Abstract

Named Entity Recognition (NER) is an important task of Natural Language Processing. Achieving good results in this task usually requires a large amount of labeled data to train models. This is especially difficult in domain-specific datasets and low-resourced languages. To mitigate the high cost of human-annotated data, data augmentation can be used. In this work, we evaluate Data Augmentation techniques for NER, focusing on domain-specific datasets in Portuguese.We employed augmentation techniques based on rules, back-translation, and large language models on four datasets of varying sizes to train Transformer-based NER models.The results showed that most techniques improved over the baseline, with the best results achieved using PP-LLM, SR, and MR.