MultiModal-Coherence-AI / src /evaluation /human_eval_analyzer.py
pratik-250620's picture
Upload folder using huggingface_hub
6835659 verified
"""
Human Evaluation Analysis Module
Analyzes human evaluation data to compute:
- Intra-rater reliability (Cohen's kappa for self-agreement)
- Inter-rater reliability (Krippendorff's alpha for multi-rater agreement)
- Descriptive statistics
- Correlation with MSCI scores (aggregated across raters)
"""
from __future__ import annotations
from collections import defaultdict
from dataclasses import dataclass
from pathlib import Path
from typing import Dict, List, Optional, Tuple, Any
import json
import numpy as np
from scipy import stats
from src.evaluation.human_eval_schema import (
EvaluationSession,
HumanEvaluation,
ReliabilityMetrics,
)
def compute_cohens_kappa(ratings1: List[int], ratings2: List[int]) -> float:
"""
Compute Cohen's kappa for two sets of ratings.
Args:
ratings1: First set of ratings
ratings2: Second set of ratings (same samples, different time)
Returns:
Cohen's kappa coefficient
"""
if len(ratings1) != len(ratings2):
raise ValueError("Rating lists must have the same length")
n = len(ratings1)
if n == 0:
return 0.0
# Create confusion matrix
categories = sorted(set(ratings1) | set(ratings2))
k = len(categories)
cat_to_idx = {cat: i for i, cat in enumerate(categories)}
confusion = np.zeros((k, k))
for r1, r2 in zip(ratings1, ratings2):
confusion[cat_to_idx[r1], cat_to_idx[r2]] += 1
# Compute observed agreement
p_o = np.trace(confusion) / n
# Compute expected agreement by chance
row_sums = confusion.sum(axis=1)
col_sums = confusion.sum(axis=0)
p_e = np.sum(row_sums * col_sums) / (n * n)
# Cohen's kappa
if p_e == 1.0:
return 1.0
return (p_o - p_e) / (1 - p_e)
def compute_weighted_kappa(
ratings1: List[int], ratings2: List[int], weights: str = "quadratic"
) -> float:
"""
Compute weighted Cohen's kappa for ordinal data.
Args:
ratings1: First set of ratings
ratings2: Second set of ratings
weights: "linear" or "quadratic" weighting scheme
Returns:
Weighted kappa coefficient
"""
if len(ratings1) != len(ratings2):
raise ValueError("Rating lists must have the same length")
n = len(ratings1)
if n == 0:
return 0.0
# Get all categories
all_ratings = set(ratings1) | set(ratings2)
min_cat, max_cat = min(all_ratings), max(all_ratings)
categories = list(range(min_cat, max_cat + 1))
k = len(categories)
cat_to_idx = {cat: i for i, cat in enumerate(categories)}
# Create confusion matrix
confusion = np.zeros((k, k))
for r1, r2 in zip(ratings1, ratings2):
confusion[cat_to_idx[r1], cat_to_idx[r2]] += 1
# Create weight matrix
weight_matrix = np.zeros((k, k))
for i in range(k):
for j in range(k):
if weights == "linear":
weight_matrix[i, j] = abs(i - j) / (k - 1)
else: # quadratic
weight_matrix[i, j] = ((i - j) ** 2) / ((k - 1) ** 2)
# Normalize confusion matrix
confusion = confusion / n
# Marginal distributions
row_sums = confusion.sum(axis=1)
col_sums = confusion.sum(axis=0)
# Expected matrix
expected = np.outer(row_sums, col_sums)
# Weighted observed and expected
w_observed = np.sum(weight_matrix * confusion)
w_expected = np.sum(weight_matrix * expected)
if w_expected == 0:
return 1.0
return 1 - (w_observed / w_expected)
def compute_intra_rater_reliability(
session: EvaluationSession,
) -> Optional[ReliabilityMetrics]:
"""
Compute intra-rater reliability from re-rated samples.
Args:
session: Evaluation session containing evaluations
Returns:
ReliabilityMetrics or None if no re-ratings available
"""
# Find paired evaluations (first rating vs re-rating)
first_ratings: Dict[str, HumanEvaluation] = {}
reratings: Dict[str, HumanEvaluation] = {}
for eval in session.evaluations:
if eval.sample_id in session.rerating_sample_ids:
if eval.is_rerating:
reratings[eval.sample_id] = eval
else:
first_ratings[eval.sample_id] = eval
# Get paired samples
paired_ids = set(first_ratings.keys()) & set(reratings.keys())
if not paired_ids:
return None
# Extract rating pairs for each dimension
dimensions = [
"text_image_coherence",
"text_audio_coherence",
"image_audio_coherence",
"overall_coherence",
]
all_first = []
all_second = []
for sample_id in paired_ids:
first = first_ratings[sample_id]
second = reratings[sample_id]
for dim in dimensions:
all_first.append(getattr(first, dim))
all_second.append(getattr(second, dim))
# Compute metrics
kappa = compute_cohens_kappa(all_first, all_second)
weighted_kappa = compute_weighted_kappa(all_first, all_second, weights="quadratic")
# Simple agreement
agreements = sum(1 for f, s in zip(all_first, all_second) if f == s)
percent_agreement = agreements / len(all_first) * 100
# Mean absolute difference
mad = np.mean([abs(f - s) for f, s in zip(all_first, all_second)])
return ReliabilityMetrics(
kappa=kappa,
percent_agreement=percent_agreement,
weighted_kappa=weighted_kappa,
mean_absolute_difference=mad,
n_reratings=len(paired_ids),
)
@dataclass
class HumanEvalSummary:
"""Summary statistics for human evaluations."""
n_samples: int
n_evaluations: int
# Per-dimension statistics
text_image_mean: float
text_image_std: float
text_audio_mean: float
text_audio_std: float
image_audio_mean: float
image_audio_std: float
overall_mean: float
overall_std: float
# Aggregated scores
mean_weighted_score: float
std_weighted_score: float
# Reliability
reliability: Optional[ReliabilityMetrics]
def to_dict(self) -> Dict[str, Any]:
"""Convert to dictionary."""
return {
"n_samples": self.n_samples,
"n_evaluations": self.n_evaluations,
"text_image": {"mean": self.text_image_mean, "std": self.text_image_std},
"text_audio": {"mean": self.text_audio_mean, "std": self.text_audio_std},
"image_audio": {"mean": self.image_audio_mean, "std": self.image_audio_std},
"overall": {"mean": self.overall_mean, "std": self.overall_std},
"weighted_score": {"mean": self.mean_weighted_score, "std": self.std_weighted_score},
"reliability": self.reliability.to_dict() if self.reliability else None,
}
def compute_human_eval_summary(session: EvaluationSession) -> HumanEvalSummary:
"""
Compute summary statistics for human evaluations.
Args:
session: Evaluation session
Returns:
HumanEvalSummary with descriptive statistics
"""
# Filter out re-ratings for summary stats
evals = [e for e in session.evaluations if not e.is_rerating]
if not evals:
raise ValueError("No evaluations found in session")
# Extract ratings
ti = [e.text_image_coherence for e in evals]
ta = [e.text_audio_coherence for e in evals]
ia = [e.image_audio_coherence for e in evals]
overall = [e.overall_coherence for e in evals]
weighted = [e.weighted_score() for e in evals]
# Compute reliability
reliability = compute_intra_rater_reliability(session)
return HumanEvalSummary(
n_samples=len(set(e.sample_id for e in evals)),
n_evaluations=len(evals),
text_image_mean=np.mean(ti),
text_image_std=np.std(ti),
text_audio_mean=np.mean(ta),
text_audio_std=np.std(ta),
image_audio_mean=np.mean(ia),
image_audio_std=np.std(ia),
overall_mean=np.mean(overall),
overall_std=np.std(overall),
mean_weighted_score=np.mean(weighted),
std_weighted_score=np.std(weighted),
reliability=reliability,
)
def compute_human_msci_correlation(
session: EvaluationSession,
msci_scores: Optional[Dict[str, float]] = None,
) -> Dict[str, Any]:
"""
Compute correlation between human ratings and MSCI scores.
Args:
session: Evaluation session with human ratings
msci_scores: Optional dict mapping sample_id to MSCI score.
If None, uses msci_score from sample metadata.
Returns:
Dictionary with correlation statistics
"""
# Get paired human scores and MSCI scores
human_weighted = []
human_overall = []
msci_values = []
sample_msci = {}
if msci_scores:
sample_msci = msci_scores
else:
# Try to extract from session samples
for sample in session.samples:
if sample.msci_score is not None:
sample_msci[sample.sample_id] = sample.msci_score
for eval in session.evaluations:
if eval.is_rerating:
continue
if eval.sample_id in sample_msci:
human_weighted.append(eval.weighted_score())
human_overall.append(eval.overall_coherence / 5.0) # Normalize to 0-1
msci_values.append(sample_msci[eval.sample_id])
if len(msci_values) < 3:
return {
"error": "Insufficient paired data for correlation",
"n_paired": len(msci_values),
}
# Spearman correlation (for ordinal human ratings)
spearman_weighted = stats.spearmanr(msci_values, human_weighted)
spearman_overall = stats.spearmanr(msci_values, human_overall)
# Pearson correlation
pearson_weighted = stats.pearsonr(msci_values, human_weighted)
pearson_overall = stats.pearsonr(msci_values, human_overall)
return {
"n_paired": len(msci_values),
"msci_vs_weighted_human": {
"spearman_rho": spearman_weighted.correlation,
"spearman_p": spearman_weighted.pvalue,
"pearson_r": pearson_weighted.statistic,
"pearson_p": pearson_weighted.pvalue,
},
"msci_vs_overall_human": {
"spearman_rho": spearman_overall.correlation,
"spearman_p": spearman_overall.pvalue,
"pearson_r": pearson_overall.statistic,
"pearson_p": pearson_overall.pvalue,
},
"interpretation": _interpret_correlation(spearman_weighted.correlation, spearman_weighted.pvalue),
}
def _interpret_correlation(rho: float, p: float, alpha: float = 0.05) -> str:
"""Generate human-readable interpretation of correlation."""
if p >= alpha:
return f"No significant correlation (ρ={rho:.3f}, p={p:.4f}{alpha})"
strength = "weak" if abs(rho) < 0.3 else "moderate" if abs(rho) < 0.6 else "strong"
direction = "positive" if rho > 0 else "negative"
return f"Significant {strength} {direction} correlation (ρ={rho:.3f}, p={p:.4f})"
def analyze_by_condition(session: EvaluationSession) -> Dict[str, Dict[str, Any]]:
"""
Analyze human ratings grouped by experimental condition.
Args:
session: Evaluation session
Returns:
Dictionary with statistics per condition
"""
# Group evaluations by condition
by_condition: Dict[str, List[HumanEvaluation]] = defaultdict(list)
# Create sample_id to condition mapping
sample_to_condition = {s.sample_id: s.condition for s in session.samples}
for eval in session.evaluations:
if eval.is_rerating:
continue
condition = sample_to_condition.get(eval.sample_id, "unknown")
by_condition[condition].append(eval)
results = {}
for condition, evals in by_condition.items():
if not evals:
continue
weighted = [e.weighted_score() for e in evals]
overall = [e.overall_coherence for e in evals]
results[condition] = {
"n": len(evals),
"weighted_score": {
"mean": np.mean(weighted),
"std": np.std(weighted),
"median": np.median(weighted),
},
"overall_coherence": {
"mean": np.mean(overall),
"std": np.std(overall),
"median": np.median(overall),
},
}
return results
def generate_analysis_report(
session: EvaluationSession,
output_path: Optional[Path] = None,
) -> Dict[str, Any]:
"""
Generate a comprehensive analysis report.
Args:
session: Evaluation session
output_path: Optional path to save JSON report
Returns:
Dictionary with complete analysis
"""
report = {
"session_id": session.session_id,
"evaluator_id": session.evaluator_id,
"started_at": session.started_at,
"completed_at": session.completed_at,
"summary": compute_human_eval_summary(session).to_dict(),
"by_condition": analyze_by_condition(session),
"msci_correlation": compute_human_msci_correlation(session),
}
if output_path:
output_path = Path(output_path)
output_path.parent.mkdir(parents=True, exist_ok=True)
with output_path.open("w", encoding="utf-8") as f:
json.dump(report, f, indent=2, ensure_ascii=False)
return report
# =========================================================================
# Multi-rater analysis (RQ3)
# =========================================================================
def compute_krippendorff_alpha(
data_matrix: np.ndarray,
level: str = "ordinal",
) -> float:
"""
Compute Krippendorff's alpha for inter-rater reliability.
Args:
data_matrix: Shape (n_raters, n_items). Use np.nan for missing values.
level: "nominal", "ordinal", or "interval" measurement level.
Returns:
Krippendorff's alpha coefficient (-1 to 1, >0.667 acceptable).
"""
n_raters, n_items = data_matrix.shape
# Collect all valid value pairs within each item
# D_o = observed disagreement, D_e = expected disagreement
all_values = []
pairs_observed = []
for item in range(n_items):
values = data_matrix[:, item]
valid = values[~np.isnan(values)]
if len(valid) < 2:
continue
all_values.extend(valid)
# All pairs within this item
for i in range(len(valid)):
for j in range(i + 1, len(valid)):
pairs_observed.append((valid[i], valid[j]))
if not pairs_observed:
return 0.0
all_values = np.array(all_values)
# Distance function
if level == "nominal":
def dist(a, b):
return 0.0 if a == b else 1.0
elif level == "ordinal":
# For ordinal: use squared rank difference
unique_vals = np.sort(np.unique(all_values))
val_to_rank = {v: i for i, v in enumerate(unique_vals)}
def dist(a, b):
return (val_to_rank[a] - val_to_rank[b]) ** 2
else: # interval
def dist(a, b):
return (a - b) ** 2
# Observed disagreement
D_o = np.mean([dist(a, b) for a, b in pairs_observed])
# Expected disagreement (all possible pairs from marginal distribution)
n_total = len(all_values)
D_e_sum = 0.0
count = 0
for i in range(n_total):
for j in range(i + 1, n_total):
D_e_sum += dist(all_values[i], all_values[j])
count += 1
D_e = D_e_sum / count if count > 0 else 0.0
if D_e == 0:
return 1.0 # Perfect agreement if no variance
alpha = 1.0 - D_o / D_e
return alpha
def aggregate_multi_rater_sessions(
sessions: List[EvaluationSession],
) -> Dict[str, Dict[str, Any]]:
"""
Aggregate evaluations across multiple raters for the same samples.
Args:
sessions: List of completed evaluation sessions (same sample set).
Returns:
Dictionary mapping sample_id to aggregated scores.
"""
# Collect all evaluations per sample (excluding re-ratings)
by_sample: Dict[str, List[HumanEvaluation]] = defaultdict(list)
for session in sessions:
for ev in session.evaluations:
if ev.is_rerating:
continue
by_sample[ev.sample_id].append(ev)
# Compute mean scores per sample
aggregated = {}
for sample_id, evals in by_sample.items():
ti = [e.text_image_coherence for e in evals]
ta = [e.text_audio_coherence for e in evals]
ia = [e.image_audio_coherence for e in evals]
overall = [e.overall_coherence for e in evals]
weighted = [e.weighted_score() for e in evals]
aggregated[sample_id] = {
"n_raters": len(evals),
"text_image": {"mean": float(np.mean(ti)), "std": float(np.std(ti))},
"text_audio": {"mean": float(np.mean(ta)), "std": float(np.std(ta))},
"image_audio": {"mean": float(np.mean(ia)), "std": float(np.std(ia))},
"overall": {"mean": float(np.mean(overall)), "std": float(np.std(overall))},
"weighted_score": {"mean": float(np.mean(weighted)), "std": float(np.std(weighted))},
"evaluator_ids": [e.evaluator_id for e in evals],
}
return aggregated
def compute_inter_rater_reliability(
sessions: List[EvaluationSession],
) -> Dict[str, Any]:
"""
Compute inter-rater reliability across multiple evaluators.
Args:
sessions: List of evaluation sessions (same sample set).
Returns:
Dictionary with Krippendorff's alpha per dimension and overall.
"""
# Get common sample IDs (rated by all raters)
sample_sets = []
for session in sessions:
ids = {e.sample_id for e in session.evaluations if not e.is_rerating}
sample_sets.append(ids)
common_ids = sorted(set.intersection(*sample_sets)) if sample_sets else []
if len(common_ids) < 3:
return {"error": "Too few common samples for reliability analysis",
"n_common": len(common_ids)}
n_raters = len(sessions)
n_items = len(common_ids)
id_to_idx = {sid: i for i, sid in enumerate(common_ids)}
dimensions = {
"text_image": "text_image_coherence",
"text_audio": "text_audio_coherence",
"image_audio": "image_audio_coherence",
"overall": "overall_coherence",
}
results = {"n_raters": n_raters, "n_common_samples": n_items}
for dim_name, attr_name in dimensions.items():
matrix = np.full((n_raters, n_items), np.nan)
for rater_idx, session in enumerate(sessions):
for ev in session.evaluations:
if ev.is_rerating:
continue
if ev.sample_id in id_to_idx:
matrix[rater_idx, id_to_idx[ev.sample_id]] = getattr(ev, attr_name)
alpha = compute_krippendorff_alpha(matrix, level="ordinal")
results[dim_name] = {
"krippendorff_alpha": round(alpha, 4),
"interpretation": _interpret_alpha(alpha),
}
# Weighted score dimension
w_matrix = np.full((n_raters, n_items), np.nan)
for rater_idx, session in enumerate(sessions):
for ev in session.evaluations:
if ev.is_rerating:
continue
if ev.sample_id in id_to_idx:
w_matrix[rater_idx, id_to_idx[ev.sample_id]] = ev.weighted_score()
alpha_w = compute_krippendorff_alpha(w_matrix, level="interval")
results["weighted_score"] = {
"krippendorff_alpha": round(alpha_w, 4),
"interpretation": _interpret_alpha(alpha_w),
}
return results
def _interpret_alpha(alpha: float) -> str:
"""Interpret Krippendorff's alpha value."""
if alpha >= 0.80:
return "good agreement"
elif alpha >= 0.667:
return "acceptable agreement"
elif alpha >= 0.40:
return "moderate agreement"
else:
return "poor agreement"
def compute_multi_rater_msci_correlation(
sessions: List[EvaluationSession],
sample_msci: Dict[str, float],
) -> Dict[str, Any]:
"""
Compute Spearman correlation between average human scores and MSCI.
Args:
sessions: List of evaluation sessions.
sample_msci: Mapping sample_id -> MSCI score.
Returns:
Correlation statistics with bootstrap 95% CI.
"""
aggregated = aggregate_multi_rater_sessions(sessions)
human_scores = []
msci_scores = []
for sample_id, agg in aggregated.items():
if sample_id in sample_msci:
human_scores.append(agg["weighted_score"]["mean"])
msci_scores.append(sample_msci[sample_id])
if len(human_scores) < 5:
return {"error": "Too few paired samples", "n_paired": len(human_scores)}
human_arr = np.array(human_scores)
msci_arr = np.array(msci_scores)
# Spearman rho
spearman = stats.spearmanr(msci_arr, human_arr)
# Pearson r
pearson = stats.pearsonr(msci_arr, human_arr)
# Bootstrap 95% CI for Spearman rho
n_boot = 10000
rng = np.random.default_rng(42)
boot_rhos = []
for _ in range(n_boot):
idx = rng.choice(len(human_arr), size=len(human_arr), replace=True)
r, _ = stats.spearmanr(msci_arr[idx], human_arr[idx])
boot_rhos.append(r)
ci_lower = float(np.percentile(boot_rhos, 2.5))
ci_upper = float(np.percentile(boot_rhos, 97.5))
return {
"n_paired": len(human_scores),
"spearman_rho": round(float(spearman.correlation), 4),
"spearman_p": float(spearman.pvalue),
"spearman_95ci": [round(ci_lower, 4), round(ci_upper, 4)],
"pearson_r": round(float(pearson.statistic), 4),
"pearson_p": float(pearson.pvalue),
"interpretation": _interpret_correlation(
float(spearman.correlation), float(spearman.pvalue)
),
}