Background Traditional medicine (TM) employs medicinal herbs (MHs) to manage diverse health conditions, yet much of this evidence remains embedded in narrative biomedical text, limiting its use in computational analysis and decision making. As integrative medicine increasingly demands reliable, machine-interpretable knowledge, there is a need for structured representations of MH–disease relationships that prioritize evidential clarity and semantic consistency. This study aims to develop an expert-annotated corpus designed to support decision-relevant knowledge extraction for integrative biomedical informatics. Methods We developed the Medicinal Herb–Disease Relationships (MHDR) corpus by systematically collecting 800 PubMed abstracts using standardized pharmacognostic names. To ensure reliability for decision-support applications, relationships were extracted exclusively from key sentences—sentences that explicitly and unambiguously state MH–disease associations and provide direct evidence for the underlying claims. Relationships expressed in non-key sentences were intentionally excluded, as they often contain implicit, speculative, or context-dependent statements that may reduce decision reliability. Three TM experts manually annotated MH entities, disease entities, and explicit relationships through a four-phase consensus-driven protocol. Baseline Transformer-based models were evaluated for entity recognition, key-sentence identification, and relation extraction to assess the corpus’s computational usability. Results The MHDR corpus contains 5,119 medicinal herb mentions, 6,621 disease mentions, and 1,314 high-confidence MH–disease relationships derived from 832 decision-relevant key sentences. Baseline Transformer models demonstrated robust performance in recognizing entities and extracting relationships, confirming that the corpus supports stable and interpretable knowledge extraction suitable for downstream decision-support and evidence-integration tasks. Conclusion The MHDR corpus represents a decision-oriented informatics resource for modeling medicinal herb–disease knowledge in TM. By restricting annotations to explicit evidence-bearing sentences, the corpus enhances semantic reliability and supports computable reasoning, enabling its use in clinical research analytics, ontology alignment, and integrative decision-making systems bridging traditional and modern medicine. The MHDR corpus is publicly available through GitHub (<a href="https://github.com/KIOM-AIDoc/MHDR">https://github.com/KIOM-AIDoc/MHDR</a>) and Figshare (<a href="https://doi.org/10.6084/m9.figshare.29555549">https://doi.org/10.6084/m9.figshare.29555549</a>).

From narrative evidence to computable knowledge: a decision-relevant corpus for medicinal herb–disease relationships

期刊论文
出版模式：

开放获取
发表日期：

2026年5月13日

查看全文

Sangjun Yea,
Ho Jang,
Jaeuk U. Kim

摘要

Background

Traditional medicine (TM) employs medicinal herbs (MHs) to manage diverse health conditions, yet much of this evidence remains embedded in narrative biomedical text, limiting its use in computational analysis and decision making. As integrative medicine increasingly demands reliable, machine-interpretable knowledge, there is a need for structured representations of MH–disease relationships that prioritize evidential clarity and semantic consistency. This study aims to develop an expert-annotated corpus designed to support decision-relevant knowledge extraction for integrative biomedical informatics.

Methods

We developed the Medicinal Herb–Disease Relationships (MHDR) corpus by systematically collecting 800 PubMed abstracts using standardized pharmacognostic names. To ensure reliability for decision-support applications, relationships were extracted exclusively from key sentences—sentences that explicitly and unambiguously state MH–disease associations and provide direct evidence for the underlying claims. Relationships expressed in non-key sentences were intentionally excluded, as they often contain implicit, speculative, or context-dependent statements that may reduce decision reliability. Three TM experts manually annotated MH entities, disease entities, and explicit relationships through a four-phase consensus-driven protocol. Baseline Transformer-based models were evaluated for entity recognition, key-sentence identification, and relation extraction to assess the corpus’s computational usability.

Results

The MHDR corpus contains 5,119 medicinal herb mentions, 6,621 disease mentions, and 1,314 high-confidence MH–disease relationships derived from 832 decision-relevant key sentences. Baseline Transformer models demonstrated robust performance in recognizing entities and extracting relationships, confirming that the corpus supports stable and interpretable knowledge extraction suitable for downstream decision-support and evidence-integration tasks.

Conclusion

The MHDR corpus represents a decision-oriented informatics resource for modeling medicinal herb–disease knowledge in TM. By restricting annotations to explicit evidence-bearing sentences, the corpus enhances semantic reliability and supports computable reasoning, enabling its use in clinical research analytics, ontology alignment, and integrative decision-making systems bridging traditional and modern medicine. The MHDR corpus is publicly available through GitHub (https://github.com/KIOM-AIDoc/MHDR) and Figshare (https://doi.org/10.6084/m9.figshare.29555549).