Spaces:
Sleeping
Sleeping
File size: 18,633 Bytes
c94f46f | 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 | # Methodology Logic Checker
# -------------------------
# The most subtle form of scientific fraud is not
# fabricating data β it is making claims that your
# method cannot actually support.
#
# "Correlation does not imply causation" is the
# famous example. But there are dozens of variations:
# underpowered claims, wrong timeframes, missing
# controls, self-report data for clinical conclusions.
#
# This module catches them β using both rule-based
# pattern matching and LLM-powered reasoning.
import re
import os
import json
import urllib.request
import urllib.error
from dataclasses import dataclass
from dotenv import load_dotenv
load_dotenv()
# ββ data structures βββββββββββββββββββββββββββββββββββββββββββ
@dataclass
class MethodologyFlag:
flag_type: str
severity: str
claim: str
issue: str
evidence: str
suggestion: str
@dataclass
class MethodologyResult:
claims_found: list
methods_found: list
flags: list
llm_assessment: str
llm_available: bool
risk_score: float
risk_level: str
summary: str
# ββ main class ββββββββββββββββββββββββββββββββββββββββββββββββ
class MethodologyChecker:
"""
Two-layer methodology analysis:
Layer 1 β Rule-based: fast, deterministic, catches
known patterns (causation language, tiny samples,
missing controls, short timeframes for long-term claims)
Layer 2 β LLM reasoning: slower, probabilistic,
catches subtle logical gaps that rules miss.
Falls back gracefully if API unavailable.
"""
# words that claim causation β need RCT to justify
CAUSATION_WORDS = [
"causes", "caused by", "leads to", "results in",
"produces", "induces", "drives", "responsible for",
"due to", "because of", "proves that", "demonstrates that"
]
# words that only justify correlation
CORRELATION_WORDS = [
"associated with", "correlated", "linked to",
"related to", "predicts", "suggests"
]
# study designs that cannot prove causation
WEAK_DESIGNS = [
"survey", "questionnaire", "self-report", "cross-sectional",
"retrospective", "observational", "case study", "anecdotal"
]
# long-term claims need long-term studies
LONGTERM_CLAIMS = [
"long-term", "chronic", "sustained", "permanent",
"lasting", "durable", "years", "lifetime"
]
def __init__(self):
self._hf_token = os.getenv("HF_API_TOKEN", "")
# free model on HuggingFace β good at reasoning
self._hf_model = "HuggingFaceH4/zephyr-7b-beta"
self._hf_api_url = (
f"https://api-inference.huggingface.co/models/{self._hf_model}"
)
# ββ public method βββββββββββββββββββββββββββββββββββββββββ
def analyze(self, text: str, abstract: str = "") -> MethodologyResult:
"""
Full methodology analysis.
Pass full paper text. Optionally pass abstract separately
for cleaner claim extraction.
"""
working_text = abstract if abstract else text
claims = self._extract_claims(working_text)
methods = self._extract_methods(text)
flags = []
flags.extend(self._check_causation_without_rct(text, claims))
flags.extend(self._check_weak_design_strong_claim(text, claims))
flags.extend(self._check_longterm_claim_shortterm_study(text, claims))
flags.extend(self._check_missing_control_group(text))
flags.extend(self._check_generalization(text, claims))
# try LLM reasoning β graceful fallback if unavailable
llm_text, llm_ok = self._llm_assess(
claims=claims,
methods=methods,
paper_snippet=text[:1500]
)
risk_score = self._calculate_risk(flags)
risk_level = self._get_risk_level(risk_score)
return MethodologyResult(
claims_found=claims,
methods_found=methods,
flags=flags,
llm_assessment=llm_text,
llm_available=llm_ok,
risk_score=round(risk_score, 3),
risk_level=risk_level,
summary=self._write_summary(flags, risk_level, llm_ok),
)
# ββ claim / method extraction βββββββββββββββββββββββββββββ
def _extract_claims(self, text: str) -> list:
"""
Pull claim-like sentences β those that assert findings,
conclusions, or implications.
Looks for language that signals a conclusion.
"""
claim_markers = [
"we found", "we show", "we demonstrate", "we conclude",
"our results", "our findings", "this study shows",
"this study demonstrates", "results indicate",
"results suggest", "data show", "analysis reveals",
"we report", "evidence suggests", "we establish"
]
claims = []
sentences = re.split(r'[.!?]', text)
for sentence in sentences:
sentence = sentence.strip()
if not sentence or len(sentence) < 20:
continue
s_lower = sentence.lower()
if any(marker in s_lower for marker in claim_markers):
claims.append(sentence)
return claims[:8] # cap at 8 β enough signal
def _extract_methods(self, text: str) -> list:
"""
Pull sentences from the methods section that describe
how the study was actually conducted.
"""
method_markers = [
"we used", "we conducted", "we recruited", "we collected",
"we measured", "we analyzed", "participants were",
"subjects were", "samples were", "data were collected",
"randomized", "controlled", "double-blind", "survey",
"questionnaire", "interview", "experiment"
]
methods = []
sentences = re.split(r'[.!?]', text)
for sentence in sentences:
sentence = sentence.strip()
if not sentence or len(sentence) < 20:
continue
s_lower = sentence.lower()
if any(marker in s_lower for marker in method_markers):
methods.append(sentence)
return methods[:8]
# ββ rule-based checks βββββββββββββββββββββββββββββββββββββ
def _check_causation_without_rct(self, text: str, claims: list) -> list:
"""
Detects causal language in claims combined with
study designs that cannot establish causation.
Classic example: observational study claiming X causes Y.
"""
flags = []
text_lower = text.lower()
has_causal_claim = any(
word in text_lower for word in self.CAUSATION_WORDS
)
has_weak_design = any(
design in text_lower for design in self.WEAK_DESIGNS
)
has_rct = any(
word in text_lower
for word in ["randomized", "randomised", "rct",
"control group", "placebo", "double-blind"]
)
if has_causal_claim and has_weak_design and not has_rct:
# find the actual causal sentence as evidence
evidence_sentence = ""
for sentence in re.split(r'[.!?]', text):
if any(w in sentence.lower() for w in self.CAUSATION_WORDS):
evidence_sentence = sentence.strip()
break
flags.append(MethodologyFlag(
flag_type="causation_without_rct",
severity="high",
claim="Causal language detected in conclusions",
issue=(
"The study uses causal language "
f"({', '.join([w for w in self.CAUSATION_WORDS if w in text_lower[:500]])}) "
"but the study design "
f"({', '.join([d for d in self.WEAK_DESIGNS if d in text_lower])}) "
"cannot establish causation."
),
evidence=evidence_sentence or "See causal language in conclusions",
suggestion=(
"Causal claims require randomized controlled trials. "
"Replace causal language with correlation language, "
"or acknowledge the design limitation explicitly."
),
))
return flags
def _check_weak_design_strong_claim(
self, text: str, claims: list
) -> list:
"""
Self-report surveys and questionnaires cannot support
strong clinical or behavioral conclusions.
"""
flags = []
text_lower = text.lower()
has_self_report = any(
w in text_lower
for w in ["self-report", "self report", "questionnaire",
"survey", "interview", "participants reported"]
)
has_strong_claim = any(
w in text_lower
for w in ["proves", "demonstrates", "establishes",
"confirms", "validates", "clinical evidence"]
)
if has_self_report and has_strong_claim:
flags.append(MethodologyFlag(
flag_type="weak_design_strong_claim",
severity="medium",
claim="Strong claim based on self-report data",
issue=(
"Self-report or questionnaire data has known limitations "
"(social desirability bias, recall bias) that undermine "
"strong conclusive claims."
),
evidence="Self-report instrument combined with conclusive language",
suggestion=(
"Acknowledge self-report limitations explicitly. "
"Soften conclusions to match data quality."
),
))
return flags
def _check_longterm_claim_shortterm_study(
self, text: str, claims: list
) -> list:
"""
Studies lasting days or weeks cannot make
long-term or chronic effect claims.
"""
flags = []
text_lower = text.lower()
has_longterm_claim = any(
w in text_lower for w in self.LONGTERM_CLAIMS
)
has_shortterm_study = any(
w in text_lower
for w in ["two weeks", "2 weeks", "one week", "1 week",
"3 days", "7 days", "short-term pilot",
"preliminary study"]
)
if has_longterm_claim and has_shortterm_study:
flags.append(MethodologyFlag(
flag_type="timeframe_mismatch",
severity="medium",
claim="Long-term claim from short-term study",
issue=(
"The study duration appears insufficient to support "
"long-term or chronic effect claims."
),
evidence="Long-term language with short study duration",
suggestion=(
"Either extend the study duration or explicitly "
"limit claims to short-term effects only."
),
))
return flags
def _check_missing_control_group(self, text: str) -> list:
"""
Studies measuring treatment effects without a
control group cannot isolate the treatment's impact.
"""
flags = []
text_lower = text.lower()
has_treatment = any(
w in text_lower
for w in ["treatment", "intervention", "drug", "therapy",
"program", "training"]
)
has_effect_claim = any(
w in text_lower
for w in ["improved", "reduced", "increased", "effective",
"significant effect"]
)
has_control = any(
w in text_lower
for w in ["control group", "control condition", "placebo",
"comparison group", "waitlist"]
)
if has_treatment and has_effect_claim and not has_control:
flags.append(MethodologyFlag(
flag_type="missing_control_group",
severity="high",
claim="Treatment effect claimed without control group",
issue=(
"Effect claims for a treatment or intervention "
"require a control group to rule out confounds, "
"placebo effects, and natural recovery."
),
evidence="Treatment + effect language with no control group mention",
suggestion=(
"Add a control/comparison condition, or acknowledge "
"that without a control group, the effect cannot be "
"attributed to the intervention specifically."
),
))
return flags
def _check_generalization(self, text: str, claims: list) -> list:
"""
Small, homogeneous samples cannot support
broad population-level generalizations.
"""
flags = []
text_lower = text.lower()
has_broad_claim = any(
w in text_lower
for w in ["all patients", "general population", "universally",
"across all", "globally applicable", "all humans"]
)
has_limited_sample = any(
w in text_lower
for w in ["undergraduate students", "college students",
"single institution", "convenience sample",
"homogeneous sample"]
)
if has_broad_claim and has_limited_sample:
flags.append(MethodologyFlag(
flag_type="overgeneralization",
severity="medium",
claim="Broad generalization from limited sample",
issue=(
"The sample characteristics (e.g., undergraduate students, "
"single institution) limit generalizability beyond "
"the studied population."
),
evidence="Broad claim language with limited sample description",
suggestion=(
"Explicitly acknowledge sampling limitations "
"and restrict claims to the studied population."
),
))
return flags
# ββ llm reasoning βββββββββββββββββββββββββββββββββββββββββ
def _llm_assess(
self, claims: list, methods: list, paper_snippet: str
) -> tuple:
"""
Ask an LLM to reason about whether the methods
logically support the claims.
Returns (assessment_text, success_bool).
Falls back gracefully if token missing or API down.
"""
if not self._hf_token or self._hf_token == "hf_xxxxxxxxxxxxxxxx":
return (
"LLM assessment unavailable β HF_API_TOKEN not configured.",
False
)
claims_text = " | ".join(claims[:3]) if claims else "Not extracted"
methods_text = " | ".join(methods[:3]) if methods else "Not extracted"
prompt = f"""[INST] You are a scientific peer reviewer.
Paper excerpt:
{paper_snippet[:800]}
Claims made: {claims_text}
Methods used: {methods_text}
In 2-3 sentences, identify the most critical logical gap between the methods and claims. Be specific and direct. [/INST]"""
try:
payload = json.dumps({
"inputs": prompt,
"parameters": {
"max_new_tokens": 150,
"temperature": 0.3,
"return_full_text": False,
}
}).encode("utf-8")
req = urllib.request.Request(
self._hf_api_url,
data=payload,
headers={
"Authorization": f"Bearer {self._hf_token}",
"Content-Type": "application/json",
},
method="POST"
)
with urllib.request.urlopen(req, timeout=30) as response:
result = json.loads(response.read().decode("utf-8"))
if isinstance(result, list) and result:
text = result[0].get("generated_text", "").strip()
return (text, True) if text else ("No assessment generated.", False)
return ("Unexpected API response format.", False)
except urllib.error.HTTPError as e:
if e.code == 503:
return ("LLM model loading β try again in 20 seconds.", False)
return (f"API error {e.code}: {str(e)}", False)
except Exception as e:
return (f"LLM unavailable: {str(e)}", False)
# ββ scoring βββββββββββββββββββββββββββββββββββββββββββββββ
def _calculate_risk(self, flags: list) -> float:
weights = {"high": 0.35, "medium": 0.20, "low": 0.08}
score = sum(weights.get(f.severity, 0) for f in flags)
return min(score, 1.0)
def _get_risk_level(self, score: float) -> str:
if score >= 0.7: return "critical"
elif score >= 0.4: return "high"
elif score >= 0.2: return "medium"
return "low"
def _write_summary(
self, flags: list, risk_level: str, llm_ok: bool
) -> str:
if not flags:
base = "No methodology logic issues detected."
else:
high = sum(1 for f in flags if f.severity == "high")
med = sum(1 for f in flags if f.severity == "medium")
parts = []
if high: parts.append(f"{high} high-severity issue{'s' if high > 1 else ''}")
if med: parts.append(f"{med} medium-severity concern{'s' if med > 1 else ''}")
base = (
f"Methodology checker flagged {', '.join(parts)}. "
f"Risk level: {risk_level.upper()}."
)
llm_note = " LLM reasoning included." if llm_ok else " LLM reasoning unavailable."
return base + llm_note |