With the rapid development of deep learning, multi-view 3D model classification has become a hot research topic. However, multiple views often contain redundant information, and complementary features are frequently overlooked, making the preservation of discriminative information a critical challenge. To address these challenges, a method of 3D model classification based on representative view selection and dynamic hard-view mining is proposed. Firstly, a three-stage strategy is adopted in data preprocessing. 12 representative views are selected from 80 original ones. Training efficiency is improved significantly. Secondly, improved ConvNeXtV2 is adopted to extract multi-scale features. Features from the penultimate and final stages are concatenated and subsequently refined by a lightweight SE-style channel recalibration module to obtain robust merged representations. Thirdly, EnhancedProtoLoss is used to generate per-view logits. It integrates SupCon, orthogonal projection loss and smooth cross-entropy. CenterLoss is introduced to reinforce intraclass compactness. Fourthly, dynamic hard-view mining is performed periodically based on per-view loss. Robustness against marginal views is enhanced. Fifthly, final prediction is computed in aggregation phase. A reliability-aware soft voting strategy is adopted, which computes the weighted sum of per-view probabilities by leveraging view reliability statistics from the validation set. Experimental results show that accuracy of ConvNeXt-Ag achieves 90.25% on ModelNet40 dataset. Ablation studies confirm that the proposed components provide substantial gains. Superior performance in accuracy and robustness is demonstrated.

Multi-view 3D model classification has achieved strong performance by leveraging convolutional neural networks on projected views. However, existing methods still suffer from view redundancy and insufficient modeling of inter-view relationships, which limit classification performance.To address these issues, this paper proposes a SEResNet-GCN framework for multi-view 3D model classification. The framework first employs a channel-attention enhanced ResNet18 to extract discriminative view-level features. Then, a KNN-based graph convolutional network is designed to model inter-view relationships and identify representative views. Finally, a probability-based voting strategy is adopted to aggregate complementary information from multiple selected views for final classification.Experiments on the ModelNet40 benchmark dataset demonstrate that the proposed method achieves a Top-1 classification accuracy of 91.5%, outperforming several representative baseline methods.The results indicate that integrating channel-aware feature enhancement with graph-based view relationship modeling can effectively improve representation quality while reducing view redundancy in multi-view 3D classification tasks.

Biomedical word sense disambiguation (WSD) faces increasing challenges due to the rapid growth of biomedical text data and the prevalence of ambiguous terms. To address these challenges, this paper proposes a tri-branch fused encoder method that integrates Electra-base-discriminator, mDeBERTa-v3-base, and Flan-T5-base to capture both local and long-range contextual dependencies. First, each encoder independently extracts contextual representations of the target term and its surrounding text, which are then fused to obtain complementary semantic features. Second, a contrastive learning module is designed using random cropped views that preserve the biomedical keyword and NT-Xent loss to enhance instance-level discrimination. Third, a χ2-guided attention mechanism highlights statistically representative tokens and strengthens the robustness of feature representations. Fourth, convergence and stability are further improved through a combination of focal loss and multi-label margin loss. Finally, a learnable voting scheme integrates the predictions of multiple classifiers to preserve interpretability, and the fused representation is input to a softmax classifier for sense prediction. The disambiguation accuracy is 95.27 on MSH dataset and 93.19 on NLM dataset. Experimental results demonstrate that the proposed method consistently surpasses existing approaches in both accuracy and F1 score on MSH-WSD and NLM-WSD datasets, validating the effectiveness of multi-encoder fusion, χ2-guided attention, composite loss function and contrastive learning.

Biomedical word sense disambiguation (WSD) is a challenging task due to the widespread semantic ambiguity of domain-specific terms. To improve semantic representation and long-range contextual modeling, this paper proposes a role-aware multi-source knowledge fusion framework for biomedical WSD, in which BERT, Word2Vec, and Reformer are assigned different roles for contextual semantic encoding, auxiliary linguistic representation, and long-range dependency modeling. To further enhance the discrimination of ambiguous samples, a hard-sample mining strategy based on prediction confidence and loss values is introduced during training. Representative keywords are extracted using YAKE, and a hybrid loss function is adopted to improve feature quality and training stability. Experiments conducted on the MSH biomedical dataset demonstrate that the proposed method outperforms baseline models in terms of Accuracy. The results verify the effectiveness of the proposed feature fusion framework and hard-sample mining strategy for biomedical word sense disambiguation.

With the rapid development of three-dimensional (3D) sensing technologies, multimodal 3D model classification has achieved significant progress. However, most existing methods are developed under closed and balanced assumptions, which limits their applicability to open long-tailed scenarios with scarce tail classes, ambiguous hard samples, and continuously emerging categories. In this work, we propose sample-enhanced category-space learning (SE-CSL) for open long-tailed multimodal 3D model classification. The proposed method first uses dual-branch modality encoders to extract point-cloud structural representations and multi-view semantic representations. Mamba is then introduced to model global dependencies across heterogeneous modalities and generate a unified global category representation. To improve the robustness of category representation, we design a category-space learning strategy that jointly integrates long-tailed learning, few-shot representation stabilization, and hard-sample enhancement. A long-tail balanced loss, a few-shot stabilization loss, and a hard-sample boundary loss are further developed to optimize intra-class compactness, inter-class separability, and boundary discrimination. To handle continually emerging classes, we introduce an incremental category-space expansion mechanism that distinguishes new classes, preserves old-class information, and supports unified classification of old and new categories. Extensive experiments on ModelNet40 and ShapeNet55 demonstrate the effectiveness and robustness of SE-CSL.

Word sense disambiguation is a fundamental problem in natural language understanding. However, many existing approaches rely on fixed-size context windows, which restrict their capacity to model multi-scale semantic cues and limit the effective exploitation of diverse contextual evidence. In this work, we propose a framework based on graph neural networks for Chinese word sense disambiguation with multi-window feature fusion. Specifically, we construct context windows of varying sizes centered on the ambiguous word to capture multi-granularity linguistic information, including lexical tokens, part-of-speech tags, semantic categories, and sentence-level rep resentations. These features are then embedded using Bidirectional Encoder Representations from Transformers and Global Vectors for Word Representation to form the node feature matrix. To model inter-feature dependen cies, we build a disambiguation feature graph whose edge weights are derived from pointwise mutual information, cosine similarity, and term frequency-inverse document frequency statistics. The resulting graph is encoded by a twin-tower architecture comprising a graph convolutional network and a graph attention network to perform complementary feature aggregation, while Kolmogorov-Arnold Network modules are incorporated into each tower to enhance nonlinear representation learning. Furthermore, we obtain the predictive distribution for each window and estimate its discriminative reliability via R & eacute;nyi entropy. Finally, an entropy-weighted soft voting strategy is employed to fuse window-level predictions, yielding the final semantic category for the ambiguous word. Experiments show that our model achieves 87.98% accuracy on Semantic Evaluation 2007 Task #5 and 97.08% on the Chinese Health Word Sense Disambiguation corpus, outperforming a strong baseline based on Bidirectional Encoder Representations from Transformers (85.68%/95.23%) by 2.30% and 1.85%, respectively.

In today's digital era, 3D models have become a key hot data form, which raises the critical issue of how to accurately and efficiently extract key features from massive 3D models. In existing view-based classification methods, views from different angles are often treated equally, ignoring the differences and importance of views at different angles. To solve these problems, this paper proposes a 3D model classification method based on multi-angle view fusion. First, multi-angle representative views of the model are obtained by setting viewpoints at 30° north latitude around the 3D model, with an interval of 60° between adjacent viewpoints. Then, the attention mechanism is introduced into the network to effectively extract deep features of the views. In the test phase, different weights are assigned to view features from different angles, and finally, a voting mechanism is used to complete the classification of 3D models. Experimental results show that under the optimal configuration (PreAct-ResNet+CBAM3+soft voting with hyperparameter tuning), the classification accuracy of this method on the ModelNet10 dataset reaches 96.61%, demonstrating outstanding classification performance.

As a crucial task in artificial intelligence, 3D model classification often suffers from an over-reliance on visual features that neglect intrinsic geometric information and creates vulnerability to adversarial perturbations. To address these challenges, this paper proposes the Selective Attention and Geometric Adversarial Network (SAGA-Net). Firstly, the method generates a set of 2D views using multi-view projections. Secondly, a dual-branch architecture is employed to extract view features using a Swin Transformer and model the shape descriptor sequence using a long short-term memory (LSTM) network to capture geometric information. Thirdly, a two-stage selective attention mechanism is used to select and fuse discriminative views, and the fused view features are multimodally integrated with global geometric features. Fourthly, an adversarial training strategy on geometric features is incorporated to enhance model robustness. Experimental results on the ModelNet10 dataset demonstrate that SAGA-Net achieves 96.37% accuracy and significantly outperforms baseline methods under adversarial attacks, which verifies its effectiveness in multimodal feature fusion and robust modeling.

Multimedia technology is widely applied to artificial intelligence and it is key to the performance of artificial intelligence systems, such as computer-aided design system, virtual reality system, the augmented reality system, medical image processing system, and game development system. With the development of multimedia technology, the number of three-dimensional (3D) models in network or database is becoming larger and larger. It is important to classify 3D models. In order to improve accuracy of three-dimensional model classification, a method of 3D model classification fusing MetaFormer with CNN (Convolutional Neural Networks) is proposed. 3D model is projected into two-dimensional (2D) views through the fixed-view projection, and representative views are selected by the clustering algorithm. Points sampled randomly from representative view are used to calculate its shape descriptors. View feature is extracted from representative view by MetaFormer. Shape feature is extracted from shape descriptors of representative view by Convolutional Neural Networks. View feature and shape feature of representative view are fused. At the same time, majority voting algorithm is used to determine category of 3D model based on the fusion of view features and shape features. Experiments are conducted on ModelNet10 dataset. Experimental results show that the proposed method achieves better results than others.

3D model classification has emerged as a significant research focus in computer vision. However, traditional convolutional neural networks (CNNs) often struggle to capture global dependencies across both height and width dimensions simultaneously, leading to limited feature representation capabilities when handling complex visual tasks. To address this challenge, we propose a novel 3D model classification network named ViT-GE (Vision Transformer with Global and Efficient Attention), which integrates Global Grouped Coordinate Attention (GGCA) and Efficient Channel Attention (ECA) mechanisms. Specifically, the Vision Transformer (ViT) is employed to extract comprehensive global features from multi-view inputs using its self-attention mechanism, effectively capturing 3D shape characteristics. To further enhance spatial feature modeling, the GGCA module introduces a grouping strategy and global context interactions. Concurrently, the ECA module strengthens inter-channel information flow, enabling the network to adaptively emphasize key features and improve feature fusion. Finally, a voting mechanism is adopted to enhance classification accuracy, robustness, and stability. Experimental results on the ModelNet10 dataset demonstrate that our method achieves a classification accuracy of 93.50%, validating its effectiveness and superior performance.

As deep learning techniques continue to advance in 3D model classification, traditional models often face challenges in resource-constrained environments, where balancing computational efficiency and classification accuracy becomes crucial. In this paper, we propose a lightweight 3D model classification framework based on the DecoupleNet network to enhance classification performance in such environments. To address the issue of lower accuracy with fewer parameters, we introduce a novel dual-branch Feature Integration Downsampling (WED) module, which combines wavelet transform downsampling, wavelet pooling, and the ECA attention mechanism. This approach preserves low-frequency structural information while enhancing high-frequency details, all while maintaining a lightweight structure. The WED module effectively balances global structural and local detailed features, improving feature discriminability and selectivity. Experimental results on the ModelNet10 dataset show a 1.72 % accuracy improvement over the original DecoupleNet, demonstrating the proposed method's efficiency and robustness in 3D model classification tasks.

3D model classification based on view has become a hot research topic. If all projection views of 3D model are treated equally, the importance and difference of different views, and complementary and correlation information between views will be ignored. In order to solve these issues, this paper proposes a method of 3D model classification based on Deep Residual Shrinkage Network (DRSN) and multi-view feature fusion. Firstly, 3D model is projected into six 2D views. Secondly, DRSN is used to extract view features from 2D views. Thirdly, shape distribution features D1, D2, D3 of 2D view are integrated with view features to get the fusion feature. Fourthly, the fusion feature is input into softmax function to get discriminative features and Shannon entropy is used to compute the uncertainty of view classification to measure view saliency. Fifthly, the fusion features of 2D views in descending order of view saliency are input into Long Short-Term Memory (LSTM) in sequence for fusing multi-view features. Finally, softmax function is adopted to classify 3D model based on multi-view fusion feature. Experimental results show that accuracy of the proposed method achieves 93.28% on ModelNet10 dataset and it demonstrates higher accuracy.

This study proposes a 3D model classification method that combines wavelet transform convolution and multi-view feature fusion, addressing the issue that traditional convolutional neural networks overlook high-frequency details and contextual information in images. The method first selects nine representative views of the 3D model and incorporates wavelet transform convolution into the EfficientNetV2 backbone network for global feature learning, enhancing the ability to capture features from different frequency domains. Additionally, a context-anchor attention mechanism is introduced to further improve detail recovery. Experimental results demonstrate that the proposed method achieves classification accuracies of $\mathbf{9 1. 1 5 \%}$ and 87.72 % on the ModelNet10 and ModelNet40 datasets, respectively, showing significant improvements over traditional convolutional neural networks and baseline models, particularly excelling in high-detail tasks.

Word Sense Disambiguation (WSD) is a fundamental task in Natural Language Processing, challenged by the presence of polysemous words in context. This paper proposes a novel WSD model, MCGF-BERT, which integrates BERT-based contextual embeddings with multi-scale convolutional features and an adaptive gating mechanism. BERT encodes rich semantic context, while parallel convolutional layers capture local features at varying granularities. A lightweight gating unit fuses these features and enhances discriminative information for sense disambiguation. We first conduct experiments on a manually annotated Chinese WSD dataset containing 40 ambiguous target words, where our model achieves an accuracy of 85.04%, outperforming BiLSTM (77.12%) and standard BERT (81.37%). To assess the model's generalizability, we extend the evaluation to two standard English benchmarks: SemCor and ALL-WSD. On SemCor, our model achieves 75.3% accuracy, surpassing GlossBERT (73.2%) and BEM (72.8%). On ALL-WSD, MCGFBERT reaches 74.1%, which demonstrates consistent advantages across benchmarks. Ablation studies further verify the contributions of the multi-scale fusion and the gating mechanism, with performance drops of 2.6% and 1.9% when these modules are removed, respectively. While our model is designed with computational efficiency in mind, detailed profiling of parameter overhead and inference latency is left for future work.

Word Sense Disambiguation (WSD) is an important research topic in the field of Natural Language Processing (NLP). Due to the strong influence of contextual words and sentence-level semantics on model performance, accurately identifying word meanings remains a significant challenge. To address this issue, this paper proposes a WSD method that combines joint learning from sequential corpora and randomly shuffled corpora. These two types of corpora are separately fed into two Transformer modules for feature extraction. An adaptive matrix and frequency-domain information are utilized to deeply analyze the context and effectively capture relevant information. Furthermore, the proposed method employs a gated fusion mechanism to integrate the learned representations and determine the semantic categories of ambiguous words. This study makes use of the SemEval-2007 Task #5 dataset and the Chinese-HealthWSD dataset, and applies the Transformer model to the WSD task for the first time. Experimental results demonstrate that this approach achieves a significant improvement in average accuracy compared to current state-of-the-art methods.

Objective In Word Sense Disambiguation(WSD),the Linguistically-motivated bidirectional Encoder Representation from Transformer(LERT)is employed to capture rich semantic representations from large-scale corpora,enabling improved contextual understanding of word meanings.However,several challenges remain.Current WSD models are not sufficiently sensitive to temporal and spatial dependencies within sequences,and single-dimensional features are inadequate for representing the diversity of linguistic expressions.To address these limitations,a hybrid network is constructed by integrating LERT,Bidirectional Gated Recurrent Units(Bi-GRU),and Graph Convolutional Network(GCN).This network enhances the modeling of structured text and contextual semantics.Nevertheless,generalization and robustness remain problematic.Therefore,an adversarial training algorithm is applied to improve the overall performance and resilience of the WSD model. Methods An adversarial WSD method is proposed based on a pre-trained model,combining Bi-GRU and GCN.First,word forms,parts of speech,and semantic categories of the neighboring words of an ambiguous term are input into the LERT model to obtain the CLS sequence and token sequence.Second,cross-attention is applied to fuse the global semantic information extracted by Bi-GRU from the token sequence with the local semantic information derived from the CLS sequence.Sentences,word forms,parts of speech,and semantic categories are then used as nodes to construct a disambiguation feature graph,which is subsequently input into GCN to update the feature information of the nodes.Third,the semantic category of the ambiguous word is determined through the interpolated prediction layer and semantic classification layer.Fourth,subtle continuous perturbations are generated by computing the gradient of the dynamic word vectors in the input.These perturbations are added to the original word vector matrix to create adversarial samples,which are used to optimize the LERT+Bi-GRU+CA+GCN(LBGCA-GCN)model.A cross-entropy loss function is applied to measure the performance of the LBGCA-GCN model on adversarial samples.Finally,the loss from the network is combined with the loss from AT to optimize the LBGCA-GCN model. Results and Discussions When the FreeLB algorithm is applied,stronger adversarial perturbations are generated,and the FreeLB algorithm achieves the best performance(Table 2).As the number of perturbation steps increases,the strength of AT improves.However,when the number of steps exceeds a certain threshold,the LBGCA-GCN+AT(LBGCA-GCN-AT)model begins to overfit.The Free Large-Batch(FreeLB)algorithm demonstrates strong robustness with three perturbation steps(Table 3).The cross-attention mechanism,which fuses the token sequence with the CLS sequence,yields significant performance gains in complex semantic scenarios(Fig.3).By incorporating AT,the LBGCA-GCN-AT model achieves notable improvements across multiple evaluation metrics(Table 4). Conclusions This study presents an adversarial WSD method based on a pre-trained model,integrating Bi-GRU and GCN to address the weak generalization ability and robustness of conventional WSD models.LERT is used to transform discriminative features into dynamic word vectors,while cross-attention fuses the global semantic information extracted by Bi-GRU from the token sequence with the local semantic information derived from the CLS sequence.This fusion generates more complete node representations for the disambiguation feature graph.A GCN is then applied to update the relationships among nodes within the feature graph.The interpolated prediction layer and semantic classification layer are used to determine the semantic category of ambiguous words.To further improve robustness,the gradient of the dynamic word vector is computed and perturbed to generate adversarial samples,which are used to optimize the LBGCA-GCN model.The network loss is combined with the AT loss to refine the model.Experiments conducted on the SemEval-2007 Task #05 and HealthWSD datasets examine multiple factors affecting model performance,including adversarial algorithms,perturbation steps,and sequence fusion methods.Results demonstrate that introducing AT improves the model's ability to handle real-world noise and perturbations.The proposed method not only enhances robustness and generalization but also strengthens the capacity of WSD models to capture subtle semantic distinctions.

At present, view-based 3D model classification has the problems of insufficient visual information for single view and redundant information for multiple views, and treating all views equally will ignore the differences between different projection angles. To solve the above problems, a 3D model classification method based on Shannon entropy representative feature and voting mechanism is proposed. Firstly, multiple angle groups are set uniformly around 3D model, and multiple view sets representing the model are obtained. In order to extract effectively deep features from view, channel attention mechanism is introduced into the feature extraction network. Secondly, based on view discriminative features output from Softmax function, Shannon entropy is used to select representative feature for avoiding redundant feature of multiple views. Finally, based on representative features from multiple angle groups, voting mechanism is used to classify 3D model. Experiments show that the classification accuracy of the proposed method on 3D model dataset ModelNet10 reaches 96.48％, and classification performance is outstanding.

3D (three-dimensional) models are widely applied in our daily life, such as mechanical manufacture, games, biochemistry, art, virtual reality, and etc. With the exponential growth of 3D models on web and in model library, there is an increasing need to retrieve the desired model accurately according to freehand sketch. Researchers are focusing on applying machine learning technology to 3D model retrieval. In this article, we combine semantic feature, shape distribution features and gist feature to retrieve 3D model based on interactive attention convolutional neural networks (CNN). The purpose is to improve the accuracy of 3D model retrieval. Firstly, 2D (two-dimensional) views are extracted from 3D model at six different angles and converted into line drawings. Secondly, interactive attention module is embedded into CNN to extract semantic features, which adds data interaction between two CNN layers. Interactive attention CNN extracts effective features from 2D views. Gist algorithm and 2D shape distribution (SD) algorithm are used to extract global features. Thirdly, Euclidean distance is adopted to calculate the similarity of semantic feature, the similarity of gist feature and the similarity of shape distribution feature between sketch and 2D view. Then, the weighted sum of three similarities is used to compute the similarity between sketch and 2D view for retrieving 3D model. It solves the problem that low accuracy of 3D model retrieval is caused by the poor extraction of semantic features. Nearest neighbor (NN), first tier (FT), second tier (ST), F-measure (E(F)), and discounted cumulated gain (DCG) are used to evaluate the performance of 3D model retrieval. Experiments are conducted on ModelNet40 and results show that the proposed method is better than others. The proposed method is feasible in 3D model retrieval.

With the development of multimedia technology, the number of 3D models on the web or in databases is becoming increasingly larger and larger. It becomes more and more important to classify and retrieve 3D models. 3D model classification plays important roles in the mechanical design field, education field, medicine field and so on. Due to the 3D model's complexity and irregularity, it is difficult to classify 3D model correctly. Many methods of 3D model classification pay attention to local features from 2D views and neglect the 3D model's contour information, which cannot express it better. So, accuracy the of 3D model classification is poor. In order to improve the accuracy of 3D model classification, this paper proposes a method based on EfficientNet and Convolutional Neural Network (CNN) to classify 3D models, in which view feature and shape feature are used. The 3D model is projected into 2D views from different angles. EfficientNet is used to extract view feature from 2D views. Shape descriptors D1, D2, D3, Zernike moment and Fourier descriptors of 2D views are adopted to describe the 3D model and CNN is applied to extract shape feature. The view feature and shape feature are combined as discriminative features. Then, the softmax function is used to determine the 3D model's category. Experiments are conducted on ModelNet 10 dataset. Experimental results show that the proposed method achieves better than other methods.