File size: 17,553 Bytes
88519e8 bf39698 88519e8 d507c32 88519e8 d507c32 88519e8 66926c8 88519e8 bf39698 88519e8 d507c32 88519e8 d507c32 88519e8 bf39698 88519e8 bf39698 88519e8 d507c32 88519e8 bf39698 88519e8 bf39698 88519e8 bf39698 88519e8 bf39698 88519e8 bf39698 88519e8 bf39698 88519e8 bf39698 88519e8 bf39698 88519e8 bf39698 88519e8 bf39698 88519e8 bf39698 88519e8 bf39698 88519e8 dbdadad 88519e8 bf39698 88519e8 bf39698 88519e8 bf39698 88519e8 bf39698 88519e8 bf39698 88519e8 bf39698 d507c32 88519e8 d507c32 88519e8 d507c32 88519e8 |
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 |
"""
Human evaluation of recommendation explanations.
Generates 50 samples from the recommendation pipeline, presents them
interactively for Likert-scale rating, and computes aggregate metrics.
Dimensions (1-5 Likert scale):
Comprehension: "I understood why this item was recommended"
Trust: "I trust this explanation is accurate"
Usefulness: "This explanation helped me make a decision"
Satisfaction: "I am satisfied with this explanation"
Usage:
python scripts/human_eval.py --generate # Generate 50 samples
python scripts/human_eval.py --annotate # Rate samples (resumable)
python scripts/human_eval.py --analyze # Compute results
python scripts/human_eval.py --status # Show progress
Run from project root.
"""
import argparse
import json
import math
import sys
from datetime import datetime
from sage.core import AggregationMethod
from sage.config import (
DATA_DIR,
EVAL_DIMENSIONS,
EVALUATION_QUERIES,
HELPFULNESS_TARGET,
MAX_EVIDENCE,
RESULTS_DIR,
get_logger,
log_banner,
save_results,
)
logger = get_logger(__name__)
SAMPLES_DIR = DATA_DIR / "human_eval"
SAMPLES_FILE = SAMPLES_DIR / "samples.json"
TARGET_SAMPLES = 50
NATURAL_QUERIES_FILE = DATA_DIR / "eval" / "eval_natural_queries.json"
# ============================================================================
# Sample Generation
# ============================================================================
def _select_diverse_natural_queries(target: int = 35) -> list[str]:
"""Select diverse queries from natural eval dataset, balanced by category."""
if not NATURAL_QUERIES_FILE.exists():
logger.error(
"Natural queries file not found: %s "
"Run 'make eval' first to build eval datasets.",
NATURAL_QUERIES_FILE,
)
return []
with open(NATURAL_QUERIES_FILE, encoding="utf-8") as f:
data = json.load(f)
# Group by category
by_category: dict[str, list[str]] = {}
for item in data:
cat = item["category"]
by_category.setdefault(cat, []).append(item["query"])
if not by_category:
return []
# Round-robin across categories
selected = []
categories = sorted(by_category.keys())
max_cat_len = max(len(v) for v in by_category.values())
idx = 0
while len(selected) < target and idx < max_cat_len:
for cat in categories:
queries = by_category[cat]
if idx < len(queries) and len(selected) < target:
q = queries[idx]
if q not in selected:
selected.append(q)
idx += 1
return selected
def _select_config_queries(exclude: set[str], target: int = 15) -> list[str]:
"""Select queries from EVALUATION_QUERIES config, excluding duplicates."""
selected = []
for q in EVALUATION_QUERIES:
if q not in exclude and len(selected) < target:
selected.append(q)
return selected
def generate_samples(force: bool = False, seed: int = 42):
"""Generate recommendation+explanation samples for human evaluation."""
import random
from sage.services.retrieval import get_candidates
from sage.services import get_explanation_services
# Protect existing rated samples from accidental overwrite
if SAMPLES_FILE.exists() and not force:
with open(SAMPLES_FILE, encoding="utf-8") as f:
existing = json.load(f)
rated = sum(1 for s in existing if s.get("rating") is not None)
if rated > 0:
logger.error(
"%s contains %d rated samples. "
"Use --force to overwrite, or run --annotate to continue.",
SAMPLES_FILE,
rated,
)
sys.exit(1)
SAMPLES_DIR.mkdir(parents=True, exist_ok=True)
RESULTS_DIR.mkdir(parents=True, exist_ok=True)
log_banner(logger, "GENERATING HUMAN EVAL SAMPLES")
logger.info("Random seed: %d", seed)
# Set seed for reproducibility
random.seed(seed)
# Select diverse query set
natural = _select_diverse_natural_queries(35)
config = _select_config_queries(set(natural), 15)
all_queries = natural + config
# Shuffle with seeded random for reproducibility
random.shuffle(all_queries)
logger.info(
"Queries: %d natural + %d config = %d total",
len(natural),
len(config),
len(all_queries),
)
if len(all_queries) < TARGET_SAMPLES:
logger.error(
"Only %d unique queries available (target: %d). "
"Results will lack statistical power. "
"Run 'make eval' to build natural query dataset.",
len(all_queries),
TARGET_SAMPLES,
)
# Initialize services
explainer, detector = get_explanation_services()
samples = []
for i, query in enumerate(all_queries, 1):
logger.info('[%d/%d] "%s"', i, len(all_queries), query)
products = get_candidates(
query=query,
k=1,
min_rating=4.0,
aggregation=AggregationMethod.MAX,
)
if not products:
logger.info(" No products found, skipping")
continue
product = products[0]
try:
expl = explainer.generate_explanation(
query,
product,
max_evidence=MAX_EVIDENCE,
)
hhem = detector.check_explanation(
expl.evidence_texts,
expl.explanation,
)
sample = {
"id": len(samples) + 1,
"query": query,
"product_id": product.product_id,
"avg_rating": round(product.avg_rating, 1),
"explanation": expl.explanation,
"evidence": expl.to_evidence_dicts(),
"hhem_score": round(hhem.score, 4),
"rating": None,
}
samples.append(sample)
logger.info(
" %s (%.1f stars) HHEM=%.3f",
product.product_id,
product.avg_rating,
hhem.score,
)
except ValueError as exc:
logger.info(" Quality gate refusal: %s", exc)
except Exception:
logger.exception(" Error generating sample")
# Save
with open(SAMPLES_FILE, "w", encoding="utf-8") as f:
json.dump(samples, f, indent=2)
logger.info("Generated %d samples -> %s", len(samples), SAMPLES_FILE)
return samples
# ============================================================================
# Interactive Annotation
# ============================================================================
def _load_samples() -> list[dict]:
"""Load samples from disk."""
if not SAMPLES_FILE.exists():
logger.error("No samples file. Run --generate first.")
sys.exit(1)
with open(SAMPLES_FILE, encoding="utf-8") as f:
return json.load(f)
def _save_samples(samples: list[dict]):
"""Save samples back to disk."""
with open(SAMPLES_FILE, "w", encoding="utf-8") as f:
json.dump(samples, f, indent=2)
def _get_likert_input(prompt: str) -> int:
"""Prompt user for a 1-5 Likert rating. Returns rating or raises KeyboardInterrupt."""
while True:
try:
raw = input(f" {prompt} [1-5]: ").strip()
except EOFError:
raise KeyboardInterrupt
if raw in ("1", "2", "3", "4", "5"):
return int(raw)
print(" Enter a number from 1 to 5.")
def annotate_samples():
"""Interactive CLI loop for rating samples."""
samples = _load_samples()
total = len(samples)
rated = sum(1 for s in samples if s["rating"] is not None)
unrated = [s for s in samples if s["rating"] is None]
log_banner(logger, "HUMAN EVALUATION ANNOTATION")
print(f"\nProgress: {rated}/{total} rated, {len(unrated)} remaining\n")
if not unrated:
print("All samples have been rated. Run --analyze to compute results.")
return
print("Rate each dimension from 1 (strongly disagree) to 5 (strongly agree).")
print("Press Ctrl+C to save progress and quit.\n")
print("-" * 60)
try:
for sample in unrated:
rated = sum(1 for s in samples if s["rating"] is not None)
print(f"\n--- Sample {sample['id']} ({rated + 1}/{total}) ---\n")
# Display product and query
print(f"PRODUCT: {sample['product_id']} ({sample['avg_rating']} stars)")
print(f"QUERY: {sample['query']}\n")
# Display explanation
print(f"EXPLANATION:\n{sample['explanation']}\n")
# Display evidence (truncated)
print("EVIDENCE:")
for ev in sample["evidence"]:
text = ev["text"]
if len(text) > 200:
text = text[:200] + "..."
print(f' [{ev["id"]}]: "{text}"')
print()
# Collect ratings
rating = {}
for dim_key, dim_prompt in EVAL_DIMENSIONS.items():
rating[dim_key] = _get_likert_input(dim_prompt)
sample["rating"] = rating
_save_samples(samples)
scores_str = ", ".join(f"{k}={v}" for k, v in rating.items())
print(f" -> Saved ({scores_str})")
print("-" * 60)
except KeyboardInterrupt:
_save_samples(samples)
rated_now = sum(1 for s in samples if s["rating"] is not None)
print(f"\n\nProgress saved: {rated_now}/{total} rated.")
print("Run --annotate again to continue.")
# ============================================================================
# Analysis
# ============================================================================
def analyze_results():
"""Compute aggregate metrics from rated samples."""
samples = _load_samples()
rated = [s for s in samples if s["rating"] is not None]
log_banner(logger, "HUMAN EVALUATION ANALYSIS")
if not rated:
logger.error("No rated samples. Run --annotate first.")
return None
logger.info("Rated samples: %d/%d", len(rated), len(samples))
# Per-dimension statistics
dimensions_results = {}
for dim_key in EVAL_DIMENSIONS:
scores = [s["rating"][dim_key] for s in rated]
n = len(scores)
mean = sum(scores) / n
variance = sum((x - mean) ** 2 for x in scores) / (n - 1) if n > 1 else 0.0
std = variance**0.5
dimensions_results[dim_key] = {
"mean": round(mean, 2),
"std": round(std, 2),
"min": min(scores),
"max": max(scores),
}
logger.info(
" %-15s mean=%.2f std=%.2f range=[%d, %d]",
dim_key + ":",
mean,
std,
min(scores),
max(scores),
)
# Overall helpfulness: mean of per-sample averages
per_sample_means = []
for s in rated:
r = s["rating"]
sample_mean = sum(r[k] for k in EVAL_DIMENSIONS) / len(EVAL_DIMENSIONS)
per_sample_means.append(sample_mean)
overall = sum(per_sample_means) / len(per_sample_means)
passed = overall >= HELPFULNESS_TARGET
logger.info("")
logger.info(
"Overall helpfulness: %.2f (target: %.1f) [%s]",
overall,
HELPFULNESS_TARGET,
"PASS" if passed else "FAIL",
)
# HHEM vs Trust correlation (Spearman)
correlation = _compute_hhem_trust_correlation(rated)
if correlation:
logger.info(
"HHEM-Trust correlation: r=%.3f, p=%.4f",
correlation["spearman_r"],
correlation["p_value"],
)
# Save results
results = {
"timestamp": datetime.now().isoformat(),
"n_samples": len(rated),
"n_total": len(samples),
"methodology": {
"evaluator": "Single rater (developer/researcher)",
"instructions": "Rate each dimension 1-5 Likert: 1=strongly disagree, 5=strongly agree",
"dimensions": {
"comprehension": "I understood why this item was recommended",
"trust": "I trust this explanation is accurate",
"usefulness": "This explanation helped me make a decision",
"satisfaction": "I am satisfied with this explanation",
},
"sample_selection": "35 natural queries (balanced by category) + 15 config queries",
"inter_annotator_agreement": "N/A (single rater)",
},
"dimensions": dimensions_results,
"overall_helpfulness": round(overall, 2),
"target": HELPFULNESS_TARGET,
"pass": passed,
}
if correlation:
results["hhem_trust_correlation"] = correlation
ts_file = save_results(results, "human_eval")
logger.info("Saved: %s", ts_file)
return results
def _compute_hhem_trust_correlation(rated: list[dict]) -> dict | None:
"""Compute Spearman correlation between HHEM score and trust rating."""
hhem_scores = [s["hhem_score"] for s in rated]
trust_scores = [s["rating"]["trust"] for s in rated]
if len(set(hhem_scores)) < 2 or len(set(trust_scores)) < 2:
return None
try:
from scipy.stats import spearmanr
r, p = spearmanr(hhem_scores, trust_scores)
return {"spearman_r": round(float(r), 4), "p_value": round(float(p), 4)}
except ImportError:
# Fall back: compute rank correlation manually
return _manual_spearman(hhem_scores, trust_scores)
def _manual_spearman(x: list[float], y: list[float]) -> dict | None:
"""Rank-based Spearman without scipy."""
n = len(x)
if n < 3:
return None
def _rank(vals):
order = sorted(range(n), key=lambda i: vals[i])
ranks = [0.0] * n
i = 0
while i < n:
j = i
while j < n - 1 and vals[order[j + 1]] == vals[order[j]]:
j += 1
avg_rank = (i + j) / 2 + 1
for k in range(i, j + 1):
ranks[order[k]] = avg_rank
i = j + 1
return ranks
rx = _rank(x)
ry = _rank(y)
d_sq = sum((rx[i] - ry[i]) ** 2 for i in range(n))
rho = 1 - (6 * d_sq) / (n * (n**2 - 1))
# Approximate p-value via t-distribution (large sample)
if abs(rho) >= 1.0:
p = 0.0
else:
t = rho * math.sqrt((n - 2) / (1 - rho**2))
# Two-tailed p-value approximation
p = 2 * (1 - _t_cdf_approx(abs(t), n - 2))
return {"spearman_r": round(rho, 4), "p_value": round(max(p, 0.0), 4)}
def _t_cdf_approx(t: float, df: int) -> float:
"""Rough t-distribution CDF approximation (good enough for p < 0.05 checks)."""
# Regularized incomplete beta function approximation
# For df > 30, normal approximation is fine
if df > 30:
z = t * (1 - 1 / (4 * df))
return 0.5 * (1 + math.erf(z / math.sqrt(2)))
# For smaller df, use a rougher bound
return 0.5 * (1 + math.erf(t / math.sqrt(2 + t * t / df)))
# ============================================================================
# Status
# ============================================================================
def show_status():
"""Show annotation progress."""
if not SAMPLES_FILE.exists():
print("No samples generated yet. Run --generate first.")
return
samples = _load_samples()
total = len(samples)
rated = sum(1 for s in samples if s["rating"] is not None)
print(f"Human Evaluation Status: {rated}/{total} samples rated")
if rated == total:
print("All samples rated. Run --analyze to compute results.")
elif rated > 0:
print(f" {total - rated} remaining. Run --annotate to continue.")
else:
print(" No ratings yet. Run --annotate to start.")
# ============================================================================
# Main
# ============================================================================
def main():
parser = argparse.ArgumentParser(
description="Human evaluation of recommendation explanations",
)
group = parser.add_mutually_exclusive_group(required=True)
group.add_argument(
"--generate", action="store_true", help="Generate recommendation samples"
)
group.add_argument(
"--annotate", action="store_true", help="Rate samples interactively (resumable)"
)
group.add_argument(
"--analyze", action="store_true", help="Compute aggregate results from ratings"
)
group.add_argument("--status", action="store_true", help="Show annotation progress")
parser.add_argument(
"--force",
action="store_true",
help="Overwrite existing rated samples (with --generate)",
)
parser.add_argument(
"--seed",
type=int,
default=42,
help="Random seed for query selection (with --generate)",
)
args = parser.parse_args()
if args.force and not args.generate:
parser.error("--force can only be used with --generate")
if args.seed != 42 and not args.generate:
parser.error("--seed can only be used with --generate")
if args.generate:
generate_samples(force=args.force, seed=args.seed)
elif args.annotate:
annotate_samples()
elif args.analyze:
analyze_results()
elif args.status:
show_status()
if __name__ == "__main__":
main()
|