Chinese spelling correction (CSC) faces unique challenges from phonetic and visual character confusions. Existing methods suffer from three critical limitations: (1) static confusion sets failing to capture evolving error patterns, (2) inadequate multi-scale contextual modeling, and (3) insufficient exploitation of Chinese phonetic features. We propose GLAD-CSC, a novel framework integrating Global-Local Attention with Dynamic confusion management. Our key innovations include: First, a global-local hybrid attention mechanism that adaptively balances semantic understanding with local context modeling through error-probability-driven window adjustment. Second, a multi-modal fusion strategy integrating textual, phonetic (Pinyin), and confusion features via dynamic weighting, with enhanced phonetic representation (40 % allocation) for handling phonetic errors. Third, a dynamic confusion management system that automatically discovers and validates character confusion patterns during training. Extensive experiments on SIGHAN benchmarks demonstrate state-of-the-art performance: 91.9% character-level correction F1-score on SIGHAN 2015 ($+3.8 \%$improvement) and 90.6 % sentence-level detection F1-score ($+8.5 \%$improvement). Our framework excels at phonetic confusion errors, achieving 96.8 % precision and 92.1% recall. Ablation studies confirm significant contributions: enhanced phonetic features ($+4.7 \%$F1), multi-scale attention ($+4.0 \%$F1), and dynamic confusion management ($+3.3 \% \text{F1}$).

Model pruning, as an effective method for compressing large language models (LLMs), has recently attracted considerable attention in the field of natural language processing. However, existing LLM pruning methods have two main drawbacks: (1) Iterative pruning for LLMs with over a billion parameters requires retraining, which leads to significant pruning costs. (2) LLMs Pruning is formalized as a weight reconstruction problem that necessitates second-order information, incurring expensive computations. To address these issues, we propose a novel pruning method named Eplra: efficient one-shot pruning of large language models with low-rank approximation, which efficiently identifies sparse networks in LLMs. Specifically, we design a novel pruning metric based on input activations for the rapid one-shot compression of LLMs. We first incorporate input activations into the calculation of weight importance to promote precise pruning of low-priority weights. Then, we perform local weight comparisons across each output of linear layers to induce uniform sparsity. Next, we expand Eplra into semi-structured pruning patterns to accommodate various acceleration scenarios. Finally, we employ low-rank parametrized update matrices to fine-tune the pruned model, facilitating a swift recovery of model performance. Experimental results on various language benchmark datasets demonstrate that Eplra outperforms the state-of-the-art methods.

Aspect-based sentiment analysis (ABSA) is a fine-grained sentiment analysis task designed to identify the polarity of a target aspect. Some works introduce various attention mechanisms to fully mine the relevant context words of different aspects, and use the traditional cross-entropy loss to fine-tune the models for the ABSA task. However, the attention mechanism paying partial attention to aspect-unrelated words inevitably introduces irrelevant noise. Moreover, the cross-entropy loss lacks discriminative learning of features, which makes it difficult to exploit the implicit information of intra-class compactness and inter-class separability. To overcome these challenges, we propose an Aspect Feature Distillation and Enhancement Network (AFDEN) for the ABSA task. We first propose a dual-feature extraction module to extract aspect-related and aspect-unrelated features through the attention mechanisms and graph convolutional networks. Then, to eliminate the interference of aspect-unrelated words, we design a novel aspect-feature distillation module containing a gradient reverse layer that learns aspect-unrelated contextual features through adversarial training, and an aspect-specific orthogonal projection layer to further project aspect-related features into the orthogonal space of aspect-unrelated features. Finally, we propose an aspect-feature enhancement module that leverages supervised contrastive learning to capture the implicit information between the same sentiment labels and between different sentiment labels. Experimental results on three public datasets demonstrate that our AFDEN model achieves state-of-the-art performance and verify the effectiveness and robustness of our model.

Summary The feature layers of different layers in the single shot multibox detector (SSD) are independently used as the input of the classification network, so it is easy to detect the same object. This article proposes an improved SSD model based on deep feature fusion. In the SSD algorithm, the deep feature fusion between the target detection layer and its adjacent feature layer is used, including convolution kernels and pooling kernels of different sizes, down‐sampling of low‐level features and up‐sampling of deconvolution of high‐level features. The network is improved by combining the target frame recommendation strategy in the SSD algorithm and the frame regression algorithm. The experimental results show that the improved SSD algorithm improves the detection accuracy and detection rate of the target, and the effect is more obvious for the relatively small‐scale target.

Abstract In sEMG (surface electromyography) pattern recognition, most of the research focuses on the static pattern recognition of different limbs, ignoring the importance of changing load intensity, and joint angle movement information. Traditional static qualitative pattern recognition cannot adjust the motion amplitude and load intensity, so it is of great significance to study the continuous prediction of wrist angle under different load intensities. Based on the correlation between the surface EMG signal and the joint angle signal, the article is based on the neural network to identify and predict the wrist angle under different loads continuously quantitatively. The sEMG signal in this article was collected with the approval and review of the Ethics Committee and the people's informed consent. Since qualitative pattern recognition cannot adjust the wrist movement range and the different load training intensity, the article establishes an angle prediction model based on a genetic algorithm to optimize the extreme learning machine (ELM). In addition, the article analyzes the influence of different loads on the continuous prediction accuracy of the wrist angle, realizes the continuous quantitative angle of the precise wrist prediction. Experimental analysis shows that the wrist joint angle predicted by the ELM optimized based on genetic algorithm is close to the actual angle, and the average error is about 5.96 degrees.

Summary Target localization in unknown environment is one of the development directions of mobile robots. Simultaneous localization and mapping (SLAM) can be used to build maps in unknown environments, but it has the problem of poor readability and interactivity. In this article, target detection and SLAM are combined to search and locate the target by using rich RGBD images information. The determined position in the global map is conducive to the follow‐up operation of the target by mobile robots. By establishing a local dense point cloud map of the target object, the current state of the target object is directly displayed, the readability of the map is improved, and the disadvantages of difficult understanding of the global sparse map and slow construction of the global dense map are avoided. A target localization algorithm under the framework of yolov4 is designed to apply in the process of SLAM global mapping. Our works are helpful for obtaining positions of objects in three‐dimensional space. The experimental results show that the time‐consuming of this method in dense mapping is reduced by 50%–70%, and the number of point clouds is also reduced by 60%–70%.

Abstract The manipulator workspace is an essential element in the field of manipulator research and is of great significance for manipulator motion planning. However, little research has been conducted on dividing the manipulator workspace into working subspaces. No precise division method has been proposed; the inverse kinematics of multiple solutions in manipulator trajectory planning may also cause abrupt joint changes, thus affecting the planned trajectory. The article proposes a working subspace division method for all ball‐wrist 6DOF(degree‐of‐freedom) manipulators that satisfy the Piper criterion to address the above problems. The kinematic model of the manipulator is established, and the Jacobi matrix of the manipulator is obtained. The space of joints of the manipulator is divided into unique domains containing only single inverse kinematic solutions by means of singular trajectory lines when the determinant of the Jacobi matrix is zero; The solution from the joint space to the workspace is achieved by a nonlinear mapping, which completes the partitioning of the work subspace, and each work subspace contains only unique inverse kinematic solutions. When trajectory planning is carried out from the independent area of a single workspace to the overlapping area of multiple workspaces, selecting the inverse kinematic solution in a single working subspace can effectively avoid abrupt changes in the joints of the manipulator and trajectory misalignment caused by numerous inverse solution selection problems and make the planned trajectory smooth and consistent with the operational requirements of each scene.

Recent dominant approaches for scene text recognition are mainly based on convolutional neural network (CNN) and recurrent neural network (RNN), where the CNN processes images and the RNN generates character sequences. Different from these methods, we propose an attention-based architecture1 which is completely based on CNNs. The distinctive characteristics of our method include: (1) the method follows encoder-decoder architecture, in which the encoder is a two-dimensional residual CNN and the decoder is a deep one-dimensional CNN. (2) An attention module that captures visual cues, and a language module that models linguistic rules are designed equally in the decoder. Therefore the attention and language can be viewed as an ensemble to boost predictions jointly. (3) Instead of using a single loss from language aspect, multiple losses from attention and language are accumulated for training the networks in an end-to-end way. We conduct experiments on standard datasets for scene text recognition, including Street View Text, IIIT5K and ICDAR datasets. The experimental results show our CNN-based method has achieved state-of-the-art performance on several benchmark datasets, even without the use of RNN.