Objectives. To study the limitations of classical approaches to generating test patterns for controlled random tests based on enumerating test set candidates through their one-dimensional scaling. To address the problem of constructing controlled random tests using an iterative method for two-dimensional scaling of initial templates. The main goal of the article is to develop a method for constructing tests based on initial templates and expanding them to the required bit size and number of test patterns using an iterative procedure.

Methods. For two-dimensional scaling of initial templates with given characteristics, scaling matrices are used, which, like templates, can also be controlled random tests. Statistical testing method was used during the experimental research.

Results. It is shown that methods for constructing controlled random tests based on the use of templates can be considered as a procedure for scaling controlled random tests to the required bit size. To construct the desired tests, both templates characterized by a minimum test suite capacity and any controllable random tests are used. This procedure allows increasing the test suite capacity while maintaining the number of their patterns. A simultaneous increase in the suite capacity and their number is achieved using the proposed approach, which is based on iterative two-dimensional scaling of templates using scaling matrices. In this case, the resulting controllable random tests are generated without the labor-intensive procedure of listing candidate test suites and calculating the difference measure(s) for them. The dependences of the main characteristics of the resulting controllable random test on the characteristics of the template and the scaling matrix are presented, which, like a template, can also represent a controllable random test. A statement is proved that determines the dependence of the characteristics of the test generated at the k-th iteration on the values of the characteristics of the test obtained at the (k–1)-th iteration and the scaling test. Useful consequences and properties of tests constructed based on the proposed procedure are presented. The performance and effectiveness of an iterative method for constructing controlled random tests are demonstrated and evaluated for binary test sets. It is shown that controlled random tests constructed using the discussed procedure have significantly larger Hamming distances compared to random tests.

Conclusion. An iterative method for constructing controlled random tests through two-dimensional scaling is considered. The basis of the proposed method is the use of initial templates and scaling matrices, which represent controlled random tests with a small number of test sets and a small bit size. It is shown that the use of various templates and their two-dimensional scaling allows for the construction of controlled random tests with the required bit size and a large number of test patterns.

Objectives. The research is conducted in the field of specialized generative neural networks for the Belarusian language. The authors aim to take the first step towards building a national generative language model.

Methods. The paper describes the development process of the BelLitGPT model (700 million parameters). It is based on a transfer learning strategy using the Russian-language model ruGPT-3 and consists of three stages: corpus preparation, tokenizer adaptation methodology and model training. The training corpus is compiled from the golden fund of classic Belarusian prose and prepared Wikipedia articles. The paper details the tokenizer adaptation method for expanding the vocabulary with specific Belarusian lexemes, as well as the model training and testing process.

Results. The research results confirm that BelLitGPT can generate coherent, grammatically and stylistically correct texts. Special attention is given to the creation of a hybrid neuro-symbolic approach for generating quatrains that adhere to rhythm and rhyme.

Conclusion. The experiment on scaling the architecture revealed difficulties in training a large model (13 billion parameters) under conditions of data scarcity.

Objectives. Modern wireless mesh networks place high demands on the adaptability of routing protocols. Standard algorithms are not always able to ensure the required quality of service (QoS) due to variability in parameters such as signal-to-noise ratio, channel load, and node mobility.

Methods. A method for multi-criteria routing is proposed using an integral quality of service criterion in a modified Dijkstra's algorithm.

Results. Routing results in an eight-node network and in a network with four clusters of five devices are analyzed. A mechanism for reducing the probability of false route discards in the multi-criteria routing method is proposed.

Conclusion. An iterative algorithm for adjusting weighting coefficients has been developed. In combination with the minimax criterion, it allows to eliminate false positives regarding the absence of a QoS-feasible route and to obtain acceptable solutions for different traffic profiles. A two-level approach and routing algorithm in a clustered network have been developed, ensuring a reduction in computational complexity and localization of route recalculation when the network state changes – specifically the quality of radio channels or the energy state of nodes.

Objectives. The problem of developing a method for spatial resolution enhancement of ambisonic audio which improves the efficiency of existing approaches to digital spatial sound representation is being solved.

Methods. The proposed approach is based on time-frequency decomposition of the audio signal with subsequent extraction of a directional plane wave from each frequency component. The approach develops the core ideas of high-resolution plane wave expansion (HARPEX) and directional audio coding (DirAC) methods, utilizing the advantages of real-valued sparse decomposition.

Results. Application of the proposed ambisonic spatial resolution enhancement method involves the use of real-valued frequency components, which, unlike complex ones, provide a simpler and more robust estimation of sound arrival direction. Sparse decomposition enables an accurate and unified approach to describing sounds of various nature – from transient to tonal.

Conclusion. Practical results confirm the method's applicability for audio processing with resolution enhancement up to seventh-order ambisonics. The disadvantage of the method is its high computational complexity; however, it is suitable for applications that do not require real-time processing.

Objectives. Improvement of speech emotion recognition accuracy using Long Short-Term Memory (LSTM) recurrent neural network (RNN) models.

Methods. The paper proposes a multi-vector attention mechanism for LSTM-based RNNs. This mechanism generalizes the classical soft attention and allows the model to simultaneously analyze different aspects of temporal dependencies. The proposed RNN architectures were applied to the task of speech emotion recognition. Input data consisted of sequences of mel-frequency cepstral coefficients (MFCCs), which reflect the time-frequency structure of the speech signal. Experiments were conducted on the publicly available RAVDESS dataset. Bayesian optimization was employed for automated hyperparameter tuning of the models.

Results. The experimental results with LSTM networks having different hidden state dimensions (64, 96, 128) demonstrate that the application of the multi-vector attention mechanism leads to a statistically significant improvement in the average accuracy metric (UAR) by 0.88 to 1.56 %.

Conclusion. The obtained results confirm the effectiveness of using the proposed multi-vector attention mechanism in LSTM-based architectures for speech emotion classification.

Objectives. Tonality as a positive-negative mood is a key parameter of text analysis and generation algorithms. Existing machine learning methods encode tonality in computationally extensive and uninterpretable ways which hampers the development of the corresponding applications. The work aims to solve this problem for Russian language.

Methods. Pre-trained vector language models are used to encode words as vectors in multidimensional spaces. In these spaces, the tonality corresponds to a specific direction which optimally discriminates the positive and negative prototypes. The tonality of word is then determined by its projection onto this direction. Adding a tonal vector to a key word defines a one-dimensional subspace, containing its positive and negative associations.

Results. The algorithm is tested on GloVe and FastText machine language models, encoding individual Russian words and morphemes with vectors in 300-dimensional space. Commonly used verbs and nouns served as key words. The average reliability of the found tonal associations estimates as 80 %.

Conclusion. The results indicate the applicability of pre-trained vector language models for fast and interpretable working with tonal information. The developed approach is applicable for the tasks of aspect-based sentiment analysis, as well as for the machine generation of object-oriented texts with a required tonality. Generalization of the tonal axis to the triple of Osgood's semantic factors allows expanding the method to a full range of affectively-semantic information.