best_similarity = 0.0
best_source_idx = -1
for source_idx, source_sent in enumerate(source_sentences):
if source_idx in matched_source_indices:
continue
ratio = SequenceMatcher(None, doc_sent.lower(), source_sent.lower()).ratio()
if ratio > best_similarity:
best_similarity = ratio
best_source_idx = source_idx
if best_similarity >= threshold and best_source_idx >= 0:
if best_source_idx not in matched_source_indices:
matched_source_indices.add(best_source_idx)
matches.append({
'doc_text': doc_sentences[doc_idx],
'source_text': source_sentences[best_source_idx],
'similarity': float(best_similarity),
'doc_index': doc_idx,
'source_index': best_source_idx
})
return matches
def calculate_semantic_similarity(text1: str, text2: str) -> float:
"""Calculate semantic similarity between two texts"""
try:
encoder = get_encoder()
embeddings = encoder.encode([text1, text2], convert_to_numpy=True, show_progress_bar=False)
similarity = np.dot(embeddings[0], embeddings[1]) / (
np.linalg.norm(embeddings[0]) * np.linalg.norm(embeddings[1]) + 1e-8
)
return float(similarity)
except:
from difflib import SequenceMatcher
return SequenceMatcher(None, text1.lower(), text2.lower()).ratio()
def compare_semantic_chunks(
doc_text: str,
source_text: str,
chunk_size: int = 200,
threshold: float = 0.65
) -> List[Dict]:
"""
Compare document chunks against source chunks using semantic similarity.
"""
def chunk_text(text, size):
words = text.split()
chunks = []
for i in range(0, len(words), size):
chunk = ' '.join(words[i:i+size])
if len(chunk.split()) >= 20:
chunks.append(chunk)
return chunks
doc_chunks = chunk_text(doc_text, chunk_size)
source_chunks = chunk_text(source_text, chunk_size)
if not doc_chunks or not source_chunks:
return []
try:
encoder = get_encoder()
doc_embeddings = encoder.encode(doc_chunks, convert_to_numpy=True, show_progress_bar=False)
source_embeddings = encoder.encode(source_chunks, convert_to_numpy=True, show_progress_bar=False)
matches = []
for doc_emb in doc_embeddings:
for source_emb in source_embeddings:
similarity = np.dot(doc_emb, source_emb) / (
np.linalg.norm(doc_emb) * np.linalg.norm(source_emb) + 1e-8
)
if similarity >= threshold:
matches.append({
'doc_text': doc_chunks[0][:200] + "...",
'source_text': source_chunks[0][:200] + "...",
'similarity': float(similarity),
})
return matches
except:
return [] |