MLlab4CS at SemEval-2023 Task 2: Named Entity Recognition in Low-resource Language Bangla Using Multilingual Language Models

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 1388–1394 July 13-14, 2023 ©2023 Association for Computational Linguistics MLlab4CS at SemEval-2023 Task 2: Named Entity Recognition in low-resource language Bangla using Multilingual Language Models Shrimon Mukherjee∗, Madhusudan Ghosh∗, Girish, Partha Basuchowdhuri Indian Association for the Cultivation of Science, Kolkata {shrimonmukherjee, madhusuda.iacs, girishg1502}@gmail.com, partha.basuchowdhuri@iacs.res.in Abstract Extracting of NERs from low-resource lan- guages and recognizing their types is one of the important tasks in the entity extraction do- main. Recently many studies have been con- ducted in this area of research. In our study, we introduce a system for identifying com- plex entities and recognizing their types from low-resource language Bangla, which was pub- lished in SemEval Task 2 MulitCoNER II 2023. For this sequence labeling task, we use a pre- trained language model built on a natural lan- guage processing framework. Our team name in this competition is MLlab4CS. Our model MuRILCRF Bangla produces a macro average F- score of 76.27%, which is a comparable result for this competition. 1 Introduction An interesting and well-studied area of natural lan- guage processing (NLP) is the extraction of entities and detecting relationships between entities from unstructured text. Named entity recognition (NER) is a sub-task of information extraction, where enti- ties are extracted from a given text and classified into some pre-defined classes. NER has many practical use cases in various NLP tasks, including text summarization (Toda and Kataoka, 2005), information retrieval, machine translation (Babych and Hartley, 2003), and ques- tion answering (Mollá et al., 2006). A wide range of methodologies have been investigated by re- searchers to perform NER tasks including rule- based techniques (Krupka and IsoQuest, 2005), su- pervised feature-based approaches (Liao and Veera- machaneni, 2009), unsupervised methods (Feld- man and Rosenfeld, 2006), deep learning-based methods (Li et al., 2020) and transformer-based approaches (Vaswani et al., 2017). Recent applications of pre-trained language mod- els (PLMs) have shown significant improvement *These authors contributed equally to this work in entity extraction and relation extraction tasks. PLMs can learn highly accurate linguistic, se- mantic, and factual information from a sizable amount of unlabeled data through self-supervised pre-training (Wang et al., 2022). The second edition of the shared task was conducted in SemEval 2023 and was named Multilingual Complex Named Entity Recognition (MultiCoNER II). The goal of this task was to build a system, which is capable of extracting fine-grained complex entities of different cate- gories, as provided in the data. The complete list of categories is - Facility, OtherLOC, Hu- manSettlement, Station, VisualWork, Musical- Work, WrittenWork, ArtWork, Software, Musical- GRP, PublicCorp, PrivateCorp, AerospaceManu- facturer, SportsGRP, CarManufacturer, ORG, Sci- entist, Artist, Athlete, Politician, Cleric, Sports- Manager, OtherPER, Clothing, Vehicle, Food, Drink, OtherPROD, Medication/Vaccine, Medical- Procedure, AnatomicalStructure, Symptom, and Disease (Fetahu et al., 2023b). The organizers (Fetahu et al., 2023b) invited the participants to submit NER models for each of the 13 languages listed on the website. They provided a different dataset (Fetahu et al., 2023a) for each language. Each dataset was made up …

Content truncated to 500 words