Recent deep learning approaches have shown promising results for named entity recognition (NER). A reasonable assumption for training robust deep learning models is that a sufficient amount of high-quality annotated training data is available. However, in many real-world scenarios, labeled training data is scarcely present. In this paper we consider two use cases: generic entity extraction from financial and from biomedical documents. First, we have developed a character based model for NER in financial documents and a word and character based model with attention for NER in biomedical documents. Further, we have analyzed how transfer learning addresses the problem of limited training data in a target domain. We demonstrate through experiments that NER models trained on labeled data from a source domain can be used as base models and then be fine-tuned with few labeled data for recognition of different named entity classes in a target domain. We also witness an interest in language models to improve NER as a way of coping with limited labeled data. The current most successful language model is BERT. Because of its success in state-of-the-art models we integrate representations based on BERT in our biomedical NER model along with word and character information. The results are compared with a state-of-the-art model applied on a benchmarking biomedical corpus.

This paper explores the question of how predictive uncertainty methods perform in practice in Natural Language Processing, specifically multi-class and multi-label text classification. We conduct benchmarking experiments with 1-D convolutional neural networks and pre-trained transformers on six real-world text classification datasets in which we empirically investigate why popular scalable uncertainty estimation strategies ( Monte-Carlo Dropout , Deep Ensemble ) and notable extensions ( Heteroscedastic , Concrete Dropout ) underestimate uncertainty. We motivate that uncertainty estimation benefits from combining posterior approximation procedures, linking it to recent research on how ensembles and variational Bayesian methods navigate the loss landscape. We find that our proposed method combination of Deep Ensemble with Concrete Dropout , by analysis of in- domain calibration, cross-domain classification, and novel class robustness, demonstrates superior performance, even at a smaller ensemble size. Our results corroborate the importance of fine-tuning dropout rate to the text classification task at hand, which individually and as an ensemble impacts model robustness. We observe in ablation that pre-trained transformers severely underperform in novelty detection, limiting the applicability of transfer learning when distribution shift from novel classes can be expected.