Simple Yet Powerful: An Overlooked Architecture for Nested Named Entity Recognition

Proceedings of the 29th International Conference on Computational Linguistics, pages 2108–2117 October 12–17, 2022. 2108 Simple yet Powerful: An Overlooked Architecture for Nested Named Entity Recognition Matías Rojas1,2, Felipe Bravo-Marquez1,3,4, and Jocelyn Dunstan2,5,6 1Department of Computer Science, University of Chile. 2Center for Mathematical Modeling (CMM), University of Chile. 3National Center for Artificial Intelligence (CENIA), Chile. 4Millennium Institute for Foundational Research on Data (IMFD), Chile. 5Initiative for Data & Artificial Intelligence, University of Chile. 6Millenium Institute for Intelligent Healthcare Engineering (IHEALTH), ANID, Chile. Abstract Named Entity Recognition (NER) is an impor- tant task in Natural Language Processing that aims to identify text spans belonging to pre- defined categories. Traditional NER systems ignore nested entities, which are entities con- tained in other entity mentions. Although sev- eral methods have been proposed to address this case, most of them rely on complex task- specific structures and ignore potentially useful baselines for the task. We argue that this creates an overly optimistic impression of their perfor- mance. This paper revisits the Multiple LSTM- CRF (MLC) model, a simple, overlooked, yet powerful approach based on training indepen- dent sequence labeling models for each entity type. Extensive experiments with three nested NER corpora show that, regardless of the sim- plicity of this model, its performance is better or at least as well as more sophisticated meth- ods. Furthermore, we show that the MLC archi- tecture achieves state-of-the-art results in the Chilean Waiting List corpus by including pre- trained language models. In addition, we imple- mented an open-source library that computes task-specific metrics for nested NER. The re- sults suggest that metrics used in previous work do not measure well the ability of a model to detect nested entities, while our metrics pro- vide new evidence on how existing approaches handle the task. 1 Introduction Named Entity Recognition (NER) is a widely stud- ied task in Natural Language Processing (NLP) that seeks to identify text spans expressing references to predefined categories such as person names, loca- tions, and organizations (Chinchor and Robinson, 1997). NER, or in general, the task of recogniz- ing entity mentions1, has drawn the attention of the community due to its relevance in several NLP applications. Nested Named Entity Recognition is 1Mentions are defined as references to entities that could be named, nominal or pronominal (Florian et al., 2004). Figure 1: An example of a multi-label entity in the Chilean Waiting List corpus, followed by a nesting of different types. The annotation was translated from its original language. a particular case of NER where entities are nested within each other (Finkel and Manning, 2009). Tra- ditional NER models simplify the nested entities by keeping the outermost entity and removing the in- ner ones. This simplified problem is better known as flat NER and is commonly regarded as a se- quence labeling problem since each token can be associated with at most one label. However, re- moving part of these entities could be a problem in model performance due to losing relevant informa- tion and …

Content truncated to 500 words