Evaluation References

The Hidden-RAD2 evaluation framework draws on methods from structured prediction, named entity recognition, text correction, semantic similarity, and radiology-report evaluation.

NTCIR-18 Hidden-RAD Evaluation

The NTCIR-18 framework evaluated free-text causal explanations using semantic similarity, diagnostic consistency, causal appropriateness, and expert review.

NTCIR-18 Hidden-RAD Evaluation Scheme

This framework provides background for evaluating clinical reasoning but is not directly reused for the structured A1–A5 outputs of Hidden-RAD2 Task 1.

Span and Type Evaluation

Exact span and typed-span precision, recall, and F1 are standard methods in named entity recognition.

Nested Named Entity Recognition: A Survey

Edit-Based Correction Evaluation

ERRANT extracts and classifies edits between original and corrected text and supports error-detection and correction evaluation.

Automatic Annotation and Evaluation of Error Types for Grammatical Error Correction

PT-M2 applies pretrained semantic metrics to edited portions rather than allowing unchanged text to dominate the score.

Revisiting Grammatical Error Correction Evaluation and Beyond

Semantic Similarity

BERTScore compares candidate and reference texts using contextual token embeddings.

BERTScore: Evaluating Text Generation with BERT

Semantic similarity is used as a supporting measure and does not by itself establish clinical factuality or fluency.

Radiology-Specific Evaluation

RadGraph F1 compares clinical entities and relations, while RadCliQ combines automatic metrics to better approximate radiologist-assessed error severity.

Evaluating Progress in Automatic Chest X-ray Radiology Report Generation

GREEN uses a radiology-specific language model to identify and explain clinically significant errors in generated reports.

GREEN: Generative Radiology Report Evaluation and Error Notation