Journal of King Saud University-Computer and Information Sciences最新文献

In the rapidly evolving field of natural language processing, the demand for efficient automated text summarization systems that not only distill extensive documents but also capture their nuanced thematic elements has never been greater. This paper introduces the FuzzyTP-BERT framework, a novel approach in extractive text summarization that synergistically combines Fuzzy Topic Modeling (FuzzyTM) with the advanced capabilities of Bidirectional Encoder Representations from Transformers (BERT). Unlike traditional extractive methods, FuzzyTP-BERT integrates fuzzy logic to refine topic modeling, enhancing the semantic sensitivity of summaries by allowing a more nuanced representation of word-topic relationships. This integration results in summaries that are not only coherent but also thematically rich, addressing a significant gap in current summarization technology. Extensive evaluations on benchmark datasets demonstrate that FuzzyTP-BERT significantly outperforms existing models in terms of ROUGE scores, effectively balancing topical relevance with semantic coherence. Our findings suggest that incorporating fuzzy logic into deep learning frameworks can markedly improve the quality of automated text summaries, potentially benefiting a wide range of applications in the information overload age.

Recent GCN-based works have achieved remarkable results for skeleton-based human action recognition. Nevertheless, while existing approaches extensively investigate pairwise joint relationships, only a limited number of models explore the intricate, high-order relationships among multiple joints. In this paper, we propose a novel hypergraph convolution method that represents the relationships among multiple joints with hyperedges, and dynamically refines the height-order relationship between hyperedges in the spatial, temporal, and channel dimensions. Specifically, our method initiates with a temporal-channel refinement hypergraph convolutional network, dynamically learning temporal and channel topologies in a data-dependent manner, which facilitates the capture of non-physical structural information inherent in the human body. Furthermore, to model various inter-joint relationships across spatio-temporal dimensions, we propose a spatio-temporal hypergraph joint module, which aims to encapsulate the dynamic spatial–temporal characteristics of the human body. Through the integration of these modules, our proposed model achieves state-of-the-art performance on RGB+D 60 and NTU RGB+D 120 datasets.

Streaming applications are being extensively run on portable embedded systems, which are battery-operated and with limited memory. Thus, minimizing the total energy consumption of such a system is important. We investigate the problem of offline scheduling for streaming applications composed of non-preemptible periodic dependent tasks on homogeneous Network-on-Chip (NoC)-based Multiprocessor System-on-Chip (MPSoCs) such that their total energy consumption is minimized under memory constraints. We propose a novel unified approach that integrates task-level software pipelining with Dynamic Voltage and Frequency Scaling (DVFS) to solve the problem. Our approach is supported by a set of novel techniques, which include constructing an initial schedule based on a list scheduling where the priority of each task is its approximate successor-tree-consistent deadline such that the workload across all the processors is balanced, a retiming heuristic to transform intra-period dependencies into inter-period dependencies for enhancing parallelism, assigning an optimal discrete frequency for each task and each message using a Non-Linear Programming (NLP)-based algorithm and an Integer-Linear Programming (ILP)-based algorithm, and an incremental approach to reduce the memory usage of the retimed schedule in case of memory size violations. Using a set of real and synthetic benchmarks, we have implemented and compared our unified approach with two state-of-the-art approaches, RDAG+GeneS (Wang et al., 2011) , and JCCTS (Wang et al., 2013a). Experimental results show that our approach’s maximum, average, and minimum improvements over RDAG+GeneS (Wang et al., 2011) are 31.72%, 14.05%, and 7.00%, respectively. Our approach’s maximum, average, and minimum improvement over JCCTS (Wang et al., 2013a) are 35.58%, 17.04%, and 8.21%, respectively.

With the COVID-19 outbreak, health-related rumors have attracted significant attention from governments and global society. These rumors often mislead the public through multimedia, amplifying their negative impact and potentially manipulating public health narratives. On social media, detecting these rumors faces unique challenges, especially for emerging health events. Existing detection algorithms struggle because they mainly learn event-specific features that are not applicable to new or unseen events. To overcome this, we developed an end-to-end framework called the Health Domain Multimodal Rumor Detection Neural Network (HDRNN). This framework extracts invariant features and effectively detects new health-related rumors. It consists of three components: a multimodal feature extractor, a rumor detector, and an event discriminator. The multimodal feature extractor extracts text and visual features from posts, working with rumor detectors to learn discriminative features. Event discriminator remove specific features while retaining shared ones across events. Extensive experiments on datasets from Tencent News and Sina Weibo show that our HDRNN model excels in multimodal health rumor detection, surpassing current methods.

Currently, many scholars have investigated reversible data hiding (RDH) based on automatic contrast enhancement (ACE). For RDH based on ACE (ACERDH), various methods have been proposed on how to better improve image quality. Preserving brightness can prevent the image from being over-enhanced. However, some ACERDH methods that can preserve brightness well cannot sufficiently enhance low-light images. Although some ACERDH methods that cannot preserve brightness well can more sufficiently enhance low-light color images, they lack the consideration of the local region, which is not conducive to enhancing the contrast of the local region. Therefore, a novel RDH method based on global ACE and local ACE is proposed. The global contrast is first improved by equalizing the global histogram. Then the high complexity region is optimized by the proposed RDH method based on double-layer ACE to further enhance the local contrast. The low-light color image can be well enhanced by applying the proposed method to enhance the R, G, and B channels sequentially. Experimental results demonstrated that the proposed method is superior to some existing advanced methods in obtaining better image quality and hiding more secret data.

As the realities of production and operation in green and intelligent workshops become more variable, the adverse risks arising from disruptions to modernized workshop energy consumption schedules and customer churn caused by dynamic events are increasing. In order to solve those problems, we take the intelligent hybrid flow shop as the research subject, use buffer capacity and automated guided vehicles (AGVs) transport devices as resource constraints, construct a multi-objective rescheduling model that considers both energy consumption and customer satisfaction. According to the model characteristics, an improved lion swarm optimization algorithm (ILSO) is designed to solve the above model. To improve the initial solution quality and global search capability of the algorithm, ILSO is improved by combining the reverse learning initialization strategy of Logistic chaotic mapping with the tabu search strategy. The results of experiments on the proposed algorithm with different sizes of arithmetic cases and real cases in the workshop indicate that ILSO can effectively solve the bi-objective rescheduling problem oriented to inserting orders, and the proposed model can provide green dynamic scheduling solutions for manufacturing enterprises to achieve the purpose of transformation to green intelligent manufacturing.

This study introduces an Adaptive K-Nearest Neighbors methodology designed for FPGA platforms, offering substantial improvements over traditional K-Nearest Neighbors implementations. By integrating a dynamic classifier selection system, our approach enhances adaptability, enabling on-the-fly adjustments of K values and subsets of training data. This flexibility results in up to a 10.66% improvement in accuracy and significantly reduces latency, rendering our system up to 3.918 times more efficient than conventional K-Nearest Neighbors techniques. The methodology’s efficacy is validated through experiments across multiple datasets, demonstrating its potential in optimizing both classification accuracy and system efficiency. The adaptive approach’s ability to improve response times, along with its flexibility, positions it as an ideal solution for real-time applications and highlights the advantages of the adaptive K-Nearest Neighbors methodology in overcoming the constraints of hardware-accelerated machine learning.

Identifying splice sites is essential for gene structure analysis and eukaryotic genome annotation. Recently, computational and deep learning approaches for splice site detection have advanced, focusing on reducing false positives by distinguishing true from pseudo splice sites. This paper introduces GraphSplice, a method using graph convolutional neural networks. It encodes DNA sequences into directed graphs to extract features and predict splice sites. Tested across multiple datasets, GraphSplice consistently achieved high accuracy (91%-94%) and F1Scores (92%-94%), outperforming state-of-the-art models by up to 9.16% for donors and 5.64% for acceptors. Cross-species experiments also show GraphSplice’s capability to annotate splice sites in under-trained genomic datasets, proving its wide applicability as a tool for DNA splice site analysis.

Discovering trending social events (e.g., major meetings, political scandals, natural disasters, etc.) from social messages is vital because it emphasizes important events and can help people comprehend the world. However, the heterogeneous semantics enrichment, severe long-tail problems, and sparse text contents of social messages pose great challenges to event detection, often leading to limited generalization ability and accuracy. In this paper, we propose a novel Multi-Relational Meta-Enhanced Network (MRME-Net) architecture to learn social events. First, we model social messages into a multi-relational message graph, incorporating abundant meta-semantics along with various meta-relations. Second, we present a multi-relational graph attention network based on Sophia by using a dual-step message aggregation mechanisms to capture the local features of neighboring messages and global semantics of mutiple relations and ultimately learn social message embeddings. We use Sophia optimizer to reduce the massive time and cost of training. Third, in order to address the long-tail problem, we introduce a locally-adapted meta-learning framework in social event detection for the first time and propose a novel META-TAILENH embedding enhancement strategy to refine tail node embeddings in multi-relational graph. Eventually, we conduct the detection of social events according to the hierarchical clustering algorithm. Extensive experiments have been carried out to evaluate MRME-Net on the MAVEN and Twitter dataset, revealing a notable improvement of 3 %–13 %, 4 %–20 % and 6 %–30 % increases on NMI, AMI and ARI in the social event detection task.

In recent years, multimodal emotion recognition models are using pre-trained networks and attention mechanisms to pursue higher accuracy, which increases the training burden and slows down the training and inference speed. In order to strike a balance between speed and accuracy, this paper proposes a speed-optimized multimodal emotion recognition architecture for speech and text emotion recognition. In the feature extraction part, a lightweight residual graph convolutional network (ResGCN) is selected as the speech feature extractor, and an efficient RoBERTa pre-trained network is used as the text feature extractor. Then, an algorithm complexity-optimized sparse cross-modal encoder (SCME) is proposed and used to fuse these two types of features. Finally, a new gated fusion module (GF) is used to weight multiple results and input them into a fully connected layer (FC) for classification. The proposed method is tested on the IEMOCAP dataset and the MELD dataset, achieving weighted accuracies (WA) of 82.4% and 65.0%, respectively. This method achieves higher accuracy than the listed methods while having an acceptable training and inference speed.