Resolving "orphaned" non-specific structures using machine learning and natural language processing methods

Abstract

Scholarly publications of biodiversity literature contain a vast amount of information in human readable format. The detailed morphological descriptions in these publications contain rich information that can be extracted to facilitate analysis and computational biology research. However, the idiosyncrasies of morphological descriptions still pose a number of challenges to machines. In this work, we demonstrate the use of two different approaches to resolve meronym (i.e. part-of) relations between anatomical parts and their anchor organs, including a syntactic rule-based approach and a SVM-based (support vector machine) method. Both methods made use of domain ontologies. We compared the two approaches with two other baseline methods and the evaluation results show the syntactic methods (92.1% F1 score) outperformed the SVM methods (80.7% F1 score) and the part-of ontologies were valuable knowledge sources for the task. It is notable that the mistakes made by the two approaches rarely overlapped. Additional tests will be conducted on the development version of the Explorer of Taxon Concepts toolkit before we make the functionality publicly available. Meanwhile, we will further investigate and leverage the complementary nature of the two approaches to further drive down the error rate, as in practical application, even a 1% error rate could lead to hundreds of errors.