Update engine/parser_fusion.py
Browse files- engine/parser_fusion.py +263 -367
engine/parser_fusion.py
CHANGED
|
@@ -1,436 +1,332 @@
|
|
| 1 |
-
# engine/
|
| 2 |
# ------------------------------------------------------------
|
| 3 |
-
#
|
| 4 |
-
# Third parser head: repair & recovery
|
| 5 |
#
|
| 6 |
-
#
|
| 7 |
-
#
|
| 8 |
-
#
|
| 9 |
-
#
|
| 10 |
-
#
|
| 11 |
-
# -
|
| 12 |
-
#
|
| 13 |
-
#
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 14 |
# ------------------------------------------------------------
|
| 15 |
|
| 16 |
from __future__ import annotations
|
| 17 |
|
| 18 |
import json
|
| 19 |
import os
|
| 20 |
-
import
|
| 21 |
-
import re
|
| 22 |
-
from typing import Dict, Any, List, Optional
|
| 23 |
-
|
| 24 |
-
import torch
|
| 25 |
-
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
|
| 26 |
-
|
| 27 |
-
# ------------------------------------------------------------
|
| 28 |
-
# Model configuration
|
| 29 |
-
# ------------------------------------------------------------
|
| 30 |
-
|
| 31 |
-
DEFAULT_MODEL = os.getenv(
|
| 32 |
-
"BACTAI_LLM_PARSER_MODEL",
|
| 33 |
-
"google/flan-t5-base",
|
| 34 |
-
)
|
| 35 |
-
|
| 36 |
-
# You asked to raise snapshots to 15
|
| 37 |
-
MAX_FEWSHOT_EXAMPLES = int(os.getenv("BACTAI_LLM_FEWSHOT", "25"))
|
| 38 |
|
| 39 |
-
|
|
|
|
| 40 |
|
| 41 |
-
#
|
| 42 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 43 |
|
| 44 |
-
|
|
|
|
| 45 |
|
| 46 |
-
|
| 47 |
-
|
| 48 |
-
_GOLD_EXAMPLES: Optional[List[Dict[str, Any]]] = None
|
| 49 |
-
|
| 50 |
-
# ------------------------------------------------------------
|
| 51 |
-
# Allowed fields
|
| 52 |
-
# ------------------------------------------------------------
|
| 53 |
-
|
| 54 |
-
ALL_FIELDS: List[str] = [
|
| 55 |
-
"Gram Stain",
|
| 56 |
-
"Shape",
|
| 57 |
-
"Motility",
|
| 58 |
-
"Capsule",
|
| 59 |
-
"Spore Formation",
|
| 60 |
-
"Haemolysis",
|
| 61 |
-
"Haemolysis Type",
|
| 62 |
-
"Media Grown On",
|
| 63 |
-
"Colony Morphology",
|
| 64 |
-
"Oxygen Requirement",
|
| 65 |
-
"Growth Temperature",
|
| 66 |
-
"Catalase",
|
| 67 |
-
"Oxidase",
|
| 68 |
-
"Indole",
|
| 69 |
-
"Urease",
|
| 70 |
-
"Citrate",
|
| 71 |
-
"Methyl Red",
|
| 72 |
-
"VP",
|
| 73 |
-
"H2S",
|
| 74 |
-
"DNase",
|
| 75 |
-
"ONPG",
|
| 76 |
-
"Coagulase",
|
| 77 |
-
"Gelatin Hydrolysis",
|
| 78 |
-
"Esculin Hydrolysis",
|
| 79 |
-
"Nitrate Reduction",
|
| 80 |
-
"NaCl Tolerant (>=6%)",
|
| 81 |
-
"Lipase Test",
|
| 82 |
-
"Lysine Decarboxylase",
|
| 83 |
-
"Ornithine Decarboxylase",
|
| 84 |
-
"Ornitihine Decarboxylase",
|
| 85 |
-
"Arginine dihydrolase",
|
| 86 |
-
"Glucose Fermentation",
|
| 87 |
-
"Lactose Fermentation",
|
| 88 |
-
"Sucrose Fermentation",
|
| 89 |
-
"Maltose Fermentation",
|
| 90 |
-
"Mannitol Fermentation",
|
| 91 |
-
"Sorbitol Fermentation",
|
| 92 |
-
"Xylose Fermentation",
|
| 93 |
-
"Rhamnose Fermentation",
|
| 94 |
-
"Arabinose Fermentation",
|
| 95 |
-
"Raffinose Fermentation",
|
| 96 |
-
"Trehalose Fermentation",
|
| 97 |
-
"Inositol Fermentation",
|
| 98 |
-
"Gas Production",
|
| 99 |
-
"TSI Pattern",
|
| 100 |
-
"Colony Pattern",
|
| 101 |
-
"Pigment",
|
| 102 |
-
"Motility Type",
|
| 103 |
-
"Odor",
|
| 104 |
-
]
|
| 105 |
-
|
| 106 |
-
SUGAR_FIELDS = [
|
| 107 |
-
"Glucose Fermentation",
|
| 108 |
-
"Lactose Fermentation",
|
| 109 |
-
"Sucrose Fermentation",
|
| 110 |
-
"Maltose Fermentation",
|
| 111 |
-
"Mannitol Fermentation",
|
| 112 |
-
"Sorbitol Fermentation",
|
| 113 |
-
"Xylose Fermentation",
|
| 114 |
-
"Rhamnose Fermentation",
|
| 115 |
-
"Arabinose Fermentation",
|
| 116 |
-
"Raffinose Fermentation",
|
| 117 |
-
"Trehalose Fermentation",
|
| 118 |
-
"Inositol Fermentation",
|
| 119 |
-
]
|
| 120 |
-
|
| 121 |
-
PNV_FIELDS = set(
|
| 122 |
-
f for f in ALL_FIELDS
|
| 123 |
-
if f not in {
|
| 124 |
-
"Media Grown On",
|
| 125 |
-
"Colony Morphology",
|
| 126 |
-
"Growth Temperature",
|
| 127 |
-
"Gram Stain",
|
| 128 |
-
"Shape",
|
| 129 |
-
"Oxygen Requirement",
|
| 130 |
-
"Haemolysis Type",
|
| 131 |
-
}
|
| 132 |
-
)
|
| 133 |
-
|
| 134 |
-
# ------------------------------------------------------------
|
| 135 |
-
# Field alias mapping (CRITICAL)
|
| 136 |
-
# ------------------------------------------------------------
|
| 137 |
|
| 138 |
-
FIELD_ALIASES: Dict[str, str] = {
|
| 139 |
-
# Gram
|
| 140 |
-
"Gram": "Gram Stain",
|
| 141 |
-
"Gram stain": "Gram Stain",
|
| 142 |
-
"Gram Stain Result": "Gram Stain",
|
| 143 |
-
|
| 144 |
-
# Salt tolerance
|
| 145 |
-
"NaCl tolerance": "NaCl Tolerant (>=6%)",
|
| 146 |
-
"NaCl Tolerant": "NaCl Tolerant (>=6%)",
|
| 147 |
-
"Salt tolerance": "NaCl Tolerant (>=6%)",
|
| 148 |
-
"Salt tolerant": "NaCl Tolerant (>=6%)",
|
| 149 |
-
"6.5% NaCl": "NaCl Tolerant (>=6%)",
|
| 150 |
-
"6% NaCl": "NaCl Tolerant (>=6%)",
|
| 151 |
-
|
| 152 |
-
# Temperature
|
| 153 |
-
"Growth temp": "Growth Temperature",
|
| 154 |
-
"Growth temperature": "Growth Temperature",
|
| 155 |
-
"Temperature growth": "Growth Temperature",
|
| 156 |
-
|
| 157 |
-
# Tests
|
| 158 |
-
"Catalase test": "Catalase",
|
| 159 |
-
"Oxidase test": "Oxidase",
|
| 160 |
-
"Indole test": "Indole",
|
| 161 |
-
"Urease test": "Urease",
|
| 162 |
-
"Citrate test": "Citrate",
|
| 163 |
-
|
| 164 |
-
# Sugars (common lowercase variants)
|
| 165 |
-
"Glucose fermentation": "Glucose Fermentation",
|
| 166 |
-
"Lactose fermentation": "Lactose Fermentation",
|
| 167 |
-
"Sucrose fermentation": "Sucrose Fermentation",
|
| 168 |
-
"Maltose fermentation": "Maltose Fermentation",
|
| 169 |
-
"Mannitol fermentation": "Mannitol Fermentation",
|
| 170 |
-
"Sorbitol fermentation": "Sorbitol Fermentation",
|
| 171 |
-
"Xylose fermentation": "Xylose Fermentation",
|
| 172 |
-
"Rhamnose fermentation": "Rhamnose Fermentation",
|
| 173 |
-
"Arabinose fermentation": "Arabinose Fermentation",
|
| 174 |
-
"Raffinose fermentation": "Raffinose Fermentation",
|
| 175 |
-
"Trehalose fermentation": "Trehalose Fermentation",
|
| 176 |
-
"Inositol fermentation": "Inositol Fermentation",
|
| 177 |
-
}
|
| 178 |
|
| 179 |
# ------------------------------------------------------------
|
| 180 |
-
#
|
| 181 |
# ------------------------------------------------------------
|
| 182 |
|
| 183 |
-
def
|
| 184 |
-
return str(s).strip() if s is not None else ""
|
| 185 |
-
|
| 186 |
-
def _normalise_pnv_value(raw: Any) -> str:
|
| 187 |
"""
|
| 188 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 189 |
"""
|
| 190 |
-
|
| 191 |
-
|
| 192 |
-
return "Unknown"
|
| 193 |
-
|
| 194 |
-
if any(x in s for x in {"positive", "pos", "+", "yes", "present", "detected", "reactive"}):
|
| 195 |
-
return "Positive"
|
| 196 |
-
|
| 197 |
-
if any(x in s for x in {"negative", "neg", "-", "no", "none", "absent", "not detected", "no growth"}):
|
| 198 |
-
return "Negative"
|
| 199 |
-
|
| 200 |
-
if "variable" in s or "mixed" in s or "inconsistent" in s:
|
| 201 |
-
return "Variable"
|
| 202 |
-
|
| 203 |
-
return "Unknown"
|
| 204 |
|
| 205 |
-
|
| 206 |
-
|
| 207 |
-
|
| 208 |
-
|
| 209 |
-
|
| 210 |
-
return
|
| 211 |
-
|
| 212 |
-
return
|
| 213 |
-
return "Unknown"
|
| 214 |
|
| 215 |
-
def _merge_ornithine_variants(fields: Dict[str, str]) -> Dict[str, str]:
|
| 216 |
-
v = fields.get("Ornithine Decarboxylase") or fields.get("Ornitihine Decarboxylase")
|
| 217 |
-
if v and v != "Unknown":
|
| 218 |
-
fields["Ornithine Decarboxylase"] = v
|
| 219 |
-
fields["Ornitihine Decarboxylase"] = v
|
| 220 |
-
return fields
|
| 221 |
|
| 222 |
-
|
| 223 |
-
|
| 224 |
-
# ------------------------------------------------------------
|
| 225 |
|
| 226 |
-
_NON_FERMENTER_PATTERNS = re.compile(
|
| 227 |
-
r"\b(non[-\s]?fermenter|non[-\s]?fermentative|asaccharolytic|"
|
| 228 |
-
r"does not ferment (sugars|carbohydrates)|no carbohydrate fermentation)\b",
|
| 229 |
-
re.IGNORECASE,
|
| 230 |
-
)
|
| 231 |
|
| 232 |
-
def
|
| 233 |
"""
|
| 234 |
-
|
| 235 |
-
|
| 236 |
"""
|
| 237 |
-
|
| 238 |
-
|
| 239 |
-
|
| 240 |
-
for sugar in SUGAR_FIELDS:
|
| 241 |
-
current = fields.get(sugar)
|
| 242 |
-
if current in {"Positive", "Variable"}:
|
| 243 |
-
continue
|
| 244 |
-
fields[sugar] = "Negative"
|
| 245 |
-
|
| 246 |
-
return fields
|
| 247 |
|
| 248 |
-
|
| 249 |
-
|
| 250 |
-
|
| 251 |
|
| 252 |
-
|
| 253 |
-
return os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
|
| 254 |
|
| 255 |
-
def _load_gold_examples() -> List[Dict[str, Any]]:
|
| 256 |
-
global _GOLD_EXAMPLES
|
| 257 |
-
if _GOLD_EXAMPLES is not None:
|
| 258 |
-
return _GOLD_EXAMPLES
|
| 259 |
|
| 260 |
-
|
| 261 |
-
|
| 262 |
-
|
| 263 |
-
|
| 264 |
-
|
| 265 |
-
|
| 266 |
-
|
| 267 |
-
|
| 268 |
-
|
| 269 |
-
|
| 270 |
-
# ------------------------------------------------------------
|
| 271 |
-
# Prompt
|
| 272 |
-
# ------------------------------------------------------------
|
| 273 |
|
| 274 |
-
PROMPT_HEADER = """
|
| 275 |
-
You are a microbiology expert assisting an automated phenotype parser.
|
| 276 |
|
| 277 |
-
|
| 278 |
-
|
|
|
|
| 279 |
|
| 280 |
-
|
| 281 |
-
|
| 282 |
-
|
| 283 |
-
|
| 284 |
-
|
| 285 |
|
| 286 |
-
|
| 287 |
-
|
| 288 |
-
|
| 289 |
-
"
|
| 290 |
-
|
| 291 |
-
|
| 292 |
-
|
| 293 |
-
"""
|
| 294 |
|
| 295 |
-
|
| 296 |
-
|
| 297 |
|
| 298 |
-
|
| 299 |
-
\"\"\"<<PHENOTYPE>>\"\"\"
|
| 300 |
|
| 301 |
-
|
| 302 |
-
|
|
|
|
|
|
|
|
|
|
| 303 |
|
| 304 |
-
|
| 305 |
-
|
| 306 |
-
|
| 307 |
-
|
| 308 |
|
| 309 |
-
|
|
|
|
| 310 |
|
| 311 |
-
|
| 312 |
-
|
| 313 |
-
exp = ex.get("expected", {})
|
| 314 |
-
if not isinstance(exp, dict):
|
| 315 |
-
exp = {}
|
| 316 |
-
blocks.append(
|
| 317 |
-
f'Input:\n"""{inp}"""\nOutput:\n'
|
| 318 |
-
f'{json.dumps({"parsed_fields": exp}, ensure_ascii=False)}\n'
|
| 319 |
-
)
|
| 320 |
|
| 321 |
-
|
| 322 |
-
|
| 323 |
|
| 324 |
-
|
| 325 |
-
# Model loader
|
| 326 |
-
# ------------------------------------------------------------
|
| 327 |
|
| 328 |
-
def _load_model() -> None:
|
| 329 |
-
global _model, _tokenizer
|
| 330 |
-
if _model is not None and _tokenizer is not None:
|
| 331 |
-
return
|
| 332 |
-
_tokenizer = AutoTokenizer.from_pretrained(DEFAULT_MODEL)
|
| 333 |
-
_model = AutoModelForSeq2SeqLM.from_pretrained(DEFAULT_MODEL).to(DEVICE)
|
| 334 |
-
_model.eval()
|
| 335 |
|
| 336 |
# ------------------------------------------------------------
|
| 337 |
-
#
|
| 338 |
# ------------------------------------------------------------
|
| 339 |
|
| 340 |
-
|
| 341 |
-
|
| 342 |
-
def _extract_first_json_object(text: str) -> Dict[str, Any]:
|
| 343 |
"""
|
| 344 |
-
|
| 345 |
"""
|
| 346 |
-
|
| 347 |
-
|
| 348 |
-
|
| 349 |
-
|
| 350 |
-
return
|
| 351 |
-
|
| 352 |
-
return
|
|
|
|
| 353 |
|
| 354 |
-
|
|
|
|
| 355 |
"""
|
| 356 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 357 |
"""
|
| 358 |
-
|
| 359 |
-
|
| 360 |
-
key = _norm_str(k)
|
| 361 |
-
if not key:
|
| 362 |
-
continue
|
| 363 |
-
mapped = FIELD_ALIASES.get(key, key)
|
| 364 |
-
out[mapped] = v
|
| 365 |
-
return out
|
| 366 |
|
| 367 |
-
|
| 368 |
-
|
| 369 |
-
# ------------------------------------------------------------
|
| 370 |
|
| 371 |
-
|
| 372 |
-
|
| 373 |
-
if not original.strip():
|
| 374 |
-
return {"parsed_fields": {}, "source": "llm_parser", "raw": original}
|
| 375 |
|
| 376 |
-
|
| 377 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 378 |
|
| 379 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 380 |
|
| 381 |
-
|
| 382 |
-
inputs = _tokenizer(prompt, return_tensors="pt", truncation=True).to(DEVICE)
|
| 383 |
|
| 384 |
-
|
| 385 |
-
|
| 386 |
-
**inputs,
|
| 387 |
-
max_new_tokens=MAX_NEW_TOKENS,
|
| 388 |
-
do_sample=False,
|
| 389 |
-
temperature=0.0,
|
| 390 |
-
)
|
| 391 |
|
| 392 |
-
|
|
|
|
| 393 |
|
| 394 |
-
|
| 395 |
-
|
| 396 |
-
|
| 397 |
-
|
|
|
|
| 398 |
|
| 399 |
-
|
| 400 |
-
|
| 401 |
-
if not isinstance(fields_raw, dict):
|
| 402 |
-
fields_raw = {}
|
| 403 |
|
| 404 |
-
|
| 405 |
-
|
|
|
|
| 406 |
|
| 407 |
-
|
| 408 |
-
|
| 409 |
-
try:
|
| 410 |
-
print(json.dumps(fields_raw, indent=2, ensure_ascii=False))
|
| 411 |
-
except Exception:
|
| 412 |
-
print(fields_raw)
|
| 413 |
-
print("===============================")
|
| 414 |
|
| 415 |
-
|
|
|
|
|
|
|
|
|
|
| 416 |
|
| 417 |
-
|
| 418 |
-
|
| 419 |
-
continue
|
| 420 |
-
raw_val = fields_raw[field]
|
| 421 |
|
| 422 |
-
|
| 423 |
-
cleaned[field] = _normalise_gram(raw_val)
|
| 424 |
-
elif field in PNV_FIELDS:
|
| 425 |
-
cleaned[field] = _normalise_pnv_value(raw_val)
|
| 426 |
-
else:
|
| 427 |
-
cleaned[field] = _norm_str(raw_val) or "Unknown"
|
| 428 |
|
| 429 |
-
|
| 430 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 431 |
|
| 432 |
return {
|
| 433 |
-
"
|
| 434 |
-
"
|
| 435 |
-
"
|
| 436 |
-
|
|
|
|
|
|
| 1 |
+
# engine/parser_fusion.py
|
| 2 |
# ------------------------------------------------------------
|
| 3 |
+
# Tri-Parser Fusion — Stage 12B (Weighted, SOTA-style)
|
|
|
|
| 4 |
#
|
| 5 |
+
# This module combines:
|
| 6 |
+
# - Rule parser (parser_rules.parse_text_rules)
|
| 7 |
+
# - Extended parser (parser_ext.parse_text_extended)
|
| 8 |
+
# - LLM parser (parser_llm.parse_llm) [optional]
|
| 9 |
+
#
|
| 10 |
+
# using per-field reliability weights learned in Stage 12A
|
| 11 |
+
# and stored in:
|
| 12 |
+
# data/field_weights.json
|
| 13 |
+
#
|
| 14 |
+
# Behaviour:
|
| 15 |
+
# - For each field, gather predictions from available parsers.
|
| 16 |
+
# - For that field, load weights:
|
| 17 |
+
# field_weights[field] (if present)
|
| 18 |
+
# else global weights
|
| 19 |
+
# else equal weights across available parsers
|
| 20 |
+
# - Discard parsers that:
|
| 21 |
+
# * did not predict the field
|
| 22 |
+
# * or only predicted "Unknown"
|
| 23 |
+
# - Group by predicted value and sum the weights of parsers
|
| 24 |
+
# that voted for each value.
|
| 25 |
+
# - Choose the value with highest total weight.
|
| 26 |
+
# Tie-break: prefer rules > extended > llm if needed.
|
| 27 |
+
#
|
| 28 |
+
# Output format:
|
| 29 |
+
# {
|
| 30 |
+
# "fused_fields": { field: value, ... }, # used by DB identifier AND genus ML
|
| 31 |
+
# "by_parser": {
|
| 32 |
+
# "rules": { ... },
|
| 33 |
+
# "extended": { ... },
|
| 34 |
+
# "llm": { ... } # may be empty
|
| 35 |
+
# },
|
| 36 |
+
# "votes": {
|
| 37 |
+
# field_name: {
|
| 38 |
+
# "per_parser": {
|
| 39 |
+
# "rules": {"value": "Positive", "weight": 0.95},
|
| 40 |
+
# "extended": {"value": "Unknown", "weight": 0.03},
|
| 41 |
+
# ...
|
| 42 |
+
# },
|
| 43 |
+
# "summed": {
|
| 44 |
+
# "Positive": 0.97,
|
| 45 |
+
# "Negative": 0.02
|
| 46 |
+
# },
|
| 47 |
+
# "chosen": "Positive"
|
| 48 |
+
# },
|
| 49 |
+
# ...
|
| 50 |
+
# },
|
| 51 |
+
# "weights_meta": {
|
| 52 |
+
# "has_weights_file": True/False,
|
| 53 |
+
# "weights_path": "data/field_weights.json",
|
| 54 |
+
# "meta": { ... } # from file if present
|
| 55 |
+
# }
|
| 56 |
+
# }
|
| 57 |
# ------------------------------------------------------------
|
| 58 |
|
| 59 |
from __future__ import annotations
|
| 60 |
|
| 61 |
import json
|
| 62 |
import os
|
| 63 |
+
from typing import Any, Dict, Optional
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 64 |
|
| 65 |
+
from engine.parser_rules import parse_text_rules
|
| 66 |
+
from engine.parser_ext import parse_text_extended
|
| 67 |
|
| 68 |
+
# Optional LLM parser
|
| 69 |
+
try:
|
| 70 |
+
from engine.parser_llm import parse_llm as parse_text_llm # type: ignore
|
| 71 |
+
HAS_LLM = True
|
| 72 |
+
except Exception:
|
| 73 |
+
parse_text_llm = None # type: ignore
|
| 74 |
+
HAS_LLM = False
|
| 75 |
|
| 76 |
+
# Path to learned weights
|
| 77 |
+
FIELD_WEIGHTS_PATH = os.path.join("data", "field_weights.json")
|
| 78 |
|
| 79 |
+
UNKNOWN = "Unknown"
|
| 80 |
+
PARSER_ORDER = ["rules", "extended", "llm"] # used for tie-breaking
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 81 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 82 |
|
| 83 |
# ------------------------------------------------------------
|
| 84 |
+
# Weights loading and helpers
|
| 85 |
# ------------------------------------------------------------
|
| 86 |
|
| 87 |
+
def _load_field_weights(path: str = FIELD_WEIGHTS_PATH) -> Dict[str, Any]:
|
|
|
|
|
|
|
|
|
|
| 88 |
"""
|
| 89 |
+
Load the JSON weights file produced by Stage 12A.
|
| 90 |
+
|
| 91 |
+
Expected structure:
|
| 92 |
+
{
|
| 93 |
+
"global": { "rules": 0.7, "extended": 0.2, "llm": 0.1 },
|
| 94 |
+
"fields": {
|
| 95 |
+
"DNase": {
|
| 96 |
+
"rules": 0.95,
|
| 97 |
+
"extended": 0.03,
|
| 98 |
+
"llm": 0.02,
|
| 99 |
+
"support": 123
|
| 100 |
+
},
|
| 101 |
+
...
|
| 102 |
+
},
|
| 103 |
+
"meta": { ... }
|
| 104 |
+
}
|
| 105 |
+
|
| 106 |
+
If the file is missing or broken, we fall back to an empty dict,
|
| 107 |
+
which triggers equal-weight behaviour later.
|
| 108 |
"""
|
| 109 |
+
if not os.path.exists(path):
|
| 110 |
+
return {}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 111 |
|
| 112 |
+
try:
|
| 113 |
+
with open(path, "r", encoding="utf-8") as f:
|
| 114 |
+
obj = json.load(f)
|
| 115 |
+
if isinstance(obj, dict):
|
| 116 |
+
return obj
|
| 117 |
+
return {}
|
| 118 |
+
except Exception:
|
| 119 |
+
return {}
|
|
|
|
| 120 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 121 |
|
| 122 |
+
FIELD_WEIGHTS_RAW: Dict[str, Any] = _load_field_weights()
|
| 123 |
+
HAS_WEIGHTS_FILE: bool = bool(FIELD_WEIGHTS_RAW)
|
|
|
|
| 124 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 125 |
|
| 126 |
+
def _normalise_scores(scores: Dict[str, float]) -> Dict[str, float]:
|
| 127 |
"""
|
| 128 |
+
Normalise a dict of parser -> score into weights summing to 1.
|
| 129 |
+
If all scores are zero or dict is empty, return equal weights.
|
| 130 |
"""
|
| 131 |
+
cleaned = {k: max(0.0, float(v)) for k, v in scores.items()}
|
| 132 |
+
total = sum(cleaned.values())
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 133 |
|
| 134 |
+
if total <= 0:
|
| 135 |
+
n = len(cleaned) or 1
|
| 136 |
+
return {k: 1.0 / n for k in cleaned}
|
| 137 |
|
| 138 |
+
return {k: v / total for k, v in cleaned.items()}
|
|
|
|
| 139 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 140 |
|
| 141 |
+
def _get_base_weights_for_parsers(include_llm: bool) -> Dict[str, float]:
|
| 142 |
+
"""
|
| 143 |
+
Get a naive equal-weight distribution across available parsers.
|
| 144 |
+
Used when no learned weights are available.
|
| 145 |
+
"""
|
| 146 |
+
parsers = ["rules", "extended"]
|
| 147 |
+
if include_llm:
|
| 148 |
+
parsers.append("llm")
|
| 149 |
+
n = len(parsers) or 1
|
| 150 |
+
return {p: 1.0 / n for p in parsers}
|
|
|
|
|
|
|
|
|
|
| 151 |
|
|
|
|
|
|
|
| 152 |
|
| 153 |
+
def _get_weights_for_field(field_name: str, include_llm: bool) -> Dict[str, float]:
|
| 154 |
+
"""
|
| 155 |
+
Get weights for a specific field.
|
| 156 |
|
| 157 |
+
Priority:
|
| 158 |
+
1) If FIELD_WEIGHTS_RAW has a 'fields[field_name]' entry,
|
| 159 |
+
use that.
|
| 160 |
+
2) Else if FIELD_WEIGHTS_RAW has 'global', use that.
|
| 161 |
+
3) Else equal weights.
|
| 162 |
|
| 163 |
+
In all cases:
|
| 164 |
+
- Drop 'llm' if include_llm == False
|
| 165 |
+
- Normalise
|
| 166 |
+
"""
|
| 167 |
+
if not FIELD_WEIGHTS_RAW:
|
| 168 |
+
base = _get_base_weights_for_parsers(include_llm)
|
| 169 |
+
return _normalise_scores(base)
|
|
|
|
| 170 |
|
| 171 |
+
fields_block = FIELD_WEIGHTS_RAW.get("fields", {}) or {}
|
| 172 |
+
global_block = FIELD_WEIGHTS_RAW.get("global", {}) or {}
|
| 173 |
|
| 174 |
+
raw: Dict[str, float] = {}
|
|
|
|
| 175 |
|
| 176 |
+
field_entry = fields_block.get(field_name)
|
| 177 |
+
if isinstance(field_entry, dict):
|
| 178 |
+
for k, v in field_entry.items():
|
| 179 |
+
if k in ("rules", "extended", "llm"):
|
| 180 |
+
raw[k] = float(v)
|
| 181 |
|
| 182 |
+
if not raw and isinstance(global_block, dict):
|
| 183 |
+
for k, v in global_block.items():
|
| 184 |
+
if k in ("rules", "extended", "llm"):
|
| 185 |
+
raw[k] = float(v)
|
| 186 |
|
| 187 |
+
if not raw:
|
| 188 |
+
raw = _get_base_weights_for_parsers(include_llm)
|
| 189 |
|
| 190 |
+
if not include_llm and "llm" in raw:
|
| 191 |
+
raw.pop("llm", None)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 192 |
|
| 193 |
+
if not raw:
|
| 194 |
+
raw = _get_base_weights_for_parsers(include_llm=False)
|
| 195 |
|
| 196 |
+
return _normalise_scores(raw)
|
|
|
|
|
|
|
| 197 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 198 |
|
| 199 |
# ------------------------------------------------------------
|
| 200 |
+
# Fusion logic
|
| 201 |
# ------------------------------------------------------------
|
| 202 |
|
| 203 |
+
def _clean_pred_value(val: Optional[str]) -> Optional[str]:
|
|
|
|
|
|
|
| 204 |
"""
|
| 205 |
+
Treat None, "", or explicit "Unknown" as missing for fusion.
|
| 206 |
"""
|
| 207 |
+
if val is None:
|
| 208 |
+
return None
|
| 209 |
+
s = str(val).strip()
|
| 210 |
+
if not s:
|
| 211 |
+
return None
|
| 212 |
+
if s.lower() == UNKNOWN.lower():
|
| 213 |
+
return None
|
| 214 |
+
return s
|
| 215 |
|
| 216 |
+
|
| 217 |
+
def parse_text_fused(text: str, use_llm: Optional[bool] = None) -> Dict[str, Any]:
|
| 218 |
"""
|
| 219 |
+
Main tri-fusion entrypoint.
|
| 220 |
+
|
| 221 |
+
Parameters
|
| 222 |
+
----------
|
| 223 |
+
text : str
|
| 224 |
+
use_llm : bool or None
|
| 225 |
+
If True → include LLM.
|
| 226 |
+
If False → skip LLM.
|
| 227 |
+
If None → include if HAS_LLM.
|
| 228 |
+
|
| 229 |
+
Returns:
|
| 230 |
+
full fusion output including votes + per-parser summaries.
|
| 231 |
"""
|
| 232 |
+
original = text or ""
|
| 233 |
+
include_llm = HAS_LLM if use_llm is None else bool(use_llm)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 234 |
|
| 235 |
+
rules_out = parse_text_rules(original) or {}
|
| 236 |
+
ext_out = parse_text_extended(original) or {}
|
|
|
|
| 237 |
|
| 238 |
+
rules_fields = dict(rules_out.get("parsed_fields", {}))
|
| 239 |
+
ext_fields = dict(ext_out.get("parsed_fields", {}))
|
|
|
|
|
|
|
| 240 |
|
| 241 |
+
llm_fields: Dict[str, Any] = {}
|
| 242 |
+
if include_llm and parse_text_llm is not None:
|
| 243 |
+
try:
|
| 244 |
+
llm_out = parse_text_llm(original)
|
| 245 |
+
if isinstance(llm_out, dict):
|
| 246 |
+
if "parsed_fields" in llm_out:
|
| 247 |
+
llm_fields = dict(llm_out.get("parsed_fields", {}))
|
| 248 |
+
else:
|
| 249 |
+
llm_fields = {str(k): v for k, v in llm_out.items()}
|
| 250 |
+
except Exception:
|
| 251 |
+
llm_fields = {}
|
| 252 |
+
else:
|
| 253 |
+
include_llm = False
|
| 254 |
|
| 255 |
+
by_parser: Dict[str, Dict[str, Any]] = {
|
| 256 |
+
"rules": rules_fields,
|
| 257 |
+
"extended": ext_fields,
|
| 258 |
+
"llm": llm_fields if include_llm else {},
|
| 259 |
+
}
|
| 260 |
|
| 261 |
+
candidate_fields = set(rules_fields.keys()) | set(ext_fields.keys()) | set(llm_fields.keys())
|
|
|
|
| 262 |
|
| 263 |
+
fused_fields: Dict[str, Any] = {}
|
| 264 |
+
votes_debug: Dict[str, Any] = {}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 265 |
|
| 266 |
+
for field in sorted(candidate_fields):
|
| 267 |
+
weights = _get_weights_for_field(field, include_llm=include_llm)
|
| 268 |
|
| 269 |
+
parser_preds: Dict[str, Optional[str]] = {
|
| 270 |
+
"rules": _clean_pred_value(rules_fields.get(field)),
|
| 271 |
+
"extended": _clean_pred_value(ext_fields.get(field)),
|
| 272 |
+
"llm": _clean_pred_value(llm_fields.get(field)) if include_llm else None,
|
| 273 |
+
}
|
| 274 |
|
| 275 |
+
per_parser_info: Dict[str, Any] = {}
|
| 276 |
+
value_scores: Dict[str, float] = {}
|
|
|
|
|
|
|
| 277 |
|
| 278 |
+
for parser_name in PARSER_ORDER:
|
| 279 |
+
if parser_name == "llm" and not include_llm:
|
| 280 |
+
continue
|
| 281 |
|
| 282 |
+
pred = parser_preds.get(parser_name)
|
| 283 |
+
w = float(weights.get(parser_name, 0.0))
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 284 |
|
| 285 |
+
per_parser_info[parser_name] = {
|
| 286 |
+
"value": pred if pred is not None else UNKNOWN,
|
| 287 |
+
"weight": w,
|
| 288 |
+
}
|
| 289 |
|
| 290 |
+
if pred is None:
|
| 291 |
+
continue
|
|
|
|
|
|
|
| 292 |
|
| 293 |
+
value_scores[pred] = value_scores.get(pred, 0.0) + w
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 294 |
|
| 295 |
+
if not value_scores:
|
| 296 |
+
fused_value = UNKNOWN
|
| 297 |
+
else:
|
| 298 |
+
max_score = max(value_scores.values())
|
| 299 |
+
best_values = [v for v, s in value_scores.items() if s == max_score]
|
| 300 |
+
|
| 301 |
+
if len(best_values) == 1:
|
| 302 |
+
fused_value = best_values[0]
|
| 303 |
+
else:
|
| 304 |
+
fused_value = best_values[0]
|
| 305 |
+
for parser_name in PARSER_ORDER:
|
| 306 |
+
if parser_name == "llm" and not include_llm:
|
| 307 |
+
continue
|
| 308 |
+
pred = parser_preds.get(parser_name)
|
| 309 |
+
if pred in best_values:
|
| 310 |
+
fused_value = pred
|
| 311 |
+
break
|
| 312 |
+
|
| 313 |
+
fused_fields[field] = fused_value
|
| 314 |
+
|
| 315 |
+
votes_debug[field] = {
|
| 316 |
+
"per_parser": per_parser_info,
|
| 317 |
+
"summed": value_scores,
|
| 318 |
+
"chosen": fused_value,
|
| 319 |
+
}
|
| 320 |
+
|
| 321 |
+
weights_meta = {
|
| 322 |
+
"has_weights_file": HAS_WEIGHTS_FILE,
|
| 323 |
+
"weights_path": FIELD_WEIGHTS_PATH,
|
| 324 |
+
"meta": FIELD_WEIGHTS_RAW.get("meta", {}) if HAS_WEIGHTS_FILE else {},
|
| 325 |
+
}
|
| 326 |
|
| 327 |
return {
|
| 328 |
+
"fused_fields": fused_fields,
|
| 329 |
+
"by_parser": by_parser,
|
| 330 |
+
"votes": votes_debug,
|
| 331 |
+
"weights_meta": weights_meta,
|
| 332 |
+
}
|