Hybrid model of speech act classification: combination of RuBERT embeddings and logical-algebraic implicants

The subject of the study is the problem of interpretable automatic classification of speech acts in Russian-language utterances, which represents one of the key tasks of modern computational linguistics and applied natural language analysis. Traditional machine learning methods and transformer models demonstrate high quality; however, they remain insufficiently transparent, hindering their application in critical areas where explainability of decisions is required. This work explores the possibility of combining contextual embeddings from RuBERT with logical-algebraic mechanisms for feature interpretation aimed at identifying stable latent structures corresponding to various types of speech acts: assertions, questions, requests, and assumptions. The study encompasses the processes of corpus formation, extraction of vector representations, construction of a classification model, and analysis of how logical rules can enhance the reliability of interpretation of results. The methodology is based on a combination of RuBERT contextual embeddings, threshold binary feature binarization, and logical-algebraic extraction of minimal implicants, applied together with a linear classifier in a hybrid architecture. The scientific novelty of the research lies in the development of a hybrid approach that combines the strengths of neural network and symbolic methods: continuous embeddings from RuBERT are used to form informative representations of text, while logical-algebraic implicants provide transparent interpretation of the classifier’s decisions. The proposed model demonstrates that it is possible to construct compact logical rules over high-dimensional embeddings, describing stable regions of the feature space and enhancing trust in the system's operation. Experimental results show that the hybrid architecture surpasses the linear classifier by 3-4 percentage points, achieving high performance. The findings confirm that logical-symbolic structures can simultaneously improve the accuracy and interpretability of models, making the approach promising for Explainable AI and the development of dialog and analytical language systems.