Semantic Consistency Auditor

Skill: Semantic Consistency Auditor

ID: 212
Name: semantic-consistency-auditor
Description: Introduces BERTScore and COMET algorithms to evaluate the semantic consistency between AI-generated clinical notes and expert gold standards from the "semantic entailment" level.

Overview

Semantic Consistency Auditor is a medical AI evaluation tool used to assess the semantic consistency between AI-generated clinical notes and expert-written gold standards from a semantic level. This tool is not limited to traditional string matching or bag-of-words models, but uses deep learning models to understand semantic entailment relationships, capable of identifying expressions with different wording but similar meaning.

Algorithms

1. BERTScore

BERTScore uses pre-trained BERT model contextual embeddings to calculate similarity between candidate text and reference text:

• Precision: How much semantics in the candidate text is covered by the reference text
• Recall: How much semantics in the reference text is covered by the candidate text
• F1 Score: Harmonic mean of Precision and Recall

2. COMET (Cross-lingual Optimized Metric for Evaluation of Translation)

COMET is a neural network-based evaluation metric originally used for machine translation evaluation, applicable to semantic entailment tasks:

• Uses XLM-RoBERTa encoder to capture deep semantics
• Outputs semantic consistency scores between 0-1
• Gives high scores to semantically equivalent but differently expressed text

Installation

# Create virtual environment (recommended)
python -m venv venv
source venv/bin/activate  # Linux/Mac
# Or venv\Scripts\activate  # Windows

# Install dependencies
pip install bertscore comet-ml transformers torch

Configuration

Configure in ~/.openclaw/skills/semantic-consistency-auditor/config.yaml:

# BERTScore Configuration
bertscore:
  model: "microsoft/deberta-xlarge-mnli"  # Or "bert-base-chinese" for Chinese
  lang: "zh"  # Language code: zh, en, etc.
  rescale_with_baseline: true
  device: "auto"  # auto, cpu, cuda

# COMET Configuration
comet:
  model: "Unbabel/wmt22-comet-da"  # COMET model
  batch_size: 8
  device: "auto"

# Evaluation Thresholds
thresholds:
  bertscore_f1: 0.85
  comet_score: 0.75
  semantic_consistency: 0.80  # Comprehensive score threshold

Usage

Command Line

# Evaluate single case pair
python scripts/main.py \
  --ai-generated "Patient presented with fever for 3 days, highest temperature 39°C, accompanied by cough." \
  --gold-standard "Patient chief complaint of fever for 3 days, highest temperature 39°C, accompanied by cough symptoms." \
  --output results.json

# Batch evaluation from JSON file
python scripts/main.py \
  --input-file batch_cases.json \
  --output results.json \
  --format detailed

# Use specific model
python scripts/main.py \
  --ai-generated "..." \
  --gold-standard "..." \
  --bert-model "bert-base-chinese" \
  --comet-model "Unbabel/wmt20-comet-da"

Python API

from semantic_consistency_auditor import SemanticConsistencyAuditor

# Initialize evaluator
auditor = SemanticConsistencyAuditor(
    bert_model="microsoft/deberta-xlarge-mnli",
    comet_model="Unbabel/wmt22-comet-da",
    lang="zh"
)

# Evaluate single case
result = auditor.evaluate(
    ai_text="Patient presented with fever for 3 days...",
    gold_text="Patient chief complaint of fever for 3 days..."
)

print(f"BERTScore F1: {result['bertscore']['f1']:.4f}")
print(f"COMET Score: {result['comet']['score']:.4f}")
print(f"Consistency: {result['consistency']:.4f}")
print(f"Passed: {result['passed']}")

# Batch evaluation
results = auditor.evaluate_batch([
    {"ai": "...", "gold": "..."},
    {"ai": "...", "gold": "..."}
])

Input Format

Single Case (Command Line)

Pass text directly through --ai-generated and --gold-standard parameters.

Batch Evaluation File (JSON)

[
  {
    "case_id": "CASE001",
    "ai_generated": "Patient presented with fever for 3 days, highest temperature 39°C, accompanied by cough.",
    "gold_standard": "Patient chief complaint of fever for 3 days, highest temperature 39°C, accompanied by cough symptoms.",
    "metadata": {
      "department": "Respiratory",
      "disease_type": "Upper respiratory infection"
    }
  },
  {
    "case_id": "CASE002",
    "ai_generated": "...",
    "gold_standard": "..."
  }
]

Output Format

Summary Mode

{
  "overall": {
    "total_cases": 100,
    "passed_cases": 85,
    "pass_rate": 0.85,
    "avg_bertscore_f1": 0.8923,
    "avg_comet_score": 0.8234,
    "avg_consistency": 0.8579
  },
  "thresholds": {
    "bertscore_f1": 0.85,
    "comet_score": 0.75,
    "semantic_consistency": 0.80
  }
}

Detailed Mode

{
  "cases": [
    {
      "case_id": "CASE001",
      "ai_generated": "Patient presented with fever for 3 days...",
      "gold_standard": "Patient chief complaint of fever for 3 days...",
      "metrics": {
        "bertscore": {
          "precision": 0.9123,
          "recall": 0.8934,
          "f1": 0.9028
        },
        "comet": {
          "score": 0.8234,
          "system_score": 0.8156
        },
        "semantic_consistency": 0.8631
      },
      "passed": true,
      "details": {
        "semantic_gaps": [],
        "matched_concepts": ["fever for 3 days", "temperature 39°C", "cough"]
      }
    }
  ],
  "summary": { ... }
}

Evaluation Criteria

Semantic Consistency Grading

Pass Criteria

Default pass conditions (configurable):

• BERTScore F1 ≥ 0.85
• COMET Score ≥ 0.75
• Comprehensive semantic consistency ≥ 0.80

Error Handling

Performance Notes

• BERTScore: First run will download model (approximately 400MB-1GB)
• COMET: First run will download model (approximately 500MB-1.5GB)
• GPU Acceleration: Significantly improves evaluation speed in CUDA environment
• Batch Processing: Recommended for batch evaluation to fully utilize GPU parallel capability

References

1. Zhang et al. "BERTScore: Evaluating Text Generation with BERT" ICLR 2020
2. Rei et al. "COMET: A Neural Framework for MT Evaluation" EMNLP 2020
3. Medical Record Standardization Evaluation Guidelines (National Health Commission)

Changelog

• v1.0.0 (2026-02-06): Initial version, supports dual-algorithm evaluation with BERTScore and COMET

Risk Assessment

Security Checklist

• No hardcoded credentials or API keys
• No unauthorized file system access (../)
• Output does not expose sensitive information
• Prompt injection protections in place
• Input file paths validated (no ../ traversal)
• Output directory restricted to workspace
• Script execution in sandboxed environment
• Error messages sanitized (no stack traces exposed)
• Dependencies audited

Prerequisites

# Python dependencies
pip install -r requirements.txt

Evaluation Criteria

Success Metrics

• Successfully executes main functionality
• Output meets quality standards
• Handles edge cases gracefully
• Performance is acceptable

Test Cases

1. Basic Functionality: Standard input → Expected output
2. Edge Case: Invalid input → Graceful error handling
3. Performance: Large dataset → Acceptable processing time

Lifecycle Status

• Current Stage: Draft
• Next Review Date: 2026-03-06
• Known Issues: None
• Planned Improvements:
  • Performance optimization
  • Additional feature support