The object of the study is the text classification and semantic search tailored to the unique linguistic features of the Kazakh language. The research addresses the challenge of improving the accuracy, relevance, and efficiency of semantic search. This study focuses on improving semantic search for the Kazakh language by analyzing computational models tailored to its unique linguistic features, such as agglutinative morphology and rich inflectional systems. The research compares traditional rule-based approaches and advanced transformer architectures, including fine-tuned models like RoBERTa, for their ability to handle semantic nuances, contextual relationships, and user intent. The results reveal that fine-tuned transformer models achieved significant advancements, with the RoBERTa model attaining a Precision@10 of 89.4 %, a Mean Reciprocal Rank (MRR) of 85.6 %, and an F1-Score of 88.0 %. Additionally, the semantic search system developed in this study demonstrated a precision of 88.4 %, recall of 87.6 %, and an F1-score of 88.0 % on a domain-specific Kazakh dataset. Key to these improvements were innovations in preprocessing pipelines, including custom tokenization and lemmatization tailored to Kazakh's agglutinative morphology, and the integration of contextual embeddings to resolve issues such as synonymy and homonymy. Computational efficiency was enhanced through resource optimization techniques, enabling the deployment of these advanced models in constrained environments. These findings underscore the potential of tailored transformer models to bridge the gap in semantic search capabilities for underrepresented languages like Kazakh, advancing the inclusivity of natural language processing technologies

With around 1 % of the population of the Republic of Kazakhstan being affected by hearing disabilities, Kazakh Sign Language holds great importance as a means of communication between citizens of the state. The limitations of tools for Kazakh Sign Language (KSL) create significant challenges for people with hearing impairments in education, employment, and daily interactions. This research addresses these challenges through the development of an automated recognition system for Kazakh Sign Language gestures, aiming to enhance accessibility and inclusivity of communication using artificial intelligence. The approach employs advanced machine learning techniques, including Convolutional Neural Networks (CNNs) for recognizing spatial gesture patterns and Recurrent Neural Networks (RNNs) for processing temporal sequences. By combining these methods, the system recognizes both hand gestures and facial expressions, providing a dual-stream model that surpasses single-stream gesture recognition systems focused solely on hand movements. A dedicated dataset was created using Mediapipe Holistic, an open-source tool that identifies 543 landmarks across hands, faces, and poses, effectively capturing the multifaceted nature of sign language. The findings showed that the hybrid model significantly outperformed standalone CNN and RNN models, achieving up to 96 % accuracy. This demonstrates that integrating facial expressions with hand gestures greatly enhances the precision of sign language recognition. This system holds immense potential to improve inclusivity and accessibility in various settings across the Republic of Kazakhstan by facilitating communication for hearing-impaired individuals, paving the way for expanded research and application in other sign languages

The linguistic diversity of the Kazakh language poses unique challenges and opportunities for the development of automatic processing systems. This research explores a comprehensive array of methodologies employed in advancing automatic processing systems tailored for the Kazakh language. The study encompasses morphological analysis, sentiment analysis, speech recognition, machine translation, and other key facets crucial for understanding and enhancing linguistic interactions. Additionally, the research delves into the creation of lexical resources, sentiment analysis tools, and named entity recognition models specific to the Kazakh language. Drawing on interdisciplinary approaches, the study investigates the integration of Kazakh language processing into e-learning systems, social media analysis, and explores challenges and opportunities in developing a Kazakh language corpus. The incorporation of cutting-edge technologies, such as neural networks and transformer-based models, is examined to address emerging trends in language processing. Through an exploration of cross-cultural collaboration, ethical considerations, and policy implications, this research contributes to the global dialogue on inclusive language technology. The findings provide insights into the current state of Kazakh language processing research, highlight persisting challenges, and suggest avenues for future exploration and development in the field.

The paper presents a pipeline for building a high-quality, multi-layer annotated corpus for Kazakh NLP. The pipeline integrates large-scale web crawling, aggressive text cleaning, transformer-based pre-annotation, linguist-in-the-loop validation, and quality assurance via inter-annotator agreement (IAA). The final release contains 20,000 documents, 350,000 sentences, and 6.2 million tokens spanning 5 domains (news, politics, science, education, culture), annotated with POS, morphology, NER, and UD dependencies. The report IAA per layer and baseline model performance (POS, NER, parsing) to demonstrate utility. The study releases the corpus, code, and guidelines to support reproducible research. The proposed methodology includes the complete workflow from data acquisition through targeted web crawling of prominent Kazakh-language news outlets, to text cleaning, automated pre-annotation utilizing transformer-based language models, and manual validation via a custom-designed annotation interface. Particular emphasis is placed on the agglutinative characteristics of Kazakh and its extensive morphological variants, which present distinct challenges in annotation and model training. The generated corpus comprises comprehensive annotations for part-of-speech (POS), named entity recognition (NER), morphological characteristics, and syntactic dependencies, establishing a fundamental dataset for several downstream NLP applications. The research additionally examines significant obstacles in the annotation process, including maintaining consistency, assessing inter-annotator agreement, and evaluating the adaptability and functionality of the annotation tools. Baseline NLP models were employed and evaluated to determine the quality and utility of the corpus. This work provides a reproducible and flexible methodology for constructing corpora in low-resource and morphologically complex languages. Its objective is to promote additional study, tool creation, and technological progress in Kazakh language processing, hence advancing the overarching purpose of multilingual NLP inclusion.

Kazakh sequence labeling is constrained by limited annotated resources, while its agglutinative morphology and productive suffixation increase data sparsity and exacerbate label inconsistency in part-of-speech (POS) tagging and named entity recognition (NER). This paper proposes an LLM-assisted weak supervision framework in which a large language model generates synthetic token-level annotations that are subsequently filtered using confidence-based criteria and combined with a smaller manually verified subset to train Transformer-based sequence taggers with Conditional Random Field (CRF) decoding. The pipeline unifies corpus construction, weak-label generation, quality filtering, word-to-subword alignment, and CRF-refined structured prediction into a reproducible workflow. Experimental results show that contextual encoders and structured decoding provide strong performance for Kazakh POS and NER, while the proposed training design enables efficient convergence with diminishing returns beyond moderate epoch budgets. Error-slice analysis indicates that residual errors are concentrated in rare tokens, morphologically complex long words, longer sentences, and the ORG entity class. Overall, the findings support the use of LLM-assisted weak supervision as a scalable strategy for low-resource Kazakh sequence labeling when synthetic labels are controlled through filtering and refined by structured decoding.

This article is devoted to the identification of early diagnosis of respiratory lung diseases, such as chronic obstructive pulmonary disease and pneumonia, to reduce mortality and prevent complications. One of the most effective methods of structuring data is the Decision Tree model. The research focuses on the development and evaluation of a decision tree model, which is used to obtain data in the form of questionnaires, text files from patients, where they describe in detail the entire process of the disease, describing their symptoms and general condition at different time periods. There are a few criteria that patients must answer for a more accurate diagnosis. The developed methodology will allow processing relevant data with various symptoms to obtain reliable identification of the signs of the disease, as well as the stages of its progression; this can be done without the use of complex and high-tech devices that make diagnosis very accessible and feasible in the shortest possible time, if resources and time are limited. The article describes the model, carefully collected, and processed the necessary data, and then the results will be described in detail, covering many indicators such as accuracy, responsiveness, F1 score and ROC-AUC. The results of this analysis strongly suggest that this model is effective enough to provide a high level of accuracy combined with extensive capabilities, which determines its practical importance for use in real conditions. It is noted that the decision tree model can significantly improve the quality of diagnostics, since it is possible to structure a large amount of data and thus collect many years of human experience.