Sartipi-Sedighin at SemEval-2023 Task 2: Fine-grained Named Entity Recognition with Pre-trained Contextual Language Models and Data Augmentation from Wikipedia

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 565–579 July 13-14, 2023 ©2023 Association for Computational Linguistics Sartipi-Sedighin at SemEval-2023 Task 2: Fine-grained Named Entity Recognition with Pre-trained Contextual Language Models and Data Augmentation from Wikipedia Amir Sartipi, Amirreza Sedighin, Afsaneh Fatemi, and Hamidreza Baradaran Kashani Faculty of Computer Engineering, University of Isfahan, Iran {amirsartipi.msc, amirreza.seddighin1376, a_fatemi, hrb.kashani}@eng.ui.ac.ir Abstract This paper presents the system developed by the Sartipi-Sedighin team for SemEval 2023 Task 2, which is a shared task focused on multilingual complex named entity recognition (NER), or MultiCoNER II. The goal of this task is to identify and classify complex named enti- ties (NEs) in text across multiple languages. To tackle the MultiCoNER II task, we leveraged pre-trained language models (PLMs) fine-tuned for each language included in the dataset. In addition, we also applied a data augmentation technique to increase the amount of training data available to our models. Specifically, we searched for relevant NEs that already existed in the training data within Wikipedia, and we added new instances of these entities to our training corpus. Our team achieved an overall F1 score of 61.25% in the English track and 71.79% in the multilingual track across all 13 tracks of the shared task that we submitted to. 1 Introduction The MultiCoNER 2023 task 2 was initiated with the purpose of developing NER systems that can accurately detect fine-grained NEs across multiple languages. The shared task was organized into 13 tracks, with 12 monolingual tracks and one mul- tilingual track, to facilitate a thorough evaluation of the participating systems (Fetahu et al., 2023b). Despite the inherent complexity and ambiguity of the dataset instances, the task presented two main features that are worth mentioning. The first feature was the identification of fine-grained NEs, which required the systems to detect and classify a wide range of entities with varying levels of specificity. The second feature involved the augmentation of test data for some languages with simulated errors to increase the difficulty and realism of the task (Fetahu et al., 2023a). These features posed signif- icant challenges for the participating systems and necessitated the use of advanced NLP techniques. The work presented in this paper makes two main Figure 1: Overall process of fine-tuning NER system contributions to the field of NER. 1. We introduce a simple yet effective method for increasing the number of instances in training datasets. 2. We fine-tune (PLMs) for each language in both the multilingual track and monolingual tracks using both the original dataset and the augmented version. The overall architecture of the model used for fine- tuning can be seen in Figure 1. 2 Related Work NER is a natural language processing (NLP) task that involves identifying and classifying NEs in text, such as person names, organization names, location names, and others, into predefined categories (Grishman and Sundheim, 1996). NER is widely used in many NLP applications, such as information extraction (Tan, 2022), text summa- rization (Khademi and Fakhredanesh, 2020), and question answering …

Content truncated to 500 words