MHDR: Annotated Corpus for Medicinal Herb-Disease Relationships in Biomedical Articles with a Focus on Traditional Medicine.

Traditional medicine (TM) has long relied on medicinal herbs (MHs) to prevent and treat various ailments. Recently, efforts to integrate MHs with modern biomedical research have gained momentum, aiming to validate traditional knowledge and identify new therapeutic agents. However, the surge in biomedical literature presents challenges in extracting meaningful relationships, such as those between MHs and diseases. While automated methods like Natural Language Processing (NLP) and machine learning (ML) offer solutions, they often struggle with ambiguity, inconsistent terminology, and implicit meanings in texts. To address these limitations, the development of high-quality, expert-annotated corpora is essential. This study introduces the Medicinal Herb-Disease Relationships (MHDR) corpus, specifically designed to capture MH-disease relationships within the context of TM integrated with modern research. Constructed from 800 PubMed abstracts, the MHDR corpus emphasizes pharmacognostic herb names, which reflect specific plant parts and processing methods—key to accurate and fine-grained entity recognition. The corpus contains 5,119 MH mentions, 6,621 disease mentions, and 1,314 annotated relationships, extracted from 832 key sentences summarizing study findings. To evaluate its utility, baseline performance tests using Transformer-based models were conducted. With its comprehensive coverage and high-quality manual annotations, the MHDR corpus serves as a foundational resource for advancing biomedical NLP, supporting more accurate knowledge extraction, and promoting deeper exploration of MHs’ therapeutic potential through computational approaches.