Kazakh Sign Language (KSL) is a crucial communication tool for individuals with hearing and speech impairments. Deep learning, particularly Transformer models, offers a promising approach to improving accessibility in education and communication. This study analyzes the syntactic structure of KSL, identifying its unique grammatical features and deviations from spoken Kazakh. A custom parser was developed to convert Kazakh text into KSL glosses, enabling the creation of a large-scale parallel corpus. Using this resource, a Transformer-based machine translation model was trained, achieving high translation accuracy and demonstrating the feasibility of this approach for enhancing communication accessibility. The research highlights key challenges in sign language processing, such as the limited availability of annotated data. Future work directions include the integration of video data and the adoption of more comprehensive evaluation metrics. This paper presents a methodology for constructing a parallel corpus through gloss annotations, contributing to advancements in sign language translation technology.

The rapid growth of scientific literature necessitates effective information extraction. However, existing methods face significant challenges, particularly when applied to multi-domain documents and low-resource languages. For Kazakh and Russian, there is a notable lack of annotated corpora and dedicated tools for scientific information extraction. To address this gap, we introduce SciMDIX (Scientific Multi-Domain Information extraction), a novel multi-domain dataset of scientific documents in Russian and Kazakh, annotated with entities and relations. Our study includes a comprehensive evaluation of entity recognition performance, comparing state-of-the-art models, such as BERT, LLaMA, GLiNER, and spaCy across four diverse domains (IT, Linguistics, Medicine, and Psychology) in both languages. The findings highlight the promise of spaCy and GLiNER for practical deployment in under-resourced language settings. Furthermore, we propose a new zero-shot relation extraction model that leverages a multimodal representation by integrating sentence context, entity mentions, and textual definitions of relation classes. Our model can predict semantic relations between entities in new documents, even for a language encountered during training. This capability is especially valuable for low-resource language scenarios.

Multimodal human–AI systems generally consider facial expressions and body motions as separate input streams, leading to disjointed interpretations and diminished emotional coherence. To overcome this issue, we offer the Engagement-Safe Expressive Alignment (ESEA) paradigm and the Unified Visual Synchrony (UVS) framework as its computational implementation. UVS models the coherence between facial expressions and gestures, offering an interpretable visual synchrony signal that can function as adaptive feedback in human–AI interactions. The framework’s key component is the Consistency Index for Affective Synchrony ($CIAS$), which correlates brief visual segments with scalar synchrony scores through a common latent representation. Facial and gestural signals are processed by modality-specific projection networks into a unified latent space, and $CIAS$ is derived from the similarity and short-term temporal consistency of these latent trajectories. The synchrony index is regarded as an estimation of affective visual coherence within the ESEA paradigm. We formalize the UVS/$CIAS$ framework and conduct a comparative experimental evaluation utilizing matched and mismatched face–gesture segments derived from rendered dialog footage. Utilizing ROC analysis, score distribution comparisons, temporal visualizations, and negative control tests, we illustrate that $CIAS$ effectively captures structured face–gesture alignment that surpasses similarity-based baselines, while also delivering a persistent, time-resolved synchronization signal. These findings establish $CIAS$ as a principled and interpretable feedback signal for future affect-aware, engagement-focused multimodal agents.

Learning meaningful representations of scientific documents is essential for information retrieval, knowledge discovery, and recommendation systems. Traditional methods such as TF-IDF rely on lexical matching and fail to capture deeper semantic relationships, while transformer-based approaches typically depend on limited supervision signals. In this work, we propose a Triple-Source automatic supervision framework for learning document embeddings from scientific corpora. The model integrates three types of supervision–title–abstract pairs, same-category document pairs, and document-level semantic relationships—within a unified contrastive learning framework based on a multilingual XLM-RoBERTa encoder. Unlike prior approaches that rely on citation graphs or manual annotations, our method enables citation-free and annotation-free representation learning using only lightweight metadata. Experiments on a publicly available arXiv dataset consisting of 98,649 documents demonstrate improved semantic retrieval performance, achieving Recall@1 = 0.6181 for same-category retrieval and outperforming both TF-IDF and single-source transformer baselines. The learned embeddings also exhibit improved clustering of scientific domains, indicating more structured semantic representations.

This study presents a comprehensive review of more than 100 research papers on sign language recognition (SLR) published between 2020 and 2026. The analysis focuses on deep learning approaches applied to video-based SLR, including spatiotemporal feature extraction, temporal modeling, attention mechanisms, motion-based representations, hybrid frameworks, transfer learning methods and other methods. Particular attention is given to how these methods model spatiotemporal dynamics and capture subtle gesture characteristics in sign language communication. The review highlights several recent developments, such as the introduction of specialized datasets, the emergence of real-time recognition systems, and the integration of multimodal fusion strategies. At the same time, persistent challenges remain, including data scarcity in low-resource sign languages, limited linguistic standardization of datasets, and insufficient model interpretability. The findings underline the importance of developing scalable and generalizable models capable of handling diverse datasets and user variability. The distinct contributions of this review are fourfold: (1) a comprehensive synthesis of over 100 studies published between 2020 and 2026, covering the full spectrum of deep learning architectures for video-based SLR; (2) a structured six-category taxonomy enabling systematic cross-architectural comparison; (3) a comprehensive focus on low-resource sign languages, which remain underrepresented in the existing literature; and (4) a critical analysis of the current benchmark landscape for low-resource sign languages, identifying key gaps and outlining strategic directions for future dataset development. These contributions are intended to guide further research toward more robust, inclusive, and universally applicable SLR systems.

We introduce the Probabilistic Multimodal Consistency Index (PMCI), a prototype measure of probabilistic semantic agreement used to quantify how much two modalities share in terms of mutually compatible affective evidence. PMCI does not model feature fusion or emotion classification directly; instead, each modality is mapped to a different probability distribution over learnable latent affective-agreement anchors, and the agreement of those distributions is quantified via the Jensen–Shannon divergence. Consequently, we propose PMCI not as a substitute for a discriminative model of pair matching, but rather as an auxiliary diagnostic index of cross-modal affective agreement. Experiments were conducted on pose-based facial and body keypoint sequences obtained from the RAVDESS and MELD datasets. In the updated RAVDESS setup, we extended the cache of preprocessed keypoint files to include all 24 actors, and we sampled four distinct pose-derived 12-frame crops per source video. This resulted in a total of 11,520 pose-derived windows, with the actor label determining the train/validation/test splits for the RAVDESS dataset. MELD contained 402 filtered pose-derived windows and served as an auxiliary in-the-wild validation setting with additional noise. When applying the updated RAVDESS standard pair-matching, DirectCosine_K0 achieved ROC–AUC = 0.943 and PR–AUC = 0.917, demonstrating that it is indeed the best exact pair-matching baseline and that PMCI-based configurations are not as accurate as DirectCosine_K0 when performing common exact face–body pair discrimination. Under this protocol, PMCI_K8, PMCI_K16, and PMCI_K32 produced ROC–AUC scores of 0.851, 0.805, and 0.887, respectively. Running a one-window-per-source-video experiment with 100 repetitions yielded similar results in the following order: DirectCosine_K0 = 0.942 plus-minus 0.009, PMCI_K16 = 0.807 plus-minus 0.016, and PMCI_K32 = 0.893 plus-minus 0.012, which shows that the results obtained were not the result of only repeated temporal crops. In the DirectCosine-mined hard negative setting, PMCI produced modest diagnostic separation above chance, where Direct+PMCI_K32 yielded ROC–AUC = 0.582. In a suite of shuffle control, permutation control, model initialization control, temperature control, anchor usage control, control latency, and pose perturbation control experiments, the diagnostic stability of PMCI was tested, confirming that PMCI is sensitive to both pose-estimation accuracy and domain shift. Finally, PMCI must be interpreted as a probabilistic diagnostic index of cross-modal affective agreement that is based on pose, not as a general-purpose emotion-recognition system.