Development of Hybrid Algorithm for Named Entity Recognition in Uzbek Public Administration Texts

The accelerating digital transformation of public administration in Uzbekistan generates large volumes of unstructured legal text that needs to be indexed, verified, and searched in real time. This paper presents a hybrid named entity recognition framework adapted for text analysis in this area. The algorithm starts with two rule-based preprocessing steps — alphabet standardization (Cyrillic → Latin) and dialect normalization — that mitigate orthographic variability and regional lexical drift. In addition, Bidirectional LSTM coupled with convolutional neural networks and a conditional random fields layer were used as neural network technologies for the named entity recognition language model. The system is trained on a curated, manually annotated corpus of Uzbek government documents that covers five main entity types: date/ time, organization, location, person, and money. It is worth noting that the sentences in the corpus are marked up using the BIOES scheme, which provides comprehensive marking of the entire text. Experiments show that each preprocessing component provides measurable performance improvements and that the hybrid architecture outperforms the baseline Transformers.

Ҳали таржима қилинмаган