ACOT-ADR - Adaptive Chain-of-Thought ADR Evaluation Dataset

ACOT-ADR is a comprehensive evaluation dataset designed to assess Adaptive Chain-of-Thought reasoning approaches in Chinese adverse drug reaction (ADR) extraction and standardization. The dataset contains real-world patient narratives with comprehensive annotations for drug entities, symptom entities, and causal relationships, featuring structured evaluations across multiple large language models (DeepSeek-V3, DeepSeek-R1, and DeepSeek-Distill-7B) using various reasoning strategies including baseline methods, rule-based pipelines, LLM-based pipelines, and integrated prompts. Each sample includes a six-dimensional feature vector capturing semantic characteristics such as text type, drug mention explicitness, symptom clarity, causal indicators, subjective expressions, and temporal cues. The evaluation framework encompasses drug entity recognition and standardization, symptom entity recognition and standardization, and drug-symptom causal relationship mining with probability assessment, providing detailed performance metrics including precision, recall, F1-scores, and execution time analysis. ACOT-ADR serves as a valuable benchmark for developing adaptive reasoning approaches in medical information extraction, facilitating reproducible research in pharmacovigilance applications and supporting the advancement of sophisticated reasoning frameworks for clinical text processing.