No 4 (2024)

The paper considers methods for formalizing the description of multi-level structural synthesis of a complex technical system in a limited natural language at different stages of conceptual design. Logical-linguistic models and tools are proposed that allow generating the text of description of a technical system structure and the design process in Russian. An information model of a terminological dictionary for describing a technical solution is constructed. A linguistic processor implementing text generation is developed. Examples of natural-language descriptions of the stages of structural synthesis of a technical device are given.

The article studies the recognition of special structural segments of genomes called promoters. To solve the problem of promoter recognition machine learning methods based on logical analysis and data classification were used for the first time. These methods are based on searching for informative fragments in feature descriptions of precedents and are focused on processing low-value integer information. The fragments found are well interpretable and allow distinguishing promoters from other regions of the genome. However, their search is time-consuming. The results of experiments on an unbalanced sample of a large volume are presented, considering both the traditional method of feature formation using k-meres and the method of direct application of the logical classifier to the original data. It is shown that in the second case, the quality of logical classification is significantly higher and amounts to 94.3% according to ROC-AUC using the ensemble approach. The best result, namely, an ROC-AUC accuracy of 95.1%, was shown by the CatBoost classifier when directly applied to the original sample. With the traditional method of feature generation, the accuracy of CatBoost is 94.8%.

Currently, one of the main directions of information technology development is graph-based modeling of complex data structures and machine learning approaches based on graph representations. The article deals with graph modeling of text data using neural networks. The aim of the paper is to develop a graph neural network for classification and clustering of texts based on semantic content. Texts are represented as graphs, where vertices are concepts and edges are links between them. Public text corpora in Russian and English were used. A new approach to analyzing text data was proposed based on their representation in the form of oriented weighted graphs and processing by graph neural networks. The graphs were processed by a neural network with three layers of graph convolutions. The obtained results show an accuracy of more than 90% for topic group classification and text clustering, outperforming RNN, CNN and doc2vec methods.

The purpose of this paper is to advance and automate language models for extracting statements related to events and factors from text documents using the designed linguistic marker system. The paper presents the outcomes of text-mining models of events and factors extraction approbation on the example of analytical research in human potential, social sciences and humanities. The testing and evaluation of the used linguistic models are performed on the basis of the results comparison obtained in automatic mode, in manual mode (with the participation of expert-analytical validation) and semiautomatic mode (using the implemented system of linguistic markers). The introduced approaches resulted in higher performance in extracting statements containing events and factors.

To increase the transparency of modern computer-aided diagnosis (CAD) systems for assessing the malignancy of lung nodules, an interpretable model based on applying the generalized additive models and the concept-based learning is proposed. The model detects a set of clinically significant attributes in addition to the final malignancy regression score and learns the association between the lung nodule attributes and a final diagnosis decision as well as their contributions into the decision. The proposed concept-based learning framework provides human-readable explanations in terms of different concepts (numerical and categorical), their values, and their contribution to the final prediction. Numerical experiments with the LIDC-IDRI dataset demonstrate that the diagnosis results obtained using the proposed model, which explicitly explores internal relationships, are in line with similar patterns observed in clinical practice. Additionally, the proposed model shows the competitive classification and the nodule attribute scoring performance, highlighting its potential for effective decision-making in the lung nodule diagnosis.