CNN to BiLSTM: Enhancing Setswana Named Entity Recognition

Abstract In order to handle Named entity recognition (NER) in the Setswana language, this study applies transfer learning from convolutional neural networks (CNN) to bidirectional long short-term memory (BiLSTM) networks. We propose a transfer learning approach that leverages an existing CNN model trained to detect features in Setswana text data. Then, utilizing these attributes, a BiLSTM model specifically designed for NER is updated with the goal of improving performance and overcoming Setswana's lack of annotated data. Our study on Setswana NER corpora shows the effectiveness of this transfer learning method by outperforming conventional NER models. The new model has an improved F1-Score of 96% over the existing ones. This research contributes to the expansion of NER methods for languages with low resources.