Named Entity: Extended Feature Analysis for Improved Recognition for Yorùbá

The research assesses the significance of four distinct features—Part of Speech (PoS), surrounding words, prefixes, and suffixes—within a Named Entity Recognition (NER) task specifically focused on the Yorùbá language. Employing a machine learning methodology, the study utilizes a Conditional Random Fields (CRF)-based approach to construct a NER system tailored for the Yorùbá language. Additionally, the research adopts a "Leave-one-out-do-the-rest" experimental design to evaluate the influence of each feature on the recognition process. A mini PoS tagger dedicated to the Yorùbá language has been trained using CRFsuite of scikit-learn as part of this research. The dataset is systematically divided into 80% for training purposes and 20% for testing. The findings indicate that the omission of PoS tags led to a marked reduction in Recall, Precision, and F-measure. In contrast, while the exclusion of surrounding words also resulted in decreases in these metrics, the impact was notably less pronounced. Performance measured by F1 score dropped by 8.07% when the context of the words was ignored while absence of POS tags reduced performance by 7.04%. The removal of prefixes and suffixes demonstrated a relatively minor effect on the overall performance of the system. Conclusively, the research asserts that PoS tags and the words within the surrounding context emerge as the most critical features for effective NER modeling in the Yorùbá language. The integration of PoS tags appears to elucidate the enhanced performance observed, even when employing a machine learning approach that may be deemed less robust.