L3I++ at SemEval-2023 Task 2: Prompting for Multilingual Complex Named Entity Recognition

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 807–814 July 13-14, 2023 ©2023 Association for Computational Linguistics L3i++ at SemEval-2023 Task 2: Prompting for Multilingual Complex Named Entity Recognition Carlos-Emiliano Gonz´alez-Gallardo1 and Thi Hong Hanh Tran1,2 and Nancy Girdhar1 and Emanuela Boros1 and Jose G. Moreno1,3 and Antoine Doucet1 1 University of La Rochelle, L3i, La Rochelle, France 2 Joˇzef Stefan Institute, Ljubljana, Slovenia 3 University of Toulouse, IRIT, Toulouse, France {firstname.lastname}@univ-lr.fr jose.moreno@irit.fr Abstract This paper summarizes the participation of the L3i laboratory of the University of La Rochelle in the SemEval-2023 Task 2, Multilingual Com- plex Named Entity Recognition (MultiCoNER II). Similar to MultiCoNER I, the task seeks to de- velop methods to detect semantic ambiguous and complex entities in short and low-context settings. However, MultiCoNER II adds a fine-grained entity taxonomy with over 30 entity types and corrupted data on the test partitions. We approach these com- plications following prompt-based learning as (1) a ranking problem using a seq2seq framework, and (2) an extractive question-answering task. Our findings show that even if prompting techniques have a similar recall to fine-tuned hierarchical lan- guage model-based encoder methods, precision tends to be more affected. 1 Introduction In SemEval-2022, Multilingual Complex Named Entity Recognition (MultiCoNER) I (Malmasi et al., 2022a) (Task 11) aimed at developing complex NER systems for 11 languages. The task focused on detecting semantically ambiguous and complex entities in short, lowercase, low-context monolin- gual, and multilingual settings. These challenges needed to be approached, while testing the domain and language adaption capability of the NER sys- tems proposed by the participants. Thus, our partic- ipation consisted of a robust approach for tackling the following challenges for the language-specific, multilingual, and mixed sub-tasks (Boros¸ et al., 2022): (1) short texts: we built multilingual con- texts with and without entities, in order to improve the focus of the NER system more on the context and less on the other surrounding entities; (2) lowercase texts: we prioritized the usage of un- cased language models; and (3) code-mixed and low-resource languages: we added a multilingual language model along with a language-specific lan- guage model. Our findings showed that, while adding contexts (from train partitions), with and without entities, was promising, the topics or do- mains overlap influenced the performance in both directions. In SemEval-2023, MultiCoNER II (Fetahu et al., 2023b) (Task 2) featured complex NER in these languages (slight changes in comparison with the previous MultiCoNER): English, Spanish, Hindi, Bangla, Chinese, Swedish, Farsi, French, Italian, Portuguese, Ukrainian, and German. In addition to multilingualism, there were two new key chal- lenges: (1) a fine-grained entity taxonomy with over 30 different classes, and (2) simulated errors added to the test set to make the task more realistic and difficult. Approaching these challenges would require an ability to easily adapt to the taxonomies of entities that are constantly evolving and to new scenarios with noise (for which there are few or no labelled data). Recently, a new paradigm based on prompt-based learning …

Content truncated to 500 words