MEERQAT-IRIT at SemEval-2023 Task 2: Leveraging Contextualized Tag Descriptors for Multilingual Named Entity Recognition

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 878–884 July 13-14, 2023 ©2023 Association for Computational Linguistics MEERQAT-IRIT at SemEval-2023 Task 2: Leveraging Contextualized Tag Descriptors for Multilingual Named Entity Recognition Jesús Lovón-Melgarejo ♠, Jose G. Moreno ♠, Romaric Besançon ♢ Olivier Ferret ♢, Lynda Tamine ♠ ♠Université Paul Sabatier, IRIT, UMR 5505 CNRS, Toulouse, France ♢Université Paris-Saclay, CEA, List, F-91120, Palaiseau, France {jesus.lovon, jose.moreno, tamine}@irit.fr {romaric.besancon, olivier.ferret}@cea.fr Abstract This paper describes the system we submitted to the SemEval 2023 Task 2 Multilingual Com- plex Named Entity Recognition (MultiCoNER II) in four monolingual tracks (English, Span- ish, French, and Portuguese). Considering the low context setting and the fine-grained tax- onomy presented in this task, we propose a system that leverages the language model rep- resentations using hand-crafted tag descriptors. We explored how integrating the contextual- ized representations of tag descriptors with a language model can help improve the model performance for this task. We performed our evaluations on the development and test sets used in the task for the Practice Phase and the Evaluation Phase respectively. 1 Introduction Named Entity Recognition (NER) (Grishman and Sundheim, 1996) consists in detecting groups of words as named entities from a given sentence and recognizing their type from an available list of en- tity tags. According to the entity tags list size, the NER task is classified as i) coarse-grained when the list is small (such as names of people, organi- zations, and locations, as proposed in the CoNLL task (Tjong Kim Sang and De Meulder, 2003)) or ii) fine-grained for a more extensive list (Ling and Weld, 2021). Additionally, we can distinguish different named entity types, referenced as tradi- tional (person, location) and non-traditional (titles of creative work, such as a book or a song). Non- traditional entities, also called complex entities, are difficult to recognize due to problems such as semantic ambiguity (Ashwini and Choi, 2014). The SemEval 2023 Task 2 Multilingual Com- plex Named Entity Recognition (MultiCoNER II) (Fetahu et al., 2023b) evaluates NER systems ca- pable of identifying semantically complex and am- biguous named entities in 12 languages (English, Spanish, Hindi, Bangla, Chinese, Swedish, Farsi, French, Italian, Portuguese, Ukrainian, and Ger- man) providing multilingual resources to this end. Additionally, this task version proposes a fine- grained entity taxonomy with more than 30 dif- ferent entity tags and a complex test set composed of low-context sentences divided into two subsets, named “noisy” and “clean”. The “noisy” subset (available for eight languages) includes corrupted sentences that add noise on context tokens or entity tokens to simulate errors. MultiCoNER II proposes two types of evaluations: 12 monolingual tracks and one multilingual track. A task constraint is that multilingual models can not participate in the monolingual tracks, and similarly, the monolingual models can not participate in the multilingual track. Also, a specific dataset is provided for each phase of the evaluation: we adopt hereafter the term de- velopment set for the Practice phase and the term test set for the …

Content truncated to 500 words