The aim of this study is to develop an approach for designing neural network architectures for text analysis based on the conceptual-graph method (CGM). Particular attention is given to the integration of semantic knowledge and domain-specific structures into neural models in order to enhance the accuracy, interpretability, and robustness of natural language processing systems. The proposed approach seeks to overcome the limitations of traditional neural architectures, which often operate as “black boxes” and fail to consider the deep semantic relationships between concepts. Despite their high performance, modern deep learning models suffer from several limitations, including limited interpretability, susceptibility to overfitting, and an inability to effectively capture contextual and semantic relationships inherent to specific domains. This is particularly problematic in text analysis tasks, where understanding of concept hierarchies, dependencies, and semantic context is critical. The absence of explicit knowledge representation complicates the use of such models in practical and expert systems that require transparency and domain-specific alignment. This paper presents a methodology for employing CGM to construct semantic graphs that represent the relationships between key concepts in a text. A mechanism has been proposed for transforming these graphs into structural components of a neural architecture, thereby enhancing the model’s capacity for semantic interpretation. Examples of conceptual modeling graphs have been provided, demonstrating how semantic structures can be integrated into the training process. The benefits of CGM have been discussed, including improved predictive accuracy, reduced overfitting due to structural constraints, and enhanced robustness to noise and incomplete data. The proposed approach is valuable for developers of artificial intelligence systems operating in highly specialized domains, such as medicine, law, education, and technical expertise, where high accuracy and interpretability are essential. By formalizing expert knowledge as conceptual graphs and embedding them into the learning process, the approach enables the creation of models capable of meaningful information processing. This paves the way for hybrid intelligent systems that combine symbolic and neural methods, which is especially relevant in the context of explainable artificial intelligence (XAI) development.
Semenov et al. (Sun,) studied this question.