D2KLab at SemEval-2023 Task 2: Leveraging T-NER to Develop a Fine-Tuned Multilingual Model for Complex Named Entity Recognition

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 836–840 July 13-14, 2023 ©2023 Association for Computational Linguistics D2KLab at SemEval-2023 Task 2: Leveraging T-NER to Develop a Fine-Tuned Multilingual Model for Complex Named Entity Recognition Thibault Ehrhart and Raphaël Troncy EURECOM, Sophia Antipolis, France thibault.ehrhart@eurcom.fr and raphael.troncy@eurecom.fr Julien Plu Buster.ai, Paris, France plu.julien@gmail.com Abstract This paper presents D2KLab’s system used for the shared task of "Multilingual Complex Named Entity Recognition (MultiCoNER II)", as part of SemEval 2023 Task 2. The sys- tem relies on a fine-tuned transformer based language model for extracting named enti- ties. We present the architecture of the sys- tem, and we discuss our results and observa- tions. Our implementation is open sourced at https://github.com/D2KLab/multiconer. 1 Introduction Named Entity Recognition (NER) is a fundamen- tal task in Natural Language Processing (NLP) that aims to identify and classify named entities in text. While NER has been extensively studied in monolingual settings and with basic named entity types (Usbeck et al., 2015), the challenges of mul- tilingual NER are significant due to variations in language structure, orthography, and morphology. The SemEval-2023 Task 2, "Multilingual Complex Named Entity Recognition (MultiCoNER II)" (Fe- tahu et al., 2023b), provides a platform for evalu- ating systems on this challenging task across 12 languages. The key challenges of this year’s task in- clude a fine-grained entity taxonomy with over 30 different classes and simulated errors added to the test set to make the task more realistic and difficult. In this paper, we present D2KLab’s system for MultiCoNER II, which relies on a fine-tuned trans- former based language model. We also report our system’s results and observations, along with a discussion of the challenges and opportunities for multilingual NER. Our system’s performance high- lights the importance of fine-tuning on target lan- guages and the effectiveness of transformer-based models for multilingual texts. 2 Related Work In the field of Named Entity Recognition (NER), researchers have made significant progress in re- cent years, particularly in addressing the challenges associated with recognizing complex entities and limited context situations. This work builds on the progress made in the Multilingual NER task started in 2022 with the first edition of MultiCoNER (Mal- masi et al., 2022b), where the key challenges in- cluded dealing with complex entities with limited context (Malmasi et al., 2022a). Meng et al. (2021) outlined the challenges of NER in recognizing com- plex entities and in low-context situations. Further- more, Fetahu et al. (2021) extended this work to multilingual and code-mixed settings. This exten- sion was also included in the MultiCoNER dataset, and our work builds on this research. 3 Data The data collection methods used to compile the dataset used in MultiCoNER is described in (Fe- tahu et al., 2023a). The dataset comprises anno- tated sentences for 12 languages, namely Bangla, German, English, Spanish, Farsi, French, Hindi, Italian, Portuguese, Swedish, Ukrainian, and Chi- nese, along a multilingual track. The data is in the CoNLL-2002 format, which consists of one token per line …

Content truncated to 500 words