Confidence-Based Thresholding with Rule-Based Model for Agricultural Entity Extraction

Entity extraction is a vital baseline task in Natural Language Processing for many applications, including relation extraction, knowledge extraction and knowledge graph construction. Entity extraction in particular domains like agriculture is explored less and needs further study. Rule-based methods are still relevant even in this machine learning and deep learning era, as they provide the possibility to start with the raw textual data without any annotated dataset. This paper proposes a rule-based approach with confidence thresholding for extracting agricultural entities from text documents and classifying them into the corresponding categories of Crop, Disease, Pathogen, and Nutrient. The Confidence score for each entity is determined based on the pattern used to extract the entity, suffix of the entity and frequency of occurrence of the entity. The entities are filtered based on the confidence score, and those with a confidence score above a threshold value are taken. The results are analyzed with a sampling approach for different threshold values and obtain a precision of 75.16% for Crop, 96.83% for Disease, 98.21% for Pathogen, and 98.11% for Nutrient with a confidence threshold value of 0.7.