Spaces:
Paused
Paused
File size: 13,949 Bytes
aceb1b2 | 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 | """
Span-specific IAA metrics.
Span annotations are unusual: annotators can disagree on (a) **where** spans
go (unitization / boundary detection) and (b) **what label** each span carries
(categorization). Token-level kappa and exact-match F1 only capture part of
this picture, which is why dedicated metrics exist:
- **Token-level Cohen / Fleiss kappa** via BIO conversion — simple,
intuitive, but penalizes near-misses harshly and ignores spans of differing
lengths.
- **Span F1 (exact, partial)** — IR-style; classic in NER literature
(MUC, CoNLL, SemEval).
- **Krippendorff's alpha_U (unitizing alpha)** — Krippendorff 2018; treats
each character/token as a unit and accounts for both boundary and
categorical disagreement.
- **Gamma (Mathet et al. 2015)** — state-of-the-art unified measure that
jointly handles unit alignment + categorization via the Hungarian algorithm.
All inputs are ``SpanAnnotation``-like objects with ``start``, ``end``, and
``name`` (label) attributes — or plain dicts/tuples with the same fields.
"""
from __future__ import annotations
from collections import Counter, defaultdict
from typing import Dict, Iterable, List, Optional, Sequence, Tuple
import logging
from potato.server_utils.iaa.nominal import cohen_kappa, fleiss_kappa
logger = logging.getLogger(__name__)
# ---------------------------------------------------------------------------
# Span representation helpers
# ---------------------------------------------------------------------------
def _span_tuple(span) -> Tuple[int, int, str]:
"""Normalise a span object to (start, end, label)."""
if isinstance(span, dict):
return int(span["start"]), int(span["end"]), str(span.get("name") or span.get("label", ""))
if isinstance(span, tuple) and len(span) == 3:
return int(span[0]), int(span[1]), str(span[2])
return int(span.start), int(span.end), str(span.name)
def _normalize(spans: Iterable) -> List[Tuple[int, int, str]]:
return [_span_tuple(s) for s in spans]
# ---------------------------------------------------------------------------
# Token-level kappa via BIO conversion
# ---------------------------------------------------------------------------
def spans_to_bio(spans: Iterable, length: int) -> List[str]:
"""
Convert spans to BIO tags over a unit sequence of length ``length``.
``length`` can be in characters or tokens depending on the unit; the
representation is the same. Overlapping spans are resolved with the rule
"longest span wins" — sufficient for IAA where overlap is rare.
"""
tags = ["O"] * length
span_list = sorted(_normalize(spans), key=lambda s: -(s[1] - s[0]))
for start, end, label in span_list:
start = max(0, start)
end = min(length, end)
if end <= start:
continue
if tags[start] != "O":
continue # respect longest-wins
tags[start] = f"B-{label}"
for i in range(start + 1, end):
if tags[i] != "O":
continue
tags[i] = f"I-{label}"
return tags
def token_level_kappa(
spans_by_user: Dict[str, Iterable],
length: int,
) -> float:
"""
Cohen's / Fleiss' kappa over the BIO tag sequence.
For 2 annotators, returns Cohen's kappa; for >=3, returns Fleiss' kappa.
"""
users = list(spans_by_user)
if len(users) < 2 or length <= 0:
return float("nan")
tag_seqs = {u: spans_to_bio(spans_by_user[u], length) for u in users}
if len(users) == 2:
return cohen_kappa(tag_seqs[users[0]], tag_seqs[users[1]])
counts_per_position = []
for i in range(length):
c: Counter = Counter()
for u in users:
c[tag_seqs[u][i]] += 1
counts_per_position.append(dict(c))
return fleiss_kappa(counts_per_position)
# ---------------------------------------------------------------------------
# Span F1 (exact and partial match)
# ---------------------------------------------------------------------------
def _overlap_len(a: Tuple[int, int, str], b: Tuple[int, int, str]) -> int:
return max(0, min(a[1], b[1]) - max(a[0], b[0]))
def span_f1_exact(spans_a: Iterable, spans_b: Iterable) -> Tuple[float, float, float]:
"""
Strict exact-match F1: (start, end, label) must match exactly.
Returns (precision, recall, F1) treating spans_b as gold.
"""
a = set(_normalize(spans_a))
b = set(_normalize(spans_b))
if not a and not b:
return 1.0, 1.0, 1.0
tp = len(a & b)
p = tp / len(a) if a else 0.0
r = tp / len(b) if b else 0.0
f1 = 2 * p * r / (p + r) if (p + r) else 0.0
return p, r, f1
def span_f1_partial(
spans_a: Iterable,
spans_b: Iterable,
label_must_match: bool = True,
threshold: float = 0.5,
) -> Tuple[float, float, float]:
"""
Partial-match F1: a span counts as TP if it overlaps a gold span by at
least ``threshold`` of either span's length (Dice-overlap convention).
label_must_match: when True (default) overlapping spans must share the
same label to count; False allows boundary-only agreement.
"""
a = _normalize(spans_a)
b = _normalize(spans_b)
if not a and not b:
return 1.0, 1.0, 1.0
matched_b = set()
tp = 0
for sa in a:
for idx, sb in enumerate(b):
if idx in matched_b:
continue
if label_must_match and sa[2] != sb[2]:
continue
ov = _overlap_len(sa, sb)
if ov <= 0:
continue
la, lb = sa[1] - sa[0], sb[1] - sb[0]
if la <= 0 or lb <= 0:
continue
if (ov / la) >= threshold or (ov / lb) >= threshold:
tp += 1
matched_b.add(idx)
break
p = tp / len(a) if a else 0.0
r = tp / len(b) if b else 0.0
f1 = 2 * p * r / (p + r) if (p + r) else 0.0
return p, r, f1
def pairwise_span_f1(
spans_by_user: Dict[str, Iterable],
partial: bool = False,
threshold: float = 0.5,
) -> float:
"""Mean pairwise span-F1 across users."""
users = list(spans_by_user)
if len(users) < 2:
return float("nan")
scores = []
for i in range(len(users)):
for j in range(i + 1, len(users)):
if partial:
_, _, f1 = span_f1_partial(
spans_by_user[users[i]], spans_by_user[users[j]], threshold=threshold,
)
else:
_, _, f1 = span_f1_exact(spans_by_user[users[i]], spans_by_user[users[j]])
scores.append(f1)
if not scores:
return float("nan")
return sum(scores) / len(scores)
# ---------------------------------------------------------------------------
# Krippendorff's alpha_U (unitizing alpha)
# ---------------------------------------------------------------------------
def krippendorff_alpha_u(
spans_by_user: Dict[str, Iterable],
length: int,
) -> float:
"""
Krippendorff's unitizing alpha for span annotation.
Implementation: assign each character/token position a categorical label
(one of the span labels or "O") per annotator, then compute Krippendorff's
alpha (nominal) over the (annotator, position) pairs. This is the
operational form recommended in Krippendorff (2018) when the unit is
fixed (per-character) rather than continuous.
For continuous-domain alpha_U (where annotators may disagree on the unit
boundary in a fundamentally continuous space such as audio), prefer gamma.
"""
users = list(spans_by_user)
if len(users) < 2 or length <= 0:
return float("nan")
rows = []
for u in users:
tags = spans_to_bio(spans_by_user[u], length)
# Map BIO -> base label (strip B-/I- prefix) so boundary placement
# within a contiguous span doesn't count as disagreement.
for pos, tag in enumerate(tags):
label = "O" if tag == "O" else tag.split("-", 1)[1]
rows.append((u, pos, label))
from potato.server_utils.iaa.alpha import krippendorff_alpha
return krippendorff_alpha(rows, level="nominal")
# ---------------------------------------------------------------------------
# Gamma (Mathet et al. 2015)
# ---------------------------------------------------------------------------
def _positional_dissimilarity(
a: Tuple[int, int, str],
b: Tuple[int, int, str],
delta_empty: float,
) -> float:
"""Positional component of the Mathet dissimilarity (normalized)."""
if a is None or b is None:
return delta_empty
# Sum of |starts diff| + |ends diff|, normalized by total span lengths.
diff = abs(a[0] - b[0]) + abs(a[1] - b[1])
total = (a[1] - a[0]) + (b[1] - b[0])
if total <= 0:
return delta_empty
return diff / total
def _categorical_dissimilarity(
a: Tuple[int, int, str],
b: Tuple[int, int, str],
delta_empty: float,
) -> float:
if a is None or b is None:
return delta_empty
return 0.0 if a[2] == b[2] else 1.0
def _pairwise_disorder(
spans_a: List[Tuple[int, int, str]],
spans_b: List[Tuple[int, int, str]],
alpha: float,
beta: float,
delta_empty: float,
) -> float:
"""
Optimal-alignment disorder between two annotators' span sets.
Uses the Hungarian algorithm (``scipy.optimize.linear_sum_assignment``)
with padded empty units so that |spans_a| != |spans_b| is handled.
"""
try:
import numpy as np
from scipy.optimize import linear_sum_assignment
except ImportError: # pragma: no cover
logger.warning("scipy unavailable; gamma falling back to NaN")
return float("nan")
n = max(len(spans_a), len(spans_b))
if n == 0:
return 0.0
# Pad shorter side with None (= empty unit)
a_padded: List[Optional[Tuple[int, int, str]]] = list(spans_a) + [None] * (n - len(spans_a))
b_padded: List[Optional[Tuple[int, int, str]]] = list(spans_b) + [None] * (n - len(spans_b))
cost = np.zeros((n, n), dtype=float)
for i in range(n):
for j in range(n):
pos = _positional_dissimilarity(a_padded[i], b_padded[j], delta_empty)
cat = _categorical_dissimilarity(a_padded[i], b_padded[j], delta_empty)
cost[i, j] = alpha * pos + beta * cat
row_ind, col_ind = linear_sum_assignment(cost)
total = float(cost[row_ind, col_ind].sum())
return total / n
def gamma(
spans_by_user: Dict[str, Iterable],
length: Optional[int] = None,
alpha: float = 1.0,
beta: float = 1.0,
n_samples: int = 30,
seed: int = 1234,
) -> float:
"""
Mathet et al. (2015) gamma agreement.
Args:
spans_by_user: annotator_id -> iterable of spans
length: total length of the unit space (characters or tokens). If
omitted, inferred from the maximum span end across annotators.
alpha: weight on positional dissimilarity.
beta: weight on categorical dissimilarity.
n_samples: number of random pairings used to estimate the
expected-by-chance disorder.
seed: RNG seed for reproducibility.
Returns:
gamma in [-1, 1] approximately, where 1 = perfect agreement, 0 =
chance-level. NaN if scipy is unavailable.
Notes:
This implementation is a faithful but simplified rendition: positional
dissimilarity is normalized by combined span length, and the chance
baseline is estimated by re-pairing spans across all annotators
``n_samples`` times. Full pygamma-agreement uses a more sophisticated
baseline (continuum-of-shuffles); the simplification is sufficient for
the relative IAA comparisons that drive routing decisions.
"""
import random as _random
users = list(spans_by_user)
if len(users) < 2:
return float("nan")
normed = {u: _normalize(spans_by_user[u]) for u in users}
# Empty-unit dissimilarity follows Mathet: a moderate constant ~ 1
delta_empty = 1.0
# Observed disorder: mean pairwise disorder across all annotator pairs
pair_disorders = []
for i in range(len(users)):
for j in range(i + 1, len(users)):
pair_disorders.append(
_pairwise_disorder(normed[users[i]], normed[users[j]], alpha, beta, delta_empty)
)
if not pair_disorders:
return float("nan")
if any(d != d for d in pair_disorders): # NaN -> bail
return float("nan")
observed = sum(pair_disorders) / len(pair_disorders)
# Expected-by-chance disorder via shuffled pairings
all_spans = [s for u in users for s in normed[u]]
if len(all_spans) < 2:
return 1.0 if observed == 0 else float("nan")
rng = _random.Random(seed)
chance_disorders = []
sizes = [len(normed[u]) for u in users]
for _ in range(n_samples):
shuffled = list(all_spans)
rng.shuffle(shuffled)
# Re-distribute back to annotators preserving original counts
idx = 0
shuffled_per_user = []
for sz in sizes:
shuffled_per_user.append(shuffled[idx:idx + sz])
idx += sz
sample_pair_disorders = []
for i in range(len(users)):
for j in range(i + 1, len(users)):
sample_pair_disorders.append(
_pairwise_disorder(
shuffled_per_user[i],
shuffled_per_user[j],
alpha, beta, delta_empty,
)
)
if sample_pair_disorders:
chance_disorders.append(sum(sample_pair_disorders) / len(sample_pair_disorders))
if not chance_disorders:
return float("nan")
expected = sum(chance_disorders) / len(chance_disorders)
if expected <= 0:
return 1.0 if observed == 0 else float("nan")
return 1.0 - (observed / expected)
|