Open Information Extraction Methodology for a New Curated Biomedical Literature Dataset

Abstract

The research articles contain a wealth of information about the interactions between biomedical entities. However, manual relation extraction processing from the literature by domain experts takes a lot of time and money. In addition, it is often prohibitively expensive and labor-intensive, especially in biomedicine where domain knowledge is required. For this reason, computer strategies that can use unlabeled data to lessen the load of manual annotation are of great relevance in biomedical relation extraction. The present study solves relation extraction tasks in a completely unsupervised scenario. This article presents an unsupervised model for relation extraction between medical entities from PubMed abstracts, after filtration and preprocessing the abstracts. The verbs and relationship types are embedded in a vector space, and each verb is mapped to the relation type with the highest similarity score. The model achieves competitive performance compared to supervised systems on the evaluation using ChemProt and DDI datasets, with an F1-score of 85.8 and 88.5 respectively. These improved results demonstrate the effectiveness of extracting relations without the need for manual annotation or human intervention.