Nested Named Entity Recognition Revisited

Proceedings of NAACL-HLT 2018, pages 861–871 New Orleans, Louisiana, June 1 - 6, 2018. c⃝2018 Association for Computational Linguistics Nested Named Entity Recognition Revisited Arzoo Katiyar and Claire Cardie Department of Computer Science Cornell University Ithaca, NY, 14853, USA {arzoo, cardie}@cs.cornell.edu Abstract We propose a novel recurrent neural network-based approach to simultane- ously handle nested named entity recogni- tion and nested entity mention detection. The model learns a hypergraph represen- tation for nested entities using features ex- tracted from a recurrent neural network. In evaluations on three standard data sets, we show that our approach significantly outperforms existing state-of-the-art meth- ods, which are feature-based. The ap- proach is also efficient: it operates lin- early in the number of tokens and the num- ber of possible output labels at any token. Finally, we present an extension of our model that jointly learns the head of each entity mention. 1 Introduction Named entity recognition (or named entity detec- tion) is the task of identifying text spans associated with proper names and classifying them according to their semantic class such as person, organiza- tion, etc. It is related to the task of mention detec- tion (or entity mention recognition) in which text spans referring to named, nominal or prominal en- tities are identified and classified according to their semantic class (Florian et al., 2004). Both named entity recognition and entity mention detection are fundamental components in information extrac- tion systems: several downstream tasks such as re- lation extraction (Mintz et al., 2009), coreference resolution (Chang et al., 2013) and fine-grained opinion mining (Choi et al., 2006) rely on both. Many approaches have been successfully em- ployed for the tasks of named entity recognition and mention detection, including linear-chain con- ditional random fields (Lafferty et al., 2001) and semi-Markov conditional random fields (Sarawagi and Cohen, 2005). However, most such methods suffer from an inability to handle nested named entities, nested entity mentions, or both. As a re- sult, the downstream tasks necessarily ignore these nested entities along with any semantic relations among them. Consider, for example, the excerpts below: (S1) Employing the [EBV - transformed [human B cell line]CELL LINE ]CELL LINE SKW6.4, we demonstrate . . . (S2) . . . [the burial site of [Sheikh Abbad]PERSON ]LOCATION is located . . . S1 shows a nested named entity from the GENIA dataset (Ohta et al., 2002): “human B cell line” and “EBV - transformed human B cell line” are both considered named entities of type CELL LINE where the former is embedded inside the latter. S2, derived from the ACE corpora1, shows a PERSON named entity (“Sheikh Abbad”) nested in an en- tity mention of type LOCATION (“the burial site of Sheikh Abbad”). Most existing methods for named entity recognition and entity mention de- tection would miss the nested entity in each sen- tence. Unfortunately, nested entities can be fairly com- mon: 17% of the entities in the GENIA corpus are embedded within another entity; in the ACE corpora, 30% of sentences …

Content truncated to 500 words