Abstract
Current state-of-the-art approaches for named entity recognition (NER) using BERT-style transformers typically use one of two different approaches: (1) The first fine-tunes the transformer itself on the NER task and adds only a simple linear layer for word-level predictions. (2) The second uses the transformer only to provide features to a standard LSTM-CRF sequence labeling architecture and thus performs no fine-tuning. In this paper, we perform a comparative analysis of both approaches in a variety of settings currently considered in the literature. In particular, we evaluate how well they work when document-level features are leveraged. Our evaluation on the classic CoNLL benchmark datasets for 4 languages shows that document-level features significantly improve NER quality and that fine-tuning generally outperforms the feature-based approaches. We present recommendations for parameters as well as several new state-of-the-art numbers. Our approach is integrated into the Flair framework to facilitate reproduction of our experiments.
Abstract (translated)
URL
https://arxiv.org/abs/2011.06993
