Zhanibek Kozhirbayev
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Zhanibek Kozhirbayev has a total of 18 co-authors. Here are the top co-authors by number of shared publications:
- Zhandos Yessenbayev
Bina Nusantara University
8 shared publications Altynbek Sharipbay
2 shared publications- Talgat Islamgozhayev
Astana IT University
2 shared publications Zhandos Yessenbayev
National Laboratory Astana 53 Kabanbay batyr ave., Astana, Kazakhstan
2 shared publicationsAibek Makazhanov
National Laboratory Astana 53 Kabanbay batyr ave., Astana, Kazakhstan
1 shared publication- Almas Shintemirov
Nazarbayev University, Astana, Kazakhstan
1 shared publication
Latest Publications
- 19 September 2026
- Published by Springer Nature
- p. 68-81
Sequential Domain Adaptation of Whisper for Kazakh Academic Lecture Speech Recognition
- 19 September 2026
- Published by Springer Nature
- p. 68-81
Sequential Domain Adaptation of Whisper for Kazakh Academic Lecture Speech Recognition
- Open Access
Enhancing Personalised Learning with Graph-Based Ensemble Prediction and Skill Cluster Mapping for Student Knowledge Completeness
The increasing adoption of data-driven educational systems requires reliable methods to predict student readiness for future coursework and support personalised learning pathways. This study proposes a graph-enhanced ensemble framework that integrates curriculum structure and skill-gap awareness to estimate student course readiness. A global prerequisite directed acyclic graph (DAG) of university subjects was constructed to model curriculum dependencies, from which structural features including the PageRank, in-degree, out-degree, and prerequisite chain depth were derived. In parallel, a domain-informed skill cluster mapping grouped subjects into nine interpretable competency domains to enable skill-gap analysis. These curriculum-aware features were combined with academic history, behavioural engagement, and demographic indicators to produce 38 engineered features for each student–subject pair. Six models (CatBoost, XGBoost, LightGBM, FT-Transformer, MLP and TabPFN) were trained and combined using a weighted ensemble. Experiments on a real-world institutional dataset containing 20,581 students and 727,168 records achieved an AUC of 0.8908 for predicting course success. Ablation experiments demonstrate that graph-derived and skill-cluster features provide modest but statistically significant incremental value. The resulting model was integrated into a prototype personalised recommender that prioritizes curriculum-consistent learning pathways. The proposed framework provides an interpretable and curriculum-aware approach for personalised learning. While the model demonstrates strong overall performance, a moderate gender disparity in the false positive rate was observed. Results were obtained on a large longitudinal dataset from a single university, and external validation at other institutions is needed to confirm generalizability.
Open AccessEnhancing Personalised Learning with Graph-Based Ensemble Prediction and Skill Cluster Mapping for Student Knowledge Completeness
The increasing adoption of data-driven educational systems requires reliable methods to predict student readiness for future coursework and support personalised learning pathways. This study proposes a graph-enhanced ensemble framework that integrates curriculum structure and skill-gap awareness to estimate student course readiness. A global prerequisite directed acyclic graph (DAG) of university subjects was constructed to model curriculum dependencies, from which structural features including the PageRank, in-degree, out-degree, and prerequisite chain depth were derived. In parallel, a domain-informed skill cluster mapping grouped subjects into nine interpretable competency domains to enable skill-gap analysis. These curriculum-aware features were combined with academic history, behavioural engagement, and demographic indicators to produce 38 engineered features for each student–subject pair. Six models (CatBoost, XGBoost, LightGBM, FT-Transformer, MLP and TabPFN) were trained and combined using a weighted ensemble. Experiments on a real-world institutional dataset containing 20,581 students and 727,168 records achieved an AUC of 0.8908 for predicting course success. Ablation experiments demonstrate that graph-derived and skill-cluster features provide modest but statistically significant incremental value. The resulting model was integrated into a prototype personalised recommender that prioritizes curriculum-consistent learning pathways. The proposed framework provides an interpretable and curriculum-aware approach for personalised learning. While the model demonstrates strong overall performance, a moderate gender disparity in the false positive rate was observed. Results were obtained on a large longitudinal dataset from a single university, and external validation at other institutions is needed to confirm generalizability.
- Open Access
An Empirical Comparison of Cascade and Direct End-to-End Speech Translation for Low-Resource Language Pair
Speech-to-text translation (S2TT) for low-resource languages remains challenging due to the scarcity of parallel speech translation data and the susceptibility of modular pipelines to error propagation. This paper presents a controlled empirical comparison of cascade and end-to-end approaches for Kazakh–Russian speech translation using the ST-kk-ru dataset (≈332 h, 140 k triplets). The cascade framework is strengthened with recent pre-trained models for automatic speech recognition and neural machine translation, achieving 21.3 BLEU on the test set. Three representative end-to-end architectures are evaluated under identical data conditions. The strongest direct model, combining a Wav2Vec 2.0 encoder with an mBART decoder augmented by a length adaptor and adapter modules, reaches 17.97 BLEU, compared with 15.35 BLEU for FAIRSEQ S2T and 16.3 BLEU for ESPnet-ST. Automatic evaluation is complemented by expert manual assessment and targeted linguistic analysis. Results indicate that, under current low-resource conditions, cascade systems provide higher translation accuracy and better morpho-syntactic fidelity, while end-to-end models remain competitive and offer advantages in architectural simplicity and potentially reduced inference latency (due to single-pass processing), although empirical measurements were not conducted in this study. This study establishes a reproducible benchmark for Kazakh–Russian speech translation and highlights practical trade-offs between modeling paradigms in low-resource, morphologically rich settings.
Open AccessAn Empirical Comparison of Cascade and Direct End-to-End Speech Translation for Low-Resource Language Pair
Speech-to-text translation (S2TT) for low-resource languages remains challenging due to the scarcity of parallel speech translation data and the susceptibility of modular pipelines to error propagation. This paper presents a controlled empirical comparison of cascade and end-to-end approaches for Kazakh–Russian speech translation using the ST-kk-ru dataset (≈332 h, 140 k triplets). The cascade framework is strengthened with recent pre-trained models for automatic speech recognition and neural machine translation, achieving 21.3 BLEU on the test set. Three representative end-to-end architectures are evaluated under identical data conditions. The strongest direct model, combining a Wav2Vec 2.0 encoder with an mBART decoder augmented by a length adaptor and adapter modules, reaches 17.97 BLEU, compared with 15.35 BLEU for FAIRSEQ S2T and 16.3 BLEU for ESPnet-ST. Automatic evaluation is complemented by expert manual assessment and targeted linguistic analysis. Results indicate that, under current low-resource conditions, cascade systems provide higher translation accuracy and better morpho-syntactic fidelity, while end-to-end models remain competitive and offer advantages in architectural simplicity and potentially reduced inference latency (due to single-pass processing), although empirical measurements were not conducted in this study. This study establishes a reproducible benchmark for Kazakh–Russian speech translation and highlights practical trade-offs between modeling paradigms in low-resource, morphologically rich settings.
- Open Access
- 6 August 2025
- Published by MDPI AG in Electronics
- Vol. 14 (15) , 3126
Fine-Tuning Methods and Dataset Structures for Multilingual Neural Machine Translation: A Kazakh–English–Russian Case Study in the IT Domain
This study explores fine-tuning methods and dataset structures for multilingual neural machine translation using the No Language Left Behind model, with a case study on Kazakh, English, and Russian. We compare single-stage and two-stage fine-tuning approaches, as well as triplet versus non-triplet dataset configurations, to improve translation quality. A high-quality, 50,000-triplet dataset in information technology domain, manually translated and expert-validated, serves as the in-domain benchmark, complemented by out-of-domain corpora like KazParC. Evaluations using BLEU, chrF, METEOR, and TER metrics reveal that single-stage fine-tuning excels for low-resource pairs (e.g., 0.48 BLEU, 0.77 chrF for Kazakh → Russian), while two-stage fine-tuning benefits high-resource pairs (Russian → English). Triplet datasets improve cross-linguistic consistency compared with non-triplet structures. Our reproducible framework offers practical guidance for adapting neural machine translation to technical domains and low-resource languages.
- 6 August 2025
- Published by MDPI AG in Electronics
- Vol. 14 (15) , 3126
Open AccessFine-Tuning Methods and Dataset Structures for Multilingual Neural Machine Translation: A Kazakh–English–Russian Case Study in the IT Domain
This study explores fine-tuning methods and dataset structures for multilingual neural machine translation using the No Language Left Behind model, with a case study on Kazakh, English, and Russian. We compare single-stage and two-stage fine-tuning approaches, as well as triplet versus non-triplet dataset configurations, to improve translation quality. A high-quality, 50,000-triplet dataset in information technology domain, manually translated and expert-validated, serves as the in-domain benchmark, complemented by out-of-domain corpora like KazParC. Evaluations using BLEU, chrF, METEOR, and TER metrics reveal that single-stage fine-tuning excels for low-resource pairs (e.g., 0.48 BLEU, 0.77 chrF for Kazakh → Russian), while two-stage fine-tuning benefits high-resource pairs (Russian → English). Triplet datasets improve cross-linguistic consistency compared with non-triplet structures. Our reproducible framework offers practical guidance for adapting neural machine translation to technical domains and low-resource languages.
- Open Access
- 29 July 2025
- Published by MDPI AG in Preprints.org
Modeling of Collatz Sequences in Base-6 Using Markov Chains
In this work, we represent and analyze the Collatz sequences in base-6 system and deduce a finite state machine that describes them. This representation of the sequences sheds the light on an interesting insight such as the existence of only four basic types of cycles which constitute any Collatz trajectory. In turn, this allows us to further treat the Collatz process as a Markov chain and, based on the dataset of the first 107 sequences, computationally show that it has a steady state. Moreover, the steady state probabilities theirselves can be directly used to show that the multiplicative factor of a trajectory for any given starting number is less than 1, that is, on average the trajectories tend to converge.
- 29 July 2025
- Published by MDPI AG in Preprints.org
Open AccessModeling of Collatz Sequences in Base-6 Using Markov Chains
In this work, we represent and analyze the Collatz sequences in base-6 system and deduce a finite state machine that describes them. This representation of the sequences sheds the light on an interesting insight such as the existence of only four basic types of cycles which constitute any Collatz trajectory. In turn, this allows us to further treat the Collatz process as a Markov chain and, based on the dataset of the first 107 sequences, computationally show that it has a steady state. Moreover, the steady state probabilities theirselves can be directly used to show that the multiplicative factor of a trajectory for any given starting number is less than 1, that is, on average the trajectories tend to converge.
- 30 March 2025
- Published by L. N. Gumilyov Eurasian National University in Bulletin of L.N. Gumilyov Eurasian National University. Mathematics, computer science, mechanics series
- Vol. 1 (1) , 6-16
Идентификация языка устной речи с использованием модели Wav2Vec2 для казахского языка
В данном исследовании представлена разработка и тонкая настройка модели идентификации языка устной речи с использованием варианта XLSR (Cross-Lingual Speech Recognition) Wav2Vec2. Обученная на богатом и разнообразном наборе данных, охватывающем шесть языков, с особым акцентом на языках с низкими ресурсами, таких как казахский, модель демонстрирует замечательные возможности в многоязычном распознавании речи. Благодаря обширной оценке тонко настроенная модель не только превосходит существующие эталонные показатели, но и превосходит другие современные модели, включая варианты Whisper. Достигнув впечатляющего результата F1 92,9% и точности 93%, модель демонстрирует свою производительность в реальных многоязычных и малоресурсных сценариях. Эта работа вносит значительный вклад в развитие технологий распознавания речи, предоставляя надежное решение для идентификации языка в различных языковых средах, особенно в недостаточно представленных языковых условиях. Его успех подчеркивает потенциал моделей на основе Wav2Vec2 в улучшении систем обработки речи в многоязычных контекстах с низкими ресурсами. Итоги данного анализа могут способствовать разработке надежных и эффективных систем автоматического распознавания речи, оптимизированных для казахского языка. Такие технологии найдут применение в различных областях, включая преобразование речи в текст, работу голосовых ассистентов и инструменты голосовой коммуникации.
- 30 March 2025
- Published by L. N. Gumilyov Eurasian National University in Bulletin of L.N. Gumilyov Eurasian National University. Mathematics, computer science, mechanics series
- Vol. 1 (1) , 6-16
Идентификация языка устной речи с использованием модели Wav2Vec2 для казахского языка
В данном исследовании представлена разработка и тонкая настройка модели идентификации языка устной речи с использованием варианта XLSR (Cross-Lingual Speech Recognition) Wav2Vec2. Обученная на богатом и разнообразном наборе данных, охватывающем шесть языков, с особым акцентом на языках с низкими ресурсами, таких как казахский, модель демонстрирует замечательные возможности в многоязычном распознавании речи. Благодаря обширной оценке тонко настроенная модель не только превосходит существующие эталонные показатели, но и превосходит другие современные модели, включая варианты Whisper. Достигнув впечатляющего результата F1 92,9% и точности 93%, модель демонстрирует свою производительность в реальных многоязычных и малоресурсных сценариях. Эта работа вносит значительный вклад в развитие технологий распознавания речи, предоставляя надежное решение для идентификации языка в различных языковых средах, особенно в недостаточно представленных языковых условиях. Его успех подчеркивает потенциал моделей на основе Wav2Vec2 в улучшении систем обработки речи в многоязычных контекстах с низкими ресурсами. Итоги данного анализа могут способствовать разработке надежных и эффективных систем автоматического распознавания речи, оптимизированных для казахского языка. Такие технологии найдут применение в различных областях, включая преобразование речи в текст, работу голосовых ассистентов и инструменты голосовой коммуникации.
- Open Access
Low-resource Machine Translation for Code-switched Kazakh-Russian Language Pair
Machine translation for low resource language pairs is a challenging task. This task could become extremely difficult once a speaker uses code switching. We propose a method to build a machine translation model for code-switched Kazakh-Russian language pair with no labeled data. Our method is basing on generation of synthetic data. Additionally, we present the first codeswitching Kazakh-Russian parallel corpus and the evaluation results, which include a model achieving 16.48 BLEU almost reaching an existing commercial system and beating it by human evaluation.
Open AccessLow-resource Machine Translation for Code-switched Kazakh-Russian Language Pair
Machine translation for low resource language pairs is a challenging task. This task could become extremely difficult once a speaker uses code switching. We propose a method to build a machine translation model for code-switched Kazakh-Russian language pair with no labeled data. Our method is basing on generation of synthetic data. Additionally, we present the first codeswitching Kazakh-Russian parallel corpus and the evaluation results, which include a model achieving 16.48 BLEU almost reaching an existing commercial system and beating it by human evaluation.
- Open Access
- 25 December 2024
- Published by Private Company Technology Center in Eastern-European Journal of Enterprise Technologies
- Vol. 6 (2 (132)) , 14-22
Enhancing skeleton-based action recognition with hybrid real and gan-generated datasets
This research addresses the critical challenge of recognizing mutual actions involving multiple individuals, an important task for applications such as video surveillance, human-computer interaction, autonomous systems, and behavioral analysis. Identifying these actions from 3D skeleton motion sequences poses significant challenges due to the necessity of accurately capturing intricate spatial and temporal patterns in diverse, dynamic, and often unpredictable environments. To tackle this, a robust neural network framework was developed that combines Convolutional Neural Networks (CNNs) for efficient spatial feature extraction with Long Short-Term Memory (LSTM) networks to model temporal dependencies over extended sequences. A distinguishing feature of this study is the creation of a hybrid dataset that which combines real-world skeleton motion data with synthetically generated samples, produced using Generative Adversarial Networks (GANs). This dataset enriches variability, enhances generalization, and mitigates data scarcity challenges. Experimental findings across three different network architectures demonstrate that our method significantly enhances recognition accuracy, mainly due to the integration of CNNs and LSTMs alongside the broadened dataset. Our approach successfully identifies complex interactions and ensures consistent performance across different perspectives and environmental conditions. The improved reliability in recognition indicates that this framework can be effectively utilized in practical applications such as security systems, crowd monitoring, and other areas where precise detection of mutual actions is critical, particularly in real-time and dynamic environments
- 25 December 2024
- Published by Private Company Technology Center in Eastern-European Journal of Enterprise Technologies
- Vol. 6 (2 (132)) , 14-22
Open AccessEnhancing skeleton-based action recognition with hybrid real and gan-generated datasets
This research addresses the critical challenge of recognizing mutual actions involving multiple individuals, an important task for applications such as video surveillance, human-computer interaction, autonomous systems, and behavioral analysis. Identifying these actions from 3D skeleton motion sequences poses significant challenges due to the necessity of accurately capturing intricate spatial and temporal patterns in diverse, dynamic, and often unpredictable environments. To tackle this, a robust neural network framework was developed that combines Convolutional Neural Networks (CNNs) for efficient spatial feature extraction with Long Short-Term Memory (LSTM) networks to model temporal dependencies over extended sequences. A distinguishing feature of this study is the creation of a hybrid dataset that which combines real-world skeleton motion data with synthetically generated samples, produced using Generative Adversarial Networks (GANs). This dataset enriches variability, enhances generalization, and mitigates data scarcity challenges. Experimental findings across three different network architectures demonstrate that our method significantly enhances recognition accuracy, mainly due to the integration of CNNs and LSTMs alongside the broadened dataset. Our approach successfully identifies complex interactions and ensures consistent performance across different perspectives and environmental conditions. The improved reliability in recognition indicates that this framework can be effectively utilized in practical applications such as security systems, crowd monitoring, and other areas where precise detection of mutual actions is critical, particularly in real-time and dynamic environments
- Open Access
- 27 November 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 12, 177844-177855
Semantically Expanded Spoken Term Detection
Spoken term detection (STD) is effectively implemented using fundamental techniques such as automatic speech recognition (ASR) and information retrieval. Through these methods, queried keywords can be identified in the decoded texts and indexed lattices produced by the ASR system. However, this approach relies heavily on the performance of the ASR; it may not produce the desired results when dealing with out-of-vocabulary (OOV) words that are not included in the ASR’s lexicon. To address this limitation, we analyzed the semantic query expansion technique through extensive and reproducible experiments to assess its impact on the search quality for OOV words. We propose an approach to enhance existing spoken content retrieval methods by searching semantically expanded query sets and leveraging the advanced features of search engines. Our experiments, conducted on the Wall Street Journal (WSJ) datasets and top Google frequent queries, demonstrate that the proposed approach significantly improves retrieval accuracy over the traditional word-based STD method for in-vocabulary (IV) terms. Specifically, the Actual Term Weighted Value (ATWV) score improved from 0 to 0.5776 for the trigram query category. Additionally, our approach outperforms the proxy-based method for OOV words. While the proxy-based technique fails to retrieve results for both bigrams and trigrams, the semantic-based approach achieves ATWV scores of 0.7143 and 0.8846 for bigrams and trigrams, respectively. Furthermore, substantial gains are observed when combining semantic-based query expansion with a full-text search engine, improving the performance of the word-based STD system by approximately 3 to 4 times on the bigram and trigram query categories.
- 27 November 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 12, 177844-177855
Open AccessSemantically Expanded Spoken Term Detection
Spoken term detection (STD) is effectively implemented using fundamental techniques such as automatic speech recognition (ASR) and information retrieval. Through these methods, queried keywords can be identified in the decoded texts and indexed lattices produced by the ASR system. However, this approach relies heavily on the performance of the ASR; it may not produce the desired results when dealing with out-of-vocabulary (OOV) words that are not included in the ASR’s lexicon. To address this limitation, we analyzed the semantic query expansion technique through extensive and reproducible experiments to assess its impact on the search quality for OOV words. We propose an approach to enhance existing spoken content retrieval methods by searching semantically expanded query sets and leveraging the advanced features of search engines. Our experiments, conducted on the Wall Street Journal (WSJ) datasets and top Google frequent queries, demonstrate that the proposed approach significantly improves retrieval accuracy over the traditional word-based STD method for in-vocabulary (IV) terms. Specifically, the Actual Term Weighted Value (ATWV) score improved from 0 to 0.5776 for the trigram query category. Additionally, our approach outperforms the proxy-based method for OOV words. While the proxy-based technique fails to retrieve results for both bigrams and trigrams, the semantic-based approach achieves ATWV scores of 0.7143 and 0.8846 for bigrams and trigrams, respectively. Furthermore, substantial gains are observed when combining semantic-based query expansion with a full-text search engine, improving the performance of the word-based STD system by approximately 3 to 4 times on the bigram and trigram query categories.
- 17 September 2024
- Published by Springer Nature
- p. 368-380
Use of Riemannian Distance Metric to Verify Topological Similarity of Acoustic and Text Domains
- 17 September 2024
- Published by Springer Nature
- p. 368-380
Use of Riemannian Distance Metric to Verify Topological Similarity of Acoustic and Text Domains