ABSTRACT The presented paper introduces a Brahui‐to‐English NMT system developed for a severely low‐resource scenario. The framework adopted for creating the NMT system is based on the Marian MT Transformer model and augmented with back‐translation. In order to develop the model, a Brahui‐to‐English parallel corpus was initially cleaned up and divided into training, validation, and test sets. After this, the pipeline generated baseline translation models in both directions. Subsequently, the English‐to‐Brahui model was utilized to generate synthetic Brahui sentences on the basis of monolingual English data while fine‐tuning the original forward translation model. In total, the experiment was executed within Google Colab TP4, and the computation was run on a CUDA‐compatible computer for up to 16 h. The experimental settings include a max length of 96 for the sequences, 8 for the batch size, a gradient accumulation of 2, the learning rate of 3e−5, and three training epochs for both baselines and augmented models. During the evaluation of the results, such measures as BLEU, ChrF, and METEOR were applied, and the obtained scores indicate a clear evolutionary superiority. The framework progressed from a traditional Bi‐LSTM baseline 45.89 BLEU, 50.68 ChrF, 56.93 METEOR, to a standard Transformer baseline 48.43 BLEU, 57.29 ChrF, 62.08 METEOR, and culminated in the proposed BT‐augmented model, which achieved peak performance 53.87 BLEU, 61.85 ChrF, 65.91 METEOR. While standard NMT models frequently struggle to surpass BLEU scores of 15–40 for severely low‐resource and morphologically complex languages, the proposed framework achieves a highly competitive BLEU score of 53.87, setting a robust new benchmark for the Natural language processing field. The results presented above were obtained by conducting single‐point estimation runs, while multiseed training and statistical significance testing have been recognized as crucial tasks for future work.
ABSTRACT Hierarchical classification (HC) is widely used in domains with pre‐defined label taxonomies, but time series classification (TSC) conventionally treats classes as flat and nonhierarchical. This paper investigates whether an automatically induced class hierarchy can improve TSC performance even when no such structure is given a priori. We propose a classifier‐informed hierarchical divisive clustering framework in which candidate splits are evaluated directly by the predictive performance of a binary classifier trained on the resulting class groups, and introduce three stochastic splitting functions (SSFs), pick‐one‐then‐regroup ( potr ), split‐randomly‐then‐regroup ( srtr ), and leave‐salient‐one‐out ( lsoo ), to explore the space of candidate hierarchies. The resulting trees are evaluated under an extended local‐classifier‐per‐parent‐node (LCPN ) scheme using nested cross‐validation (NCV), and compared against flat classification (FC) and four established hierarchy‐generation (HG) baselines from the literature, across 46 UCR archive datasets, three modern base classifiers (MiniRocket, QUANT, and Crossfire), and four complementary metrics (accuracy, macro‐F1, AUC, and negative log‐likelihood). Results show that the benefit of hierarchy imposition is strongly classifier‐ and metric‐dependent: srtr ‐based HC significantly outperforms FC and all established baselines for MiniRocket on accuracy and F1, and is statistically competitive with leading state‐of‐the‐art (SOTA) TSC methods, but offers no consistent advantage over FC for QUANT or Crossfire. Across all three classifiers, hierarchy imposition does not improve AUC over FC, indicating that accuracy gains do not necessarily translate into better discrimination confidence. Dataset meta‐feature analysis further shows that improvement correlates with class cardinality, series length, and baseline accuracy in an interpretable way for MiniRocket, but weakly or inconsistently for the other classifiers. These findings position automated hierarchy induction as a targeted, classifier‐specific tool rather than a general‐purpose enhancement for TSC, and establish the first systematic benchmark for this problem, along with practical guidance on when hierarchy imposition is worth its computational cost.
ABSTRACT This paper proposes the dynamic adaptive regularization soft actor‐critic (DAR‐SAC) framework. The framework is based on the SAC algorithm, which is a reinforcement learning framework incorporating dynamic environment detection and adaptive optimization, to address the fixed exploration noise stiffness of the traditional SAC algorithm in dynamic environments. For the problems of fixed exploration noise rigidity, strategy oscillation aggravation and Q‐value over‐estimation in traditional SAC algorithms in dynamic environments, this paper firstly introduces a dynamic environment detection module to enable the robot to distinguish dynamic environments, and then combines the hybrid regularization mechanism and adaptive gradient trimming to accelerate convergence while suppressing abrupt changes in the strategy. In the constructed 3D point cloud simulation environment, the DAR‐SAC algorithm achieves significantly higher cumulative rewards and lower collision rates compared with SAC, SARC, DDPG, and TD3 algorithms. To further evaluate the robustness of the proposed framework against non‐stationary environments, we conduct extensive comparisons with state‐of‐the‐art variants, including SAC with adaptive entropy scheduling (SAC + AES), distributional SAC with three refinements (DSAC‐T), and SAC with delayed actor updates and replay mixing (SAC + DA + RM). The experimental results demonstrate that the proposed method exhibits superior path‐planning performance in complex static and dynamic environments.
Gait, the distinctive and natural pattern of an individual's walking behavior, has emerged as an effective biometric for user authentication. Its ability to operate without active user involvement, along with its resistance to replication, makes it particularly well-suited for continuous smartphone authentication. However, real-world deployment remains challenging due to diverse user activities and varying environmental conditions. Therefore, this study proposes a data synthesis approach to enhance the robustness and generalization of gait-based authentication systems in real-world environment. We introduce a Data Synthetic Model that combines a Variational Autoencoder (VAE) with a Bidirectional Long Short-Term Memory (Bi-LSTM) network. This architecture generates synthetic gait sequences that are temporally consistent, reflecting real-world variability while maintaining essential gait dynamics. Evaluations on the ETRI Lifelog dataset demonstrate that the proposed method enhances performance of the smartphone authentication system, improving accuracy from 0.91 to 0.951 and reducing the Equal Error Rate (EER) from 0.101 to 0.040. These results highlight the potential of data synthesis for advancing real-world gait authentication. While further improvements are needed for deployment in high-security applications, this study provides a strong foundation for developing robust and scalable gait-based biometric systems.
Events with their triggers and entity mentions are significant in multi-document summarization, which provide assistance to the recognition of the vital content of plenty of documents and guarantee the logic of the text. Many existing methods pay attention to leveraging fine-grained and noun-centered units like entities to interact with different documents rather than verb-centered units, where the state changes and salient semantic associations in documents are neglected, causing a lack of logic in summarization. To address these issues, we propose an innovative abstractive multi-document summarization model focused on events comprised of dynamic triggers and static entities, referred to as Event-focused Abstractive Multi-document Summarization (EVESUM), to form high-quality summaries by explicitly introducing event information from documents. Specifically, we first apply an elaborately designed event-focused graph attention network to fuse information from events and documents from different encoders, which particularly considers the function of edges in the graph. Additionally, we construct a hierarchical attention mechanism in the decoder to incorporate crucial event messages into summaries, making the generated summaries more succinct, informative, and logical. Experimental results illustrate that compared with state-of-the-art baselines, our model achieves remarkable improvements on the WikiSum and MultiNews datasets.
