Grouped semantic-feature relation extraction from texts to represent medicinal-plant property knowledge on social...

This research aims to extract a grouped semantic-feature relation, particularly a PlantPart-MedicinalPropertyGroup relation which is a semantic relation between an element of a plant-part concept set and a group of medicinal-property concept features of various herbs or medicinal plants, including indigenous medicinal plants, to graphically represent medicinal-plant property knowledge from documents available on pharmacy academic websites. The medicinal-plant property knowledge representation particularly benefits native users and patients seeking alternative medical therapies during pandemics, such as COVID-19, due to limited access to medicines, physicians and hospitals. Medicinal-property expressions on the documents, particularly in Thai, are often structured as event expressions conveyed through verb phrases within Elementary Discourse Units (EDUs) or simple sentences. There are three research problems in extracting the PlantPart-MedicinalPropertyGroup relations from the documents: how to identify EDU occurrences with medicinal-property concepts, how to extract medicinal-property concept features from medicinal-property concept EDU occurrences without concept annotations, and how to extract the PlantPart-MedicinalPropertyGroup relation without relation-class labeling from the documents with the high dimensional and correlated feature consideration. To address these problems, we apply a Solving-Verb Concept set primarily sourced from translated terms on HerbMed, an American Botanical Council resource, to identify a medicinal-property concept EDU. Additionally, a word co-occurrence (word-co) pattern is applied as a compound variable on the translated terms to construct a medicinal-property-concept (MPC) table. The MPC table is employed to extract the medicinal-property concept features from the medicinal-property concept EDUs through a string-matching method. We then propose using structural equation modeling to automatically extract the PlantPart-MedicinalPropertyGroup relations from the documents. Thus, the proposed approach enables the extraction of PlantPart-MedicinalPropertyGroup relations with high qualities to represent medicinal-plant property knowledge on social media.