Diffusion models have demonstrated impressive performance in text-to-image generation and image editing. However, in instruction-based image editing, they often encounter two challenges: (1) inaccurate localization of the editing targets and (2) unintended modifications in nontarget regions. These issues stem from the global processing of diffusion models due to attention mechanisms. To address these limitations, we conduct a systematic analysis of attention maps under editing instructions and design localization instructions to obtain the desired attention. We propose Instruction Attention Maps (IAM)-Edit, a localized image editing framework that explicitly decouples an editing pipeline into two stages: region localization followed by region-aware editing. Specifically, to localize the editing region, a mask is generated by clustering patches of self-attention maps and combining them with the focal points of cross-attention maps under the editing instruction. To preserve nonediting regions, we apply an attention modulation method that adjusts cross-attention weights at each denoising step based on the generated mask, enabling the denoising process to focus on the editing region. Experiments show that IAM-Edit outperforms state-of-the-art methods both qualitatively and quantitatively.

Prototypical networks are a typical method in few-shot learning. The instance-based training causes the model to learn specific domain features of the support set, leading to inadequate cross-domain generalization capabilities when faced with new domain data. To this end, this paper proposes a cross-domain Invariant Prototypical Contrastive Learning (IPCL) method. We extract invariant features (content features) and abstract invariant features (emotional features) from sample characteristics to construct invariant prototypes, which are domain-independent. Based on invariant representation learning, the model decouples latent features to obtain invariant features, which form the core basis for classification decisions. Abstract invariant features, refined from variable features (style features), reveal the underlying commonalities behind these variable features. Furthermore, we propose constructing auxiliary classification tasks that directly mine invariant prototype's semantic features, combined with metric tasks to achieve cross-task information complementarity and enhance the model's generalization ability to new domains. Experiments show that IPCL's decoupling of latent variables is recognizable, achieving accuracies of 92.23% and 93.45% on the ARSC and SST-2 datasets, respectively. In cross-domain tasks, IPCL outperforms previous method by up to 1.11%, demonstrating IPCL's superior cross-domain generalization capabilities.

Data assets have become a core driver of societal development; however, data originates from diverse channels and varies significantly in quality, making it unsuitable for direct analysis and decision-making without enhancement through data governance. To address the limitations of traditional data integration techniques in meeting the multi-modal data governance requirements of massive data environments, this paper proposes a data governance model based on multi-agents system. Leveraging the powerful computational and reasoning capabilities of large language Model, this model constructs specialized intelligent agents for tasks including metadata management, data standards management, data integration, data classification and grading, and data quality management, facilitating the evolution of data management from the integration phase towards intelligent governance. The paper further illustrates the model’s application within the context of university data governance and concludes with a summary of related work.

In recent years, vision-language pretraining (VLP) models have become a crucial driving force in the advancement of artificial intelligence. Besides, studies such as contrastive language-image pretraining (CLIP) have demonstrated that incorporating prompt learning within VLP models can significantly enhance the performance of downstream tasks. However, we believe that CLIP’s visual encoder suffers from feature extraction bias in image classification tasks, which is because of the uneven quantity and distribution of image features CLIP learned between the pretraining and fine-tuning stages. This can be further summarized as an inherent bias in feature extraction for differently distributed samples during the pretraining phase. To address the above problem this paper proposes (i) text-semantic hierarchical injection prompt learning method, which constructs self-attention layers and prompt mapping structures and injects text semantic features into the visual encoder layer by layer to generate visual prompt features and (ii) visual-semantic attention interactive prompt learning method, which further integrates text embeddings with the output features of the visual encoder through cross-attention and constructs instance-level text prompt features for each image. Based on the two above methods, this paper further proposes the multimodal coupling prompt learning CLIP (MCPL-CLIP) to enhance CLIP’s performance in image classification tasks. Experiments conducted on 15 image classification datasets demonstrate that MCPL-CLIP outperforms baseline models such as MaPLe, CoCoOp, and CoOp in cross-dataset transfer, domain generalization, and base-to-novel class generalization tasks, showcasing its superior text semantic representation and visual feature extraction capabilities.

To address problems such as ineffective real-time monitoring and weak decision-making support in the development of disciplines in universities, this study proposes a monitoring system for discipline development level based on big data, which is grounded in ShanghaiRanking's Academic Ranking of World Universities and Double First-class discipline evaluation indicators and integrates the discipline evaluation indicators and methods of Times and QS Higher Education World University Rankings to establish a multidimensional discipline development evaluation system. From macro overview to micro insight, this system offers three-dimensional evaluation and analysis of key discipline indicators through vertical in-depth exploration and horizontal multi-dimensional comparison. It realizes the digital real-time monitoring of discipline construction, tracks discipline development, analyzes gap between disciplines, and provides visual monitoring and analysis-based early warning of discipline development at the university,faculty and discipline levels, providing strong data support for the relevant decisions of discipline construction and planning.

In response to the requirements for intelligence, collaboration, and interactivity in undergraduate teaching evaluation under the background of big data technology, this paper proposed an undergraduate education teaching evaluation system based on big data technology. Based on the Analytic Hierarchy Process and Entropy Weight Method, the system linearly fits the two weights to calculate the indicator system weights for teaching evaluation. Under the indicator system framework, the system can achieve multi-form, multi-dimensional, and three-dimensional undergraduate evaluation data collection, monitoring, and analysis to assist schools in providing accurate support for training programs, teaching investment, scientific innovation, and employment. The system focuses on the real-time and dynamic nature of self-evaluation in universities, integrating data collection, intelligent analysis, and precise decision support, enhancing the role of evaluation in promoting teaching, optimizing the decision-making ability of universities, and realizing the re-evaluation of the value of evaluation in universities.

Pre-trained Visual Language Models (VLMs) like CLIP have shown great potential in the multimodal domain. Among this, using different modal contexts and interaction features to construct prompt can stimulate the model’s prior knowledge circuit more accurately, thus generating better outputs. However, in CLIP, the formal mismatch of textual descriptions between the pre-training and inference phases results in a suboptimal representation ability of prompt, which is detrimental to model alignment learning. Therefore, R egion- A ttention P rompt (RAP) is proposed, which introduces region features to enrich the semantic representation of prompt. RAP is acquired by the Cross-Attention mechanism between images and texts, and it is essentially a region-level prompt with category-sensitive properties. For each category, RAP adaptively assigns greater attention weight to image regions that are more semantically relevant to the category. Besides, CLIP is equipped with RAP (called RA-CLIP) to improve image classification performance in generalization scenarios. Extensive experiments demonstrate that RA-CLIP outperforms the current SOTA CoCoOp 0.4% - 4.16% on base classes and 0.25% - 11.34% on new classes, across 7 datasets. In addition, we show that focusing on category-related regions to construct prompt can further improve the model’s alignment ability.

As prototype-based Few-Shot Learning methods, Prototypical Network generates prototypes for each class in a low-resource state and classify by a metric module. Therefore, the quality of prototypes matters but they are inaccurate from the few support instances, and the domain-specific information of training data are harmful to the generalizability of prototypes. We propose a Conceptual Prototype (CP), which contains both rich instance and concept features. The numerous query data can inspire the few support instances. An interactive network is designed to leverage the interrelation between support set and query-detached set to acquire a rich Instance Prototype which is typical on the whole data. Besides, class labels are introduced to prototype by prompt engineering, which makes it more conceptual. The label-only concept makes prototype immune to domain-specific information in training phase to improve its generalizability. Based on CP, Conceptual Prototypical Contrastive Learning (CPCL) is proposed where PCL brings instances closer to its corresponding prototype and pushes away from other prototypes. “2-way 5-shot” experiments show that CPCL achieves 92.41% accuracy on ARSC dataset, 2.30% higher than other prototype-based models. Meanwhile, the 0-shot performance of CPCL is comparable to Induction Network in the 5-shot way, indicating that our model is adequate for 0-shot tasks.

为了对学生一卡通数据、教务数据、图书馆数据进行相关性研究,挖掘这些数据的隐藏价值,文中提出基于聚类算法的学生消费数据与成绩数据的关联性分析.首先对原始数据进行预处理,使用Cubic插值法对缺失值进行插补处理,并采用主成分分析法将原始数据降至一维.然后通过改进的聚类算法K-Means++算法对降维后的数据进行了初步的聚类,并对不同k值取得的聚类结果进行性能比较,最后对8287个实验样本进行关联性计算.所得出实验的结果可以为学校提供决策帮助.

Contrastive self-supervised learning has recently shown promising progress in representation learning, such as SimCLR, wherein negative samples (negatives) play a significant role in learning complex semantic representation. However, increasing the number of negatives leads to a significant increase in computational resources. Meanwhile, only a few negatives–hard negatives–that are closer to positives are crucial for model training. Therefore, we propose Hard-Negatives Focused Strategy (HNFS) to improve the quality of samples by focusing on hard negatives in the training of SimCLR. Specifically, HNFS calculates the Impact of negatives by power function distribution, which dynamically assigns a higher weight to harder negatives and replaces the similarity matrix with the Impact matrix for contrastive learning. The non-linear incremental nature of power function distribution can widen the focus difference between hard negatives and other negatives, thus enhancing the model’s willingness to effectively learn semantic representations from hard negatives. We propose Hard-Negatives Focused SimCLR (HNF-SimCLR) for text classification and similar sentence pair tasks. Extensive experiments demonstrate that HNF-SimCLR outperforms all baselines across SST-2, ARSC, and QQP datasets (for example, Precision improvement from 0.42% to 4.35%). Meanwhile, in the ablation study we propose two linear weighting algorithms for comparison and suggest that HNFS facilitates more effective learning of hard negatives than both linear weighting algorithms, thus improving the model’s performance.

Instance-level contrastive learning such as SimCLR has been successful as a powerful method for representation learning. However, SimCLR suffers from problems of sampling bias, feature bias and model collapse. A set-level based S ampling E nhanced C ontrastive L earning (SECL) method based on SimCLR is proposed in this paper. We use the proposed super-sampling method to expand the augmented samples into a contrastive-positive set, which can learn class features of the target sample to reduce the bias. The contrastive-positive set includes Augmentations (the original augmented samples) and Neighbors (the super-sampled samples). We also introduce a samples-correlation strategy to prevent model collapse, where a positive correlation loss or a negative correlation loss is computed to adjust the balance of model’s Alignment and Uniformity. SECL reaches 94.14% classification precision on SST-2 dataset and 89.25% on ARSC dataset. For the multi-class classification task, SECL achieves 90.99% on AGNews dataset. They are all about 1% higher than the precision of SimCLR. Experiments show that the training convergence of SECL is faster, and SECL reduces the risk of bias and model collapse.

The applications of large language models (LLMs) such as ChatGPT exhibit impressive comprehension and generative capabilities in dialogue task. LLMs require massive high-quality data and computational cost, which limits their application to low-resource tasks. Dialogue generation when using smaller language models like GPT-2 encounters difficulties in maintaining context consistency. To address the problem of dialogue generation under resource constraints, we propose an LLM-guided Generative Prompt method (LGP). LGP enhances the relevance and coherence of generated dialogues through a smaller model GPT-2 and generative prompt (GP). GP is produced by the proposed Prompt Network, which leverages prompt encoder to learn dialogue history features and utilizes LSTM to extract contextual temporal features. Therefore, GP shown as the simple fixed-length learnable embeddings can replace the original complex and redundant context in GPT-2. The few-shot training of GP is guided by the LLM’s responses, which facilitates GPT-2 in generating more contextually consistent and comprehensive responses. Experiments on the DailyDialog and MultiWOZ datasets show that LGP achieves high improvements in BLEU, NIST, METEOR and ROUGE-L metrics. Remarkably, LGP achieves these results with approximately 18% of the training data, surpassing other full-data-finetuning methods in automatic evaluation metrics.

少样本文本分类中,原型网络对语义利用不足、可迁移特征挖掘不够,导致模型泛化能力不强,在新任务空间中分类性能不佳.从模型结构、编码网络、度量网络等角度提高模型泛化性,提出多任务原型网络(multi-ple-task prototypical network,MTPN).结构上,基于原型网络度量任务增加辅助分类任务约束训练目标,提高了模型的语义特征抽取能力,利用多任务联合训练,获得与辅助任务更相关的语义表示.针对编码网络,提出LF-Transformer编码器,使用层级注意力融合底层通用编码信息,提升特征的可迁移性.度量网络使用基于BiG-RU的类原型生成器,使类原型更具代表性,距离度量更加准确.实验表明,MTPN在少样本文本情感分类任务中取得了 91.62％的准确率,比现有最佳模型提升了 3.5％以上;在新领域的情感评论中,基于五条参考样本,模型对查询样本可获得超过90％的分类准确率.

传统的属性级别情感分析方法缺乏对属性实体与前后文之间交互关系的研究,导致情感分类结果的正确率不高.为了有效提取文本特征,提出了一种利用多头注意力机制学习属性实体与前后文之间关系的属性级别情感分析模型(intra &inter multi-head attention network,IIMAN),从而提高情感极性判断结果.该模型首先利用BERT预训练完成输入语句的词向量化;通过注意力网络中的内部多头注意力与联合多头注意力学习属性实体与前后文以及前后文内部间的关系;最后通过逐点卷积变换层、面向属性实体的注意力层和输出层完成情感极性分类.通过在三个公开的属性级别情感分析数据集Twitter、laptop、restaurant上的实验证明,IIMAN 相较于其他基线模型,正确率和F1 值有了进一步的提升,能够有效提高情感极性分类结果.

In contrastive self-supervised learning, such as Sim-CLR and MOCO, negative samples play an important role on the model's robustness. Instead of increasing the negatives' number, we promote the in-batch negatives' quality by adjusting their embedding. In-Batch Negatives' Enhanced Self-Supervised Learning (IBN-SSL) focus on negatives' quality by an importance-weighted algorithm and an online boundary. The importance-weighted negatives' quality algorithm decreases the negatives' relative locations from positive in projection space, which pushes the model to learn to distinguish positives from harder negatives. And the online boundary compresses negatives' vector space and keep negatives farther from positive. A better representation model can be leaned by IBN-SSL, and experiments show that it outperforms both RoBERTa and SimCLR in text classification task and similar sentence pair task.

近年来,恶意软件大多数通过僵尸网络技术攻击,并在通信过程中利用命令与控制服务器进行信息交互,从而窃取个入信息,甚至实施大规模网络攻击.由于网络安全人员可以利用流量封锁、域名黑名单等对恶意域名进行限制和过滤,所以攻击者开始采用Domain Flux技术,以及域名生成算法(Domain Generation Algorithm,DGA).

Generic language models pretrained on large unspecific domains are currently the foundation of NLP. Labeled data are limited in most model training due to the cost of manual annotation, especially in domains including massive Proper Nouns such as mathematics and biology, where it affects the accuracy and robustness of model prediction. However, directly applying a generic language model on a specific domain does not work well. This paper introduces a BERT-based text classification model enhanced by unlabeled data (UL-BERT) in the LaTeX formula domain. A two-stage Pretraining model based on BERT(TP-BERT) is pretrained by unlabeled data in the LaTeX formula domain. A double-prediction pseudo-labeling (DPP) method is introduced to obtain high confidence pseudo-labels for unlabeled data by self-training. Moreover, a multi-rounds teacher–student model training approach is proposed for UL-BERT model training with few labeled data and more unlabeled data with pseudo-labels. Experiments on the classification of the LaTex formula domain show that the classification accuracies have been significantly improved by UL-BERT where the F1 score has been mostly enhanced by 2.76%, and lower resources are needed in model training. It is concluded that our method may be applicable to other specific domains with enormous unlabeled data and limited labelled data.

Botnets have become the infrastructure of cryptocurrency in recent years, but traditional graph-based detection methods ignore multiple flows and their features. We propose a botnet detection method (ME-LGCN) by node classification based on the fine-grained multilateral attribute graph (fMAG). Multiple flows and their features are appended on the simple graph of network topology as multilateral structures and attributes in fMAG. Latent Graph Convolutional Neural Network (Latent-GCN) is used for node classification, where multi-edge embedding learns the multilateral attributes as an interaction vector, direct on-vertex embedding extends node representation, and GCN aggregates information of neighborhoods. Experiments on real datasets show that ME-LGCN provides significant improvements compared to other methods with a more than 3% improvement in F1.

针对长文本自动摘要任务中抽取式模型摘要较为冗余,而生成式摘要模型时常有关键信息丢失、摘要不准确和生成内容重复等问题,提出一种面向长文本的基于优势演员-评论家算法的强化自动摘要模型(A2C-RLAS).首先,用基于卷积神经网络(CNN)和循环神经网络(RNN)的混合神经网络的抽取器(extractor)来提取原文关键句;然后,用基于拷贝机制和注意力机制的重写器(rewriter)来精炼关键句;最后,使用强化学习的优势演员-评论家(A2C)算法训练整个网络,把重写摘要和参考摘要的语义相似性(BERTScore值)作为奖励(reward)来指导抽取过程,从而提高抽取器提取句子的质量.在CNN/Daily Mail数据集上的实验结果表明,与基于强化学习的抽取式摘要(Refresh)模型、基于循环神经网络的抽取式摘要序列模型(SummaRuNNer)和分布语义奖励(DSR)模型等模型相比,A2C-RLAS的最终摘要内容更加准确、语言更加流畅,冗余的内容有效减少,且A2C-RLAS的ROUGE和BERTScore指标均有提升.相较于Refresh模型和SummaRuNNer模型,A2C-RLAS模型的ROUGE-L值分别提高了6.3％和10.2％;相较于DSR模型,A2C-RLAS模型的F1值提高了30.5％.