File size: 30,183 Bytes
3190e1e |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 |
from typing import Dict, List
import numpy as np
import tensorflow as tf
import tensorflow_hub as hub
import re
from pipeline_config import PipelineConfig
from quality_metrics import QualityMetrics
from paraphraser import Paraphraser
from back_translator import BackTranslator
import nlpaug.augmenter.word as naw
from concurrent.futures import ThreadPoolExecutor
from functools import lru_cache
from sklearn.metrics.pairwise import cosine_similarity
class DialogueAugmenter:
"""
Optimized dialogue augmentation with quality control and complexity management.
"""
def __init__(self, nlp, config: PipelineConfig):
self.nlp = nlp
self.config = config
self.quality_metrics = QualityMetrics(config)
self.use_model = hub.load('https://tfhub.dev/google/universal-sentence-encoder/4')
# Advanced augmentation techniques
self.paraphraser = Paraphraser()
self.back_translator = BackTranslator()
# Basic augmentation techniques
self.word_augmenter = naw.SynonymAug(aug_src='wordnet')
self.spelling_augmenter = naw.SpellingAug()
self.augmenters = {
'advanced': [self.paraphraser, self.back_translator],
'basic': [
('synonym', self.word_augmenter),
('spelling', self.spelling_augmenter)
]
}
# Initialize cache
self.embedding_cache = {}
self.perplexity_cache = {}
# Compile regex patterns
self.spelling_pattern = re.compile(r'[a-zA-Z]{3,}')
# GPU memory management
gpus = tf.config.list_physical_devices('GPU')
if gpus:
try:
for gpu in gpus:
tf.config.experimental.set_memory_growth(gpu, True)
except RuntimeError as e:
print(e)
@lru_cache(maxsize=1024)
def _compute_embedding(self, text: str) -> np.ndarray:
"""Cached computation of text embedding"""
return self.use_model([text])[0].numpy()
def _compute_batch_embeddings(self, texts: List[str]) -> np.ndarray:
"""Compute embeddings for multiple texts at once"""
return self.use_model(texts).numpy()
def _quick_quality_check(self, variation: str, original: str) -> bool:
"""
Simplified preliminary quality check with minimal standards
"""
if self.config.debug:
print(f"\nQuick check for variation: {variation}")
# Only reject if length is extremely different
orig_len = len(original.split())
var_len = len(variation.split())
# For very short texts (1-3 words), allow more variation
if orig_len <= 3:
if var_len > orig_len * 4: # Allow up to 4x length for short texts
if self.config.debug:
print(f"Failed length check (short text): {var_len} vs {orig_len}")
return False
else:
if var_len > orig_len * 3: # Allow up to 3x length for longer texts
if self.config.debug:
print(f"Failed length check (long text): {var_len} vs {orig_len}")
return False
# Basic content check - at least one word in common (excluding stop words)
stop_words = {'a', 'an', 'the', 'and', 'or', 'but', 'in', 'on', 'at', 'to', 'for', 'is', 'are'}
orig_words = set(w.lower() for w in original.split() if w.lower() not in stop_words)
var_words = set(w.lower() for w in variation.split() if w.lower() not in stop_words)
if not orig_words.intersection(var_words):
if self.config.debug:
print("Failed content check: no content words in common")
return False
if self.config.debug:
print("Passed all quick checks")
return True
def _compute_metrics_parallel(self, original: str, candidates: List[str]) -> List[Dict[str, float]]:
"""Compute quality metrics for multiple candidates in parallel"""
with ThreadPoolExecutor(max_workers=4) as executor:
futures = [
executor.submit(self.quality_metrics.compute_metrics, original, candidate)
for candidate in candidates
]
return [future.result() for future in futures]
def _filter_variations_batch(self, variations: List[str], context: List[str], original_turn: str) -> List[str]:
"""
Filter variations using batched computations with detailed logging
"""
if not variations:
return []
if self.config.debug:
print(f"\nStarting filtration of {len(variations)} variations")
print(f"Context length: {len(context)}")
print(f"Original turn: {original_turn}")
words = original_turn.split()
if len(words) < 3:
if self.config.debug:
print("Short text detected, using predefined variations")
short_text_variations = self._augment_short_text({'text': original_turn, 'speaker': ''})
return [var['text'] for var in short_text_variations]
# If this is the first turn (no context), be more lenient
if not context:
preliminary_filtered = variations
if self.config.debug:
print("First turn - skipping preliminary filtering")
else:
# Quick preliminary filtering against original turn
preliminary_filtered = []
for var in variations:
passed = self._quick_quality_check(var, original_turn)
if self.config.debug:
print(f"\nVariation: {var}")
print(f"Passed quick check: {passed}")
if passed:
preliminary_filtered.append(var)
if self.config.debug:
print(f"Variations after quick check: {len(preliminary_filtered)}")
if not preliminary_filtered:
return []
# Only use last turn for coherence
recent_context = [context[-1]] if context else []
context_text = ' '.join(recent_context) if recent_context else ''
# Even more lenient thresholds
min_similarity = 0.1 # Further reduced
min_coherence = 0.05 # Further reduced
if context_text:
if self.config.debug:
print(f"\nContext text: {context_text}")
all_texts = [context_text] + preliminary_filtered
all_embeddings = self._compute_batch_embeddings(all_texts)
context_embedding = all_embeddings[0]
variation_embeddings = all_embeddings[1:]
# Vectorized similarity computation
context_similarities = cosine_similarity([context_embedding], variation_embeddings)[0]
# Response coherence check
if recent_context:
prev_embedding = self._compute_embedding(recent_context[-1])
response_coherence = cosine_similarity([prev_embedding], variation_embeddings)[0]
else:
response_coherence = np.ones_like(context_similarities)
# Combined scoring with detailed logging
filtered_variations = []
for i, (variation, sim, coh) in enumerate(zip(
preliminary_filtered, context_similarities, response_coherence)):
# Use absolute values for scoring
combined_score = (
self.config.context_similarity_weight * abs(sim) +
self.config.response_coherence_weight * abs(coh)
)
if self.config.debug:
print(f"\nVariation: {variation}")
print(f"Context similarity: {sim:.3f}")
print(f"Response coherence: {coh:.3f}")
print(f"Combined score: {combined_score:.3f}")
# Accept if EITHER score is good enough
if (combined_score >= min_similarity or abs(coh) >= min_coherence):
filtered_variations.append(variation)
if self.config.debug:
print("ACCEPTED")
else:
if self.config.debug:
print("REJECTED")
# If we have enough variations, stop
if len(filtered_variations) >= self.config.max_variations_per_turn:
break
else:
filtered_variations = preliminary_filtered[:self.config.max_variations_per_turn]
if self.config.debug:
print(f"\nFinal filtered variations: {len(filtered_variations)}")
return filtered_variations
def _generate_variations_progressive(self, text: str, needed: int) -> List[str]:
"""
Generate variations progressively until we have enough good ones
"""
variations = set()
if self.config.debug:
print(f"\nAttempting to generate {needed} variations for text: {text}")
# Try advanced augmenters first
for augmenter in self.augmenters['advanced']:
if len(variations) >= needed:
break
try:
if isinstance(augmenter, Paraphraser):
if self.config.debug:
print("Trying paraphrase augmentation...")
new_vars = augmenter.paraphrase(text, num_return_sequences=needed-len(variations))
if self.config.debug:
print(f"Paraphraser generated {len(new_vars)} variations")
else:
if self.config.debug:
print("Trying back translation...")
new_vars = [augmenter.back_translate(text)]
if self.config.debug:
print(f"Back translator generated {len(new_vars)} variations")
valid_vars = [v for v in new_vars if v.strip() and v != text]
variations.update(valid_vars)
if self.config.debug:
print(f"Current unique variations: {len(variations)}")
except Exception as e:
print(f"Error in advanced augmentation: {str(e)}")
continue
# Try basic augmenters if needed
if len(variations) < needed:
if self.config.debug:
print("Not enough variations, trying basic augmenters...")
for aug_type, augmenter in self.augmenters['basic']:
if len(variations) >= needed:
break
try:
if aug_type == 'spelling' and self._is_technical_or_formal_text(text):
if self.config.debug:
print("Skipping spelling augmentation for technical text")
continue
if self.config.debug:
print(f"Trying {aug_type} augmentation...")
new_vars = augmenter.augment(text, n=2)
if isinstance(new_vars, list):
valid_vars = [v for v in new_vars if v.strip() and v != text]
variations.update(valid_vars)
else:
if new_vars.strip() and new_vars != text:
variations.add(new_vars)
if self.config.debug:
print(f"After {aug_type}, total variations: {len(variations)}")
except Exception as e:
print(f"Error in {aug_type} augmentation: {str(e)}")
continue
variations_list = list(variations)
if self.config.debug:
print(f"Final number of variations generated: {len(variations_list)}")
if not variations_list:
print("WARNING: No variations were generated!")
return variations_list
def augment_dialogue(self, dialogue: Dict) -> List[Dict]:
"""
Create augmented versions of the dialogue with optimized processing
"""
# Early dialogue length check
original_length = len(dialogue['turns'])
if original_length > self.config.max_turns_per_dialogue:
if self.config.debug:
print(f"Truncating dialogue from {original_length} to {self.config.max_turns_per_dialogue} turns")
dialogue['turns'] = dialogue['turns'][:self.config.max_turns_per_dialogue]
turn_variations = []
context = []
# Process each turn with progressive generation
for turn in dialogue['turns']:
original_text = turn['text'] # Store original turn text
variations = self._generate_variations_progressive(
original_text,
self.config.max_variations_per_turn
)
# Batch filter variations with original text
filtered_variations = self._filter_variations_batch(
variations,
context,
original_text # Pass the original turn text
)
# Create turn variations with speaker info
turn_vars = [{'speaker': turn['speaker'], 'text': v} for v in filtered_variations]
if self.config.debug:
print(f"Turn {len(turn_variations)}: Generated {len(turn_vars)} variations")
turn_variations.append(turn_vars)
context.append(original_text)
# Generate combinations with sampling
augmented_dialogues = self._generate_dialogue_combinations(
dialogue['dialogue_id'],
turn_variations
)
# Add original dialogue
result = [{
'dialogue_id': f"{dialogue['dialogue_id']}_original",
'turns': dialogue['turns']
}]
# Add unique augmentations
result.extend(augmented_dialogues[:self.config.augmentation_factor])
if self.config.debug:
print(f"Generated {len(result)-1} unique augmented dialogues")
return result
def _generate_dialogue_combinations(self, dialogue_id: str, turn_variations: List[List[Dict]]) -> List[Dict]:
"""
Generate dialogue combinations using sampling
"""
augmented_dialogues = []
used_combinations = set()
def generate_dialogues(current_turns=None, turn_index=0):
if current_turns is None:
current_turns = []
if len(augmented_dialogues) >= self.config.augmentation_factor:
return
if turn_index == len(turn_variations):
dialogue_fingerprint = " | ".join(turn['text'] for turn in current_turns)
if dialogue_fingerprint not in used_combinations:
used_combinations.add(dialogue_fingerprint)
augmented_dialogues.append({
'dialogue_id': f"{dialogue_id}_aug_{len(augmented_dialogues)}",
'turns': current_turns.copy()
})
return
variations = list(turn_variations[turn_index])
np.random.shuffle(variations)
for variation in variations[:self.config.max_sampled_variations]:
if len(augmented_dialogues) >= self.config.augmentation_factor:
return
current_turns.append(variation)
generate_dialogues(current_turns, turn_index + 1)
current_turns.pop()
try:
generate_dialogues()
except Exception as e:
print(f"Error in dialogue generation: {str(e)}")
return []
return augmented_dialogues
def _is_dialogue_duplicate(self, dialogue1: Dict, dialogue2: Dict) -> bool:
"""
Check if two dialogues are duplicates.
"""
text1 = " ".join(turn['text'] for turn in dialogue1['turns'])
text2 = " ".join(turn['text'] for turn in dialogue2['turns'])
return text1 == text2
# def _augment_turn(self, turn: Dict, context: List[str]) -> List[Dict]:
# """
# Generate augmented versions of the turn using multiple strategies.
# """
# text = turn['text']
# words = text.split()
# # Special handling for very short texts
# if len(words) < 3:
# return self._augment_short_text(turn)
# all_variations = set()
# # Advanced augmentations (paraphrase and back-translation)
# for augmenter in self.augmenters['advanced']:
# try:
# if isinstance(augmenter, Paraphraser):
# variations = augmenter.paraphrase(text)
# all_variations.update(variations)
# elif isinstance(augmenter, BackTranslator):
# aug_text = augmenter.back_translate(text)
# if aug_text:
# all_variations.add(aug_text)
# except Exception as e:
# print(f"Error in advanced augmentation: {str(e)}")
# continue
# # Basic nlpaug augmentations
# for aug_type, augmenter in self.augmenters['basic']:
# try:
# if aug_type == 'spelling' and self._is_technical_or_formal_text(text):
# continue
# aug_texts = augmenter.augment(text, n=2)
# if isinstance(aug_texts, list):
# all_variations.update(aug_texts)
# else:
# all_variations.add(aug_texts)
# except Exception as e:
# print(f"Error in {aug_type} augmentation: {str(e)}")
# continue
# # Remove exact duplicates and empty strings
# augmented_texts = [t for t in list(all_variations) if t.strip()]
# # Apply context filtering
# if context:
# augmented_texts = self._filter_by_context(augmented_texts, context)
# print(f"After context filtering: {len(augmented_texts)} variations")
# # Select best variations
# best_variations = self._select_best_augmentations(
# text,
# augmented_texts,
# num_to_select=self.config.augmentation_factor,
# min_quality_score=0.7
# )
# # Create variations with speaker info
# variations = [{'speaker': turn['speaker'], 'text': text} for text in best_variations]
# return variations
def _augment_short_text(self, turn: Dict) -> List[Dict]:
"""
Special handling for very short texts with predefined variations.
Args:
turn (Dict): Original dialogue turn
Returns:
List[Dict]: List of variations for the short text
"""
text = turn['text']
common_variations = {
'goodbye': [
'Bye!', 'Farewell!', 'See you!', 'Take care!',
'Goodbye!', 'Bye for now!', 'Until next time!'
],
'hello': [
'Hi!', 'Hey!', 'Hello!', 'Greetings!',
'Good day!', 'Hi there!', 'Hello there!'
],
'yes': [
'Yes!', 'Correct!', 'Indeed!', 'Absolutely!',
'That\'s right!', 'Definitely!', 'Sure!'
],
'no': [
'No!', 'Nope!', 'Not at all!', 'Negative!',
'Unfortunately not!', 'I\'m afraid not!'
],
'thanks': [
'Thank you!', 'Thanks a lot!', 'Many thanks!',
'I appreciate it!', 'Thank you so much!'
],
'ok': [
'Okay!', 'Alright!', 'Sure!', 'Got it!',
'Understood!', 'Fine!', 'Great!', 'Perfect!',
'That works!', 'Sounds good!'
],
'good': [
'Great!', 'Excellent!', 'Perfect!', 'Wonderful!',
'Fantastic!', 'Amazing!', 'Terrific!'
]
}
# Try to find matching variations
text_lower = text.lower().rstrip('!.,?')
variations = []
# Check if text matches any of our predefined categories
for key, predefined_vars in common_variations.items():
if key in text_lower or text_lower in key:
variations.extend(predefined_vars)
# If no predefined variations found, generate simple variants
if not variations:
# Add punctuation variations
variations = [
text.rstrip('!.,?') + '!',
text.rstrip('!.,?') + '.',
text.rstrip('!.,?')
]
# Add capitalization variations
variations.extend([
v.capitalize() for v in variations
if v.capitalize() not in variations
])
# Filter variations for uniqueness and quality
unique_variations = list(set(variations))
quality_variations = []
for var in unique_variations:
metrics = self.quality_metrics.compute_metrics(text, var)
quality_score = (
0.35 * metrics['semantic_similarity'] +
0.30 * (1.0 - metrics['perplexity'] / 100) +
0.15 * (1.0 - metrics['grammar_errors'] / 10) +
0.15 * metrics['content_preservation'] +
0.10 * metrics['type_token_ratio']
)
# More lenient quality threshold for short texts
if quality_score >= 0.5: # Lower threshold for short texts
quality_variations.append(var)
# Ensure we have at least some variations
if not quality_variations:
quality_variations = [text]
# Return the variations with original speaker
return [{'speaker': turn['speaker'], 'text': v} for v in quality_variations[:self.config.augmentation_factor]]
def _is_technical_or_formal_text(self, text: str) -> bool:
"""
Check if text is formal/technical and shouldn't have spelling variations.
"""
formal_indicators = {
'technical_terms': {'api', 'config', 'database', 'server', 'system'},
'formal_phrases': {'please advise', 'regarding', 'furthermore', 'moreover'},
'professional_context': {'meeting', 'conference', 'project', 'deadline'}
}
text_lower = text.lower()
words = set(text_lower.split())
for category in formal_indicators.values():
if words.intersection(category):
return True
return False
# def _filter_by_context(self, variations: List[str], context: List[str]) -> List[str]:
# """
# Filter variations based on conversation context using config parameters.
# """
# # Manage context window using config
# recent_context = context[-self.config.context_window_size:] if len(context) > self.config.context_window_size else context
# filtered_variations = []
# context_embedding = self.use_model([' '.join(recent_context)])[0].numpy()
# prev_turn = recent_context[-1] if recent_context else ''
# for variation in variations:
# var_embedding = self.use_model([variation])[0].numpy()
# # Overall context similarity
# context_similarity = cosine_similarity([context_embedding], [var_embedding])[0][0]
# # Direct response coherence
# response_coherence = 1.0
# if prev_turn:
# prev_embedding = self.use_model([prev_turn])[0].numpy()
# response_coherence = cosine_similarity([prev_embedding], [var_embedding])[0][0]
# # Use weights from config
# combined_similarity = (
# self.config.context_similarity_weight * context_similarity +
# self.config.response_coherence_weight * response_coherence
# )
# if (combined_similarity >= self.config.semantic_similarity_threshold and
# response_coherence >= self.config.min_response_coherence):
# filtered_variations.append(variation)
# if self.config.debug:
# print(f"Accepted variation: {variation}")
# print(f"Context similarity: {context_similarity:.3f}")
# print(f"Response coherence: {response_coherence:.3f}")
# print(f"Combined score: {combined_similarity:.3f}\n")
# else:
# if self.config.debug:
# print(f"Rejected variation: {variation}")
# print(f"Combined score {combined_similarity:.3f} below threshold "
# f"{self.config.semantic_similarity_threshold}")
# print(f"Response coherence {response_coherence:.3f} below threshold "
# f"{self.config.min_response_coherence}\n")
# return filtered_variations or variations # Fallback to original
# def _select_best_augmentations(self, original: str, candidates: List[str], used_variations: set = None,
# num_to_select: int = 3, min_quality_score: float = 0.7) -> List[str]:
# """
# Select the best augmentations using a quality score.
# Args:
# original (str): The original text
# candidates (List[str]): List of candidate augmented texts
# used_variations (set): Set of already used variations
# num_to_select (int): Number of variations to select
# min_quality_score (float): Minimum quality score threshold
# """
# if used_variations is None:
# used_variations = set()
# candidates = [c for c in candidates if c.strip()]
# # Skip short text
# if len(original.split()) < 3:
# print(f"Text too short for augmentation: {original}")
# return [original]
# scored_candidates = []
# for candidate in candidates:
# if candidate in used_variations:
# continue
# metrics = self.quality_metrics.compute_metrics(original, candidate)
# # Add contextual penalty for inappropriate audience terms
# audience_terms = {'everyone', 'everybody', 'folks', 'all', 'guys', 'people'}
# has_audience_term = any(term in candidate.lower() for term in audience_terms)
# audience_penalty = 0.2 if has_audience_term else 0.0
# # Weighted quality score
# quality_score = (
# 0.40 * metrics['semantic_similarity'] + # Semantic preservation
# 0.25 * (1.0 - metrics['perplexity'] / 100) + # Fluency
# 0.15 * (1.0 - metrics['grammar_errors'] / 10) + # Grammar
# 0.15 * metrics['content_preservation'] + # Content preservation
# 0.05 * metrics['type_token_ratio'] # Lexical diversity
# )
# quality_score -= audience_penalty
# if (metrics['semantic_similarity'] < 0.5 or # Reject on semantic threshold miss
# metrics['rouge1_f1'] < 0.2): # Enforce minimum lexical overlap
# continue
# # Bonus points for:
# # Length similarity to original
# if 0.75 <= metrics['length_ratio'] <= 1.25:
# quality_score += 0.05
# # Correct grammar
# if metrics['grammar_errors'] == 0:
# quality_score += 0.025
# print(f"Candidate: {candidate}")
# print(f"Quality score: {quality_score:.2f}, Metrics: {metrics}")
# # Consider the augmentationif meets basic quality threshold
# if quality_score >= min_quality_score:
# print('Candidate accepted\n')
# scored_candidates.append((candidate, quality_score, metrics))
# else:
# print('Candidate rejected\n')
# # Sort by quality score with small random factor for diversity
# scored_candidates.sort(key=lambda x: x[1], reverse=True)
# selected = []
# for candidate, score, metrics in scored_candidates:
# # Check diversity against already selected
# if len(selected) == 0:
# selected.append(candidate)
# continue
# # Compute average similarity to already selected
# avg_similarity = np.mean([
# self.quality_metrics.compute_semantic_similarity(candidate, prev)
# for prev in selected
# ])
# # Add if sufficiently different (similarity < 0.98)
# if avg_similarity < 0.98:
# selected.append(candidate)
# if len(selected) >= num_to_select:
# break
# return selected |