LitTx: A New Treatment Relation Extraction Dataset

Area of Science:

• Biomedical informatics
• Natural Language Processing
• Knowledge Representation

Background:

• Biomedical relation extraction (RE) is crucial for building knowledge graphs, which enhance Large Language Model (LLM) applications by preventing hallucinations.
• Therapy-disease treatment relations are vital for identifying emerging therapeutic hypotheses and off-label drug use from scientific literature.
• Manual extraction of these relations is infeasible due to the exponential growth of biomedical publications.

Purpose of the Study:

• To address the lack of recent expert-annotated datasets for biomedical RE.
• To introduce a novel dataset, LitTx, specifically designed for identifying treatment relationships in biomedical literature.
• To include a new 'conditional treatment' relation type to capture hedging and potential therapeutic associations.

Main Methods:

• Development of the LitTx dataset, featuring expert annotations for treatment relationships.
• Inclusion of confirmed, implied positive, and conditional treatment relation types.
• Establishment of baseline RE models trained and evaluated on the LitTx dataset.

Main Results:

• Baseline RE models demonstrated promising performance on the LitTx dataset.
• The results indicate clear opportunities for further improvement in biomedical RE.
• The study highlights the utility of the LitTx dataset for advancing RE research.

Conclusions:

• The LitTx dataset provides a valuable resource for the biomedical RE community.
• The dataset facilitates the automatic extraction of evolving knowledge bases of therapy-disease treatment relations.
• Public release of the dataset, code, and guidelines aims to foster innovation and ensure reproducibility.