A comprehensive dataset and neural network approach for named entity recognition in the Uzbek language
Abstract
In this study, the authors presented a dataset for named entity recognition in the Uzbek language. The dataset consists of 2000 sentences and 25,865 words, and the sources were legal documents and hand-crafted sentences annotated using the BIOES scheme. The study is complemented by the fact that the authors demonstrated the applications of the created dataset by training a language model using the CNN + LSTM architecture, which achieves high accuracy in NER tasks, with an F1 score of 90.8 %, precision of 93.9 %, and recall of 88.0 % on the test set. The proposed dataset and trained model contribute to the development of natural language processing in the Uzbek language. In addition, the authors also conducted an analysis of existing works, as well as a comparative analysis, which will help to identify the distinctive features and novelty of the proposed work. Moreover, in conclusion, the authors propose possible scenarios for the development of the work, in the form of further scaling of the dataset, as well as the use of other neural network architectures.