Evaluation of Neural Network Transformer Models for Named-Entity Recognition on Low-Resourced Languages

Abstract

Neural Network (NN) models produce state-of-the-art results for natural language processing tasks. Further, NN models are used for sequence tagging tasks on low-resourced languages with good results. However, the findings are not consistent for all low-resourced languages, and many of these languages have not been sufficiently evaluated. Therefore, in this paper, transformer NN models are used to evaluate named-entity recognition for ten low-resourced South African languages. Further, these transformer models are compared to other NN models and a Conditional Random Fields (CRF) Machine Learning (ML) model. The findings show that the transformer models have the highest F-scores with more than a 5% performance difference from the other models. However, the CRF ML model has the highest average F-score. The transformer model's greater parallelization allows low-resourced languages to be trained and tested with less effort and resource costs. This makes transformer models viable for low-resourced languages. Future research could improve upon these findings by implementing a linear-complexity recurrent transformer variant.