FII_Better at SemEval-2023 Task 2: MultiCoNER II Multilingual Complex Named Entity Recognition

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 1107–1113 July 13-14, 2023 ©2023 Association for Computational Linguistics FII_Better at SemEval-2023 Task 2: MultiCoNER II Multilingual Complex Named Entity Recognition Viorica-Camelia Lupancu1 Alexandru-Gabriel Pl˘atic˘a1 Cristian-Mihai Ros,u1 Daniela Gîfu1,2 Diana Trandab˘at, 1 1Faculty of Computer Science, Alexandru Ioan Cuza University of Iasi, Romania 2Institute of Computer Science, Romanian Academy - Iasi Branch {lupancu.camelia.99, platicaalex, cristian.rosu453}@gmail.com daniela.gifu@iit.academiaromana-is.ro diana.trandabat@gmail.com Abstract The “MultiCoNER II Multilingual Complex Named Entity Recognition” task1 at SemEval 2023 competition focuses on identifying com- plex named entities (NEs), such as the titles on creative works (songs, books, movies), in sev- eral languages. In the context of SemEval, our team, FII_Better, presents an exploration of a base transformer model’s capabilities regarding the task, focused more specifically on five lan- guages (English, Spanish, Swedish, German, Italian). We take DistilBERT and BERT as two examples of basic transformer models, us- ing DistilBERT as a baseline and BERT as the platform to create an improved model. In this process, we learned a lot about working with transformers and, on top of that, we managed to get fair results on the chosen languages. 1 Introduction Named entity recognition (NER) is a subtask of nat- ural language processing (NLP), being still a cru- cial learning problem (Lample et al. (2016), Zhang et al. (2022)). NER is used especially for relevant information extraction from text data (Shao et al. (2016), Gifu and Vasilache (2014)). It is impressive that state-of-the-art NER systems rely heavily on hand-crafted features and domain-specific knowl- edge (Cristea et al., 2016). 1https://multiconer.github.io The MultiCoNER II shared task (Fetahu et al., 2023b) aims at building NER systems for 12 lan- guages, namely English, Spanish, Hindi, Bangla, Chinese, Swedish, Farsi, French, Italian, Por- tuguese, Ukrainian and German. The task has 12 monolingual tracks and a multilingual one. NER helps one easily identify key elements in a text, like athletes, politicians, human settlements, clothing pieces, medications and more. Extracting the main entities in a text helps sorting unstructured data and detecting important information, which is crucial if one has to deal with large datasets. The datasets mainly contain sentences from three domains: Wikipedia, web questions and user queries which are usually short and low-context sentences (Malmasi et al. (2022a), Fetahu et al. (2023a)). Moreover, these short sentences usually contain semantically ambiguous and complex enti- ties, which makes the problem more difficult. Usu- ally, retrieving knowledge related to such ambigu- ous concepts in any form (like an article or results of a web search) is a definite method of understand- ing and disambiguating them. Thus, the ideal NER model would be capable of taking on hard samples if the option of additional context information was available. The rest of the paper is organized as follows: section 2 briefly presents studies related to NER, either in a multilingual context or not, section 3 1107 presents the dataset, the required pre-processing and plausible methods for it, section 4 resumes the results of the conducted experiments, with their …

Content truncated to 500 words