DaN+: Danish Nested Named Entities and Lexical Normalization

Proceedings of the 28th International Conference on Computational Linguistics, pages 6649–6662 Barcelona, Spain (Online), December 8-13, 2020 6649 DAN+: Danish Nested Named Entities and Lexical Normalization Barbara Plank, Kristian Nørgaard Jensen and Rob van der Goot Department of Computer Science ITU Copenhagen, Denmark bplank@itu.dk, krnj@itu.dk, robv@itu.dk Abstract This paper introduces DAN+, a new multi-domain corpus and annotation guidelines for Dan- ish nested named entities (NEs) and lexical normalization to support research on cross-lingual cross-domain learning for a less-resourced language. We empirically assess three strategies to model the two-layer Named Entity Recognition (NER) task. We compare transfer capabilities from German versus in-language annotation from scratch. We examine language-specific versus multilingual BERT, and study the effect of lexical normalization on NER. Our results show that 1) the most robust strategy is multi-task learning which is rivaled by multi-label decoding, 2) BERT-based NER models are sensitive to domain shifts, and 3) in-language BERT and lexical normalization are the most beneficial on the least canonical data. Our results also show that an out-of-domain setup remains challenging, while performance on news plateaus quickly. This highlights the importance of cross-domain evaluation of cross-lingual transfer. 1 Introduction Named Entity Recognition (NER) is the task of finding entities in text, such as locations, organizations, and persons. NER is a key step towards natural language understanding, for instance for question answer- ing and information extraction. The task has received a substantial amount of attention, particularly for English. Most research so far, including for Danish, focused on newswire data and flat entities. It ignores nested entities, like ‘Australian Open’ (illustrated in Figure 1) being both an event and a location-derived entity. There is also little prior work on transfer learning for nested NER. In this paper we introduce DAN+, a novel resource for Danish Nested Named entities and lexi- cal Normalization, covering texts from canonical data from newswire and non-canonical social media sources. Danish bears interesting challenges for NER similar to German, which we capture by drawing inspiration from the NOSTA-D (Benikova et al., 2014) NER annotation scheme. In particular, location adjectives like ‘dansk’ (Danish) or ‘hollandske‘ (Dutch) are not capitalized, and there are tokens which are only partially named entities, like ‘Baltica-aktierne’ (the Baltica shares). Such entities were mostly ignored so far. Full annotation guidelines for both tasks are provided in the appendix. Wozniacki vinder iik australian open Wozniacki vinder ikke Australian Open PER MISC LOCderiv Figure 1: Example with (a) nested entities and (b) lexical normalization. Contributions We present 1) DAN+, a new multi-domain dataset for nested NER and lexical nor- malization; 2) an evaluation of various models for Danish nested NER, including BERT variants and in-language versus cross-language experiments; 3) first experiments of lexical normalization on Danish and its downstream impact on NER.1 This work is licensed under a Creative Commons Attribution 4.0 International License. License details: http://creativecommons.org/licenses/by/4.0/. 1All code and data to reproduce the experiments is available at https://github.com/bplank/DaNplus 6650 2 Related Work Nested NEs have received less research focus in contrast to flat entities (Grishman …

Content truncated to 500 words