DEVELOPMENT OF A PROTOTYPE OF A LOCAL NEURAL NETWORK FOR AUTOMATIC ANALYSIS OF THE EMOTIONAL COLORING OF CITIZENS’ APPEALS TO LOCAL GOVERNMENT BODIES

Abstract

The article presents a prototype of a local neural network for the automatic detection of the emotional tone of citizens’ appeals to local self-government bodies and Administrative Service Centers (CNAPs). The system operates in a fully autonomous mode, requires minimal computational resources, runs on a standard office computer, and does not need an internet connection. A fully connected feed-forward neural network model was developed with a one-layer architecture containing 9924 trainable parameters. The model was implemented in Python using Google Colab environment and the TensorFlow library. For training, a dataset of 10000 examples was created through automatic generation of Ukrainian-language texts based on pre-prepared dictionaries, followed by data augmentation. Input text structuring was performed using the TF-IDF method, with the vocabulary limited to the 150 most informative features. The network training process was completed after achieving optimal values of average accuracy and prediction confidence. A desktop application was developed in C++ using Qt Creator, providing a user interface for interaction with the neural network. During the prototyping stage, a simulation mode of the neural network operation was implemented to accelerate demon strations and allow presentation and interface testing without additional computational resources. Testing of the prototype on control examples simulating real citizen appeals showed stable classification quality with confidence levels ranging from 82% to 95%. The model correctly identifies the priority of keywords corresponding to the emotional tone of the text. A comparative analysis of the developed model with alternative architectures was conducted. The proposed model has advantages in terms of minimal hardware requirements, full operational autonomy, high response speed (1–5 ms), and result interpretability. The application with the prototype confirms the feasibility of using specialized neural networks for automated analysis of the emotional tone of texts in the field of public administration and serves as a foundation for further implementation.