Method for constructing neural network tools for recognizing the emotional tone of texts in universal computing devices

Abstract

This paper proposes a method for constructing neural-network tools to recognize the emotional coloring of Ukrainian-language texts on “universal” computing platforms (CPU/mobile devices/GPU). The method standardizes the stages of input encoding and neural-network output decoding into probabilistic estimates, and formalizes an algorithm for selecting the architecture based on resource constraints and accuracy requirements. Three branches are provided: lightweight models for tight computational budgets, sequential recurrent models for moderate resources, and Transformer-based models for quality-first scenarios. A unified pipeline for processing text fragments is presented: tokenization/vectorization, neural transformation, normalization, and selection of the dominant emotion. Experimental evaluation of two prototypes confirms the “quality–efficiency” trade-off: RoBERTa outperforms FastText in accuracy, but requires more parameters and exhibits higher inference latency (milliseconds versus microseconds). The method targets Ukrainian and mixed social-media content, and supports optimizations for resource-constrained platforms. Its practical value lies in a guided architecture choice and a reproducible development process that balance accuracy, speed, and energy consumption. Future directions include ensembles/hybrids, handling sarcasm, and expanding the set of emotions. The results show that a systematized approach enables adaptive solutions for real-time mood monitoring and integration into mobile and embedded systems. The proposed methodology is compatible with existing corpora, metrics, and the standards of reproducibility and open science.