A model for constructing neural network systems forrecognizing emotions of text fragments

Abstract

Emotion Recognition in text is a crucial task in Natural Language Processing, particularly relevant given the exponential growth of textual data from social media and voice interfaces. However, developing effective emotion recognition systems for low-resource languages, such as Ukrainian, faces significant challenges, including linguistic informality, dialectal variations, and cultural specificities. This paper introduces a modular model (framework) for developing neural network-based tools for recognizing emotions in Ukrainian text fragments. The model encompasses a comprehensive data preprocessing pipeline, flexible architectural choices (including approaches based on Word to Vector, Long Short-Term Memory, and Transformers), and rigorous validation using standard metrics and interpretability methods. As part of an experimental study, two prototypes were implemented and compared: a lightweight classifier based on FastText and a more powerful classifier based on pretrained RoBERTa-base, both trained to recognize seven basic emotions. The results demonstrate that RoBERTa-base achieves high accuracy, significantly outperforming FastText and a baseline translation-based approach, yet it demands substantially more computational resources for inference. The study underscores the importance of creating Ukrainian-language corpora to enhance recognition capabilities and highlights the critical trade-off between accuracy and efficiency. It provides practical recommendations for model selection based on resource constraints and performance requirements for emotion analysis tasks in the Ukrainian language.