Where are we in Named Entity Recognition from Speech?

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 4514–4520 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 4514 Where are we in Named Entity Recognition from Speech? Antoine Caubri`ere1, Sophie Rosset2, Yannick Est`eve3, Antoine Laurent1, Emmanuel Morin4 1LIUM - Le Mans University, France, 2LIMSI - University of Paris-Saclay, CNRS, France, 3LIA - Avignon University, France, 4LS2N - University of Nantes, CNRS, France antoine.caubriere@univ-lemans.fr, sophie.rosset@limsi.fr, yannick.esteve@univ-avignon.fr, antoine.laurent@univ-lemans.fr, emmanuel.morin@univ-nantes.fr Abstract Named entity recognition (NER) from speech is usually made through a pipeline process that consists in (i) processing audio using an automatic speech recognition system (ASR) and (ii) applying a NER to the ASR outputs. The latest data available for named entity extraction from speech in French were produced during the ETAPE evaluation campaign in 2012. Since the publication of ETAPE’s campaign results, major improvements were done on NER and ASR systems, especially with the development of neural approaches for both of these components. In addition, recent studies have shown the capability of End-to-End (E2E) approach for NER / SLU tasks. In this paper, we propose a study of the improvements made in speech recognition and named entity recognition for pipeline approaches. For this type of systems, we propose an original 3-pass approach. We also explore the capability of an E2E system to do structured NER. Finally, we compare the performances of ETAPE’s systems (state-of-the-art systems in 2012) with the performances obtained us- ing current technologies. The results show the interest of the E2E approach, which however remains below an updated pipeline approach. Keywords: Named Entity Recognition, Automatic Speech Recognition, Tree-structured Named Entity, End-to-End 1. Introduction Named entity recognition seeks to locate and classify named entity mentions in unstructured text into pre-defined categories (such as person names, organizations, locations, ...). Quaero project (Grouin et al., 2011) is at the initia- tive of an extended definition of named entity for French data. This extended version has a multilevel tree structure, where base entities are combined to define more complex ones. With the extended definition, named entity recogni- tion consists in the detection, the classification and the de- composition of the entities. This new definition was used for the French evaluation campaign ETAPE (Galibert et al., 2014). Since the ETAPE’s results publication in 2012, no new work were published, to the best of our knowledge, on named entity recognition from speech for Quaero-like tree- structured French data. Tree-structured named entities can not be tackled as a simple sequence labeling task. At the time of the ETAPE campaign, state-of-the-art works fo- cused on multiple processing steps before rebuilding a tree structure. Conditional Random Field (Lafferty et al., 2001) (CRF) are in the core of these previous sequence labeling approaches. Some approaches (Dinarelli and Rosset, 2012; Dinarelli and Rosset, 2011) used Probabilistic Context-Free Grammar (Johnson, 1998) (PCFG) in complement of CRF to implement a cascade model. CRF was trained on compo- nents information and PCFG was used to predict the whole entity tree. The ETAPE winning …

Content truncated to 500 words