Semi-supervised Named Entity Recognition for Low-Resource Languages Using Dual PLMs

Named Entity Recognition (NER) plays a crucial role in natural language processing (NLP) tasks by identifying and classifying named entities. However, developing high-performing NER models for low-resource languages remains challenging due to the limited availability of labeled data. This paper proposes a semi-supervised data augmentation approach that combines two state-of-the-art pre-trained language models (PLMs). Our method first fine-tunes two PLMs using a small set of labeled data, then uses them to generate weakly supervised data from unlabeled data through collaborative learning. These predictions are then evaluated using confidence scores and an agreement measurement by both models to generate a high-quality dataset. We perform experiments using seven low-resource, but widely spoken African languages, demonstrating that augmented datasets generated by our approach achieve better results in six out of the seven languages. Furthermore, we conduct cross-lingual zero-shot experiments between language pairs and multi-lingual experiments to validate the robustness of our method.