Classifying job offers into occupational categories is a fundamental task in human resource information systems,as it improves and streamlines indexing,search,and matching between openings and job seekers.Comprehensiv...Classifying job offers into occupational categories is a fundamental task in human resource information systems,as it improves and streamlines indexing,search,and matching between openings and job seekers.Comprehensive occupational databases such as O∗NET or ESCO provide detailed taxonomies of interrelated positions that can be leveraged to align the textual content of postings with occupational categories,thereby facilitating standardization,cross-system interoperability,and access to metadata for each occupation(e.g.,tasks,knowledge,skills,and abilities).In this work,we explore the effectiveness of fine-tuning existing language models(LMs)to classify job offers with occupational descriptors from O∗NET.This enables a more precise assessment of candidate suitability by identifying the specific knowledge and skills required for each position,and helps automate recruitment processes by mitigating human bias and subjectivity in candidate selection.We evaluate three representative BERT-like models:BERT,RoBERTa,and DeBERTa.BERT serves as the baseline encoder-only architecture;RoBERTa incorporates advances in pretraining objectives and data scale;and DeBERTa introduces architectural improvements through disentangled attention mechanisms.The best performance was achieved with the DeBERTa model,although the other models also produced strong results,and no statistically significant differences were observed acrossmodels.We also find that these models typically reach optimal performance after only a few training epochs,and that training with smaller,balanced datasets is effective.Consequently,comparable results can be obtained with models that require fewer computational resources and less training time,facilitating deployment and practical use.展开更多
Long-term traffic flow prediction is a crucial component of intelligent transportation systems within intelligent networks,requiring predictive models that balance accuracy with low-latency and lightweight computation...Long-term traffic flow prediction is a crucial component of intelligent transportation systems within intelligent networks,requiring predictive models that balance accuracy with low-latency and lightweight computation to optimize trafficmanagement and enhance urban mobility and sustainability.However,traditional predictivemodels struggle to capture long-term temporal dependencies and are computationally intensive,limiting their practicality in real-time.Moreover,many approaches overlook the periodic characteristics inherent in traffic data,further impacting performance.To address these challenges,we introduce ST-MambaGCN,a State-Space-Based Spatio-Temporal Graph Convolution Network.Unlike conventionalmodels,ST-MambaGCN replaces the temporal attention layer withMamba,a state-space model that efficiently captures long-term dependencies with near-linear computational complexity.The model combines Chebyshev polynomial-based graph convolutional networks(GCN)to explore spatial correlations.Additionally,we incorporate a multi-temporal feature capture mechanism,where the final integrated features are generated through the Hadamard product based on learnable parameters.This mechanism explicitly models shortterm,daily,and weekly traffic patterns to enhance the network’s awareness of traffic periodicity.Extensive experiments on the PeMS04 and PeMS08 datasets demonstrate that ST-MambaGCN significantly outperforms existing benchmarks,offering substantial improvements in both prediction accuracy and computational efficiency for long-term traffic flow prediction.展开更多
Laser additively manufactured microscale metallic lattices show great potential for high-performance applications,yet trade-offs among geometric precision,structural integrity,and computational efficiency still persis...Laser additively manufactured microscale metallic lattices show great potential for high-performance applications,yet trade-offs among geometric precision,structural integrity,and computational efficiency still persist.Here,we introduce a stereolithography file format-free(STL-free)hybrid toolpath generation method for laser-based powder bed fusion(PBF-LB)that synergizes implicit geometric modeling with optimized laser scanning strategy,overcoming these limitations.By circumventing traditional mesh-based workflows,our method directly translates implicit lattice geometries into laser toolpaths while precisely regulating energy deposition trajectories.This mesh-free process enables the fabrication of complex shell lattices with ultra-thin walls and enhanced surface quality.In addition to reducing memory usage and processing time by up to 90%,the method yields a synergistic enhancement in mechanical performance,notably improving both strength and toughness.By bridging computational design and fabrication,this framework enables the scalable production of high-performance microscale lattices and unlocks their potential for industrial applications.展开更多
In the field of smart agriculture,accurate and efficient object detection technology is crucial for automated crop management.A particularly challenging task in this domain is small object detection,such as the identi...In the field of smart agriculture,accurate and efficient object detection technology is crucial for automated crop management.A particularly challenging task in this domain is small object detection,such as the identification of immature fruits or early stage disease spots.These objects pose significant difficulties due to their small pixel coverage,limited feature information,substantial scale variations,and high susceptibility to complex background interference.These challenges frequently result in inadequate accuracy and robustness in current detection models.This study addresses two critical needs in the cashew cultivation industry—fruitmaturity and anthracnose detection—by proposing an improved YOLOv11-NSDDil model.The method introduces three key technological innovations:(1)The SDDil module is designed and integrated into the backbone network.This module combines depthwise separable convolution with the SimAM attention mechanism to expand the receptive field and enhance contextual semantic capture at a low computational cost,effectively alleviating the feature deficiency problem caused by limited pixel coverage of small objects.Simultaneously,the SDmodule dynamically enhances discriminative features and suppresses background noise,significantly improving the model’s feature discrimination capability in complex environments;(2)The introduction of the DynamicScalSeq-Zoom_cat neck network,significantly improving multi-scale feature fusion;and(3)The optimization of the Minimum Point Distance Intersection over Union(MPDIoU)loss function,which enhances bounding box localization accuracy byminimizing vertex distance.Experimental results on a self-constructed cashew dataset containing 1123 images demonstrate significant performance improvements in the enhanced model:mAP50 reaches 0.825,a 4.6% increase compared to the originalYOLOv11;mAP50-95 improves to 0.624,a 6.5% increase;and recall rises to 0.777,a 2.4%increase.This provides a reliable technical solution for intelligent quality inspection of agricultural products and holds broad application prospects.展开更多
文摘Classifying job offers into occupational categories is a fundamental task in human resource information systems,as it improves and streamlines indexing,search,and matching between openings and job seekers.Comprehensive occupational databases such as O∗NET or ESCO provide detailed taxonomies of interrelated positions that can be leveraged to align the textual content of postings with occupational categories,thereby facilitating standardization,cross-system interoperability,and access to metadata for each occupation(e.g.,tasks,knowledge,skills,and abilities).In this work,we explore the effectiveness of fine-tuning existing language models(LMs)to classify job offers with occupational descriptors from O∗NET.This enables a more precise assessment of candidate suitability by identifying the specific knowledge and skills required for each position,and helps automate recruitment processes by mitigating human bias and subjectivity in candidate selection.We evaluate three representative BERT-like models:BERT,RoBERTa,and DeBERTa.BERT serves as the baseline encoder-only architecture;RoBERTa incorporates advances in pretraining objectives and data scale;and DeBERTa introduces architectural improvements through disentangled attention mechanisms.The best performance was achieved with the DeBERTa model,although the other models also produced strong results,and no statistically significant differences were observed acrossmodels.We also find that these models typically reach optimal performance after only a few training epochs,and that training with smaller,balanced datasets is effective.Consequently,comparable results can be obtained with models that require fewer computational resources and less training time,facilitating deployment and practical use.
基金supported byNationalNatural Science Foundation of China,GrantNo.62402046the Beijing Forestry University Science and Technology Innovation Project under Grant No.BLX202358.
文摘Long-term traffic flow prediction is a crucial component of intelligent transportation systems within intelligent networks,requiring predictive models that balance accuracy with low-latency and lightweight computation to optimize trafficmanagement and enhance urban mobility and sustainability.However,traditional predictivemodels struggle to capture long-term temporal dependencies and are computationally intensive,limiting their practicality in real-time.Moreover,many approaches overlook the periodic characteristics inherent in traffic data,further impacting performance.To address these challenges,we introduce ST-MambaGCN,a State-Space-Based Spatio-Temporal Graph Convolution Network.Unlike conventionalmodels,ST-MambaGCN replaces the temporal attention layer withMamba,a state-space model that efficiently captures long-term dependencies with near-linear computational complexity.The model combines Chebyshev polynomial-based graph convolutional networks(GCN)to explore spatial correlations.Additionally,we incorporate a multi-temporal feature capture mechanism,where the final integrated features are generated through the Hadamard product based on learnable parameters.This mechanism explicitly models shortterm,daily,and weekly traffic patterns to enhance the network’s awareness of traffic periodicity.Extensive experiments on the PeMS04 and PeMS08 datasets demonstrate that ST-MambaGCN significantly outperforms existing benchmarks,offering substantial improvements in both prediction accuracy and computational efficiency for long-term traffic flow prediction.
基金financial support of the Hong Kong Special Administrative Region University Grants Committee—General Research Fund CUHK14209523Collaborative Research Fund C4074-22G,C4002-22Y and C7074-23Gsupport by the University of Massachusetts Amherst。
文摘Laser additively manufactured microscale metallic lattices show great potential for high-performance applications,yet trade-offs among geometric precision,structural integrity,and computational efficiency still persist.Here,we introduce a stereolithography file format-free(STL-free)hybrid toolpath generation method for laser-based powder bed fusion(PBF-LB)that synergizes implicit geometric modeling with optimized laser scanning strategy,overcoming these limitations.By circumventing traditional mesh-based workflows,our method directly translates implicit lattice geometries into laser toolpaths while precisely regulating energy deposition trajectories.This mesh-free process enables the fabrication of complex shell lattices with ultra-thin walls and enhanced surface quality.In addition to reducing memory usage and processing time by up to 90%,the method yields a synergistic enhancement in mechanical performance,notably improving both strength and toughness.By bridging computational design and fabrication,this framework enables the scalable production of high-performance microscale lattices and unlocks their potential for industrial applications.
基金supported by Hebei North University Doctoral Research Fund Project(No.BSJJ202315)the Youth Research Fund Project of Higher Education Institutions in Hebei Province(No.QN2024146).
文摘In the field of smart agriculture,accurate and efficient object detection technology is crucial for automated crop management.A particularly challenging task in this domain is small object detection,such as the identification of immature fruits or early stage disease spots.These objects pose significant difficulties due to their small pixel coverage,limited feature information,substantial scale variations,and high susceptibility to complex background interference.These challenges frequently result in inadequate accuracy and robustness in current detection models.This study addresses two critical needs in the cashew cultivation industry—fruitmaturity and anthracnose detection—by proposing an improved YOLOv11-NSDDil model.The method introduces three key technological innovations:(1)The SDDil module is designed and integrated into the backbone network.This module combines depthwise separable convolution with the SimAM attention mechanism to expand the receptive field and enhance contextual semantic capture at a low computational cost,effectively alleviating the feature deficiency problem caused by limited pixel coverage of small objects.Simultaneously,the SDmodule dynamically enhances discriminative features and suppresses background noise,significantly improving the model’s feature discrimination capability in complex environments;(2)The introduction of the DynamicScalSeq-Zoom_cat neck network,significantly improving multi-scale feature fusion;and(3)The optimization of the Minimum Point Distance Intersection over Union(MPDIoU)loss function,which enhances bounding box localization accuracy byminimizing vertex distance.Experimental results on a self-constructed cashew dataset containing 1123 images demonstrate significant performance improvements in the enhanced model:mAP50 reaches 0.825,a 4.6% increase compared to the originalYOLOv11;mAP50-95 improves to 0.624,a 6.5% increase;and recall rises to 0.777,a 2.4%increase.This provides a reliable technical solution for intelligent quality inspection of agricultural products and holds broad application prospects.
