Document image retrieval (DIR) aims to retrieve document images from a gallery according to a given query. Existing DIR methods are primarily based on image queries that retrieve documents within the same coarse semantic category, e.g., newspapers or receipts. However, these methods struggle to effectively retrieve document images in real-world scenarios where textual queries with fine-grained semantics are usually provided. To bridge this gap, we introduce a new Natural Language-based Document Image Retrieval (NL-DIR) benchmark with corresponding evaluation metrics. In this work, natural language descriptions serve as semantically rich queries for the DIR task. The NL-DIR dataset contains 41K authentic document images, each paired with five high-quality, fine-grained semantic queries generated and evaluated through large language models in conjunction with manual verification. We perform zero-shot and fine-tuning evaluations of existing mainstream contrastive vision-language models and OCR-free visual document understanding (VDU) models. A two-stage retrieval method is further investigated for performance improvement while achieving both time and space efficiency. We hope the proposed NL-DIR benchmark can bring new opportunities and facilitate research for the VDU community. Datasets and codes will be publicly available at huggingface.co/datasets/nianbing/NL-DIR.

Multimodal relation extraction (MRE) aims at predicting the semantic relation between two entities given a hybrid context of a text and its related image. Though existing MRE methods have explored different strategies to fuse multimodal information, they suffer from two limitations. First, they ignore fine-grained visual relations between objects which can provide important hints for inferring the correct relation. Second, they neglect informative textual evidence from the image, leading to a performance decline when processing text-intensive images. To address above issues, we propose a novel MRE model, named VRTE, which takes full advantage of both Visual Relations and Textual Evidence to determine the final relation label. Specifically, the input image-text pair is transformed into two scene graphs, which are further bridged into a unified multimodal graph. Next, the relation-aware Transformer is utilized to propagate information in the multimodal graph while explicitly encoding diverse relations among visual objects and textual tokens via learnable relation embeddings. Besides, a cross-attention mechanism is also used to capture valuable textual information in the OCR results and image captions, which is combined with the representations of entity nodes and original text to predict the final relation label. Experimental results on the MNRE dataset demonstrate the effectiveness of the proposed model. Extensive ablation studies are also conducted to analyse contributions of different modules.

Given a text and its related image, the multimodal relation extraction (MRE) task aims at predicting the correct semantic relation between two entities in the input text. Though certain advances have been made by recent MRE approaches, they still suffer from two drawbacks. First, they ignore fine-grained visual relations between image objects which can serve as important clues for inferring the correct relation. Second, they are unable to utilize useful external knowledge about entities and the input sentence, leading to a sub-optimal performance when processing examples that demand related background or commonsense knowledge. To alleviate above limitations, we propose a novel method, named VRMK, which exploits both Visual Relations and Multi-grained Knowledge for the MRE task. Specifically, the input image-text pair is converted into a unified multimodal graph. Then, the relation-aware transformer is adopted to update node representations while explicitly encoding diverse relations among visual and textual nodes. Based on the input text, the powerful large language model (LLM) is used to generate entity-level and sentence-level knowledge with the in-context learning. The most relevant knowledge information is captured by a cross-attention mechanism and is further combined with the representations of entity nodes and original text to predict the final relation label. On the MNRE dataset, VRMK outperforms recent state-of-the-art baselines including LLM-based methods by 2.71% (82.55%→85.26% in F1 score). We also conduct extensive ablation experiments to reveal contributions of different modules and provide useful insights for future research.

Due to the flexible representation of arbitrary-shaped scene text and simple pipeline, bottom-up segmentation-based methods begin to be mainstream in real-time scene text detection. Despite great progress, these methods show deficiencies in robustness and still suffer from false positives and instance adhesion. Different from existing methods which integrate multiple-granularity features or multiple outputs, we resort to the perspective of representation learning in which auxiliary tasks are utilized to enable the encoder to jointly learn robust features with the main task of per-pixel classification during optimization. For semantic representation learning, we propose global-dense semantic contrast (GDSC), in which a vector is extracted for global semantic representation, then used to perform element-wise contrast with the dense grid features. To learn instance-aware representation, we propose to combine top-down modeling (TDM) with the bottom-up framework to provide implicit instance-level clues for the encoder. With the proposed GDSC and TDM, the encoder network learns stronger representation without introducing any parameters and computations during inference. Equipped with a very light decoder, the detector can achieve more robust real-time scene text detection. Experimental results on four public datasets show that the proposed method can outperform or be comparable to the state-of-the-art on both accuracy and speed. Specifically, the proposed method achieves 87.2% F-measure with 48.2 FPS on Total-Text and 89.6% F-measure with 36.9 FPS on MSRA-TD500 on a single GeForce RTX 2080 Ti GPU.

It aims to improve the efficiency of information collection and extraction in the current intelligent transportation system, and accurately mine the vehicle trajectory data By using Artificial Intelligence (AI) and Deep Learning methods, the trajectory data generated during vehicle driving are deeply mined and analyzed, and the characteristics of driving behavior of vehicle drivers are modeled and analyzed in detail. Then, a method of mining driving behavior characteristics based on Convolutional Neural Network (CNN) and vehicle trajectory is proposed. Based on the mathematical principle of wavelet packet and Least Square Support Vector Machine (LSSVM), a combined model of trajectory mining is constructed and applied to the short-term prediction of traffic flow. The traffic flow of Binjiang Road and Renmin Road in Guangzhou, Guangdong Province from August 19 to August 21, 2021 is predicted to verify the accuracy of the trajectory mining combined model. The results show that the combination model of data mining has good fitting effect, and the average accuracy is above 0.8. Besides, the effectiveness of the Deep Learning model in driver behavior classification is verified. The accuracy of the classification model is 75.2% for trajectory, and that is 76.8% for driver behavior characteristics. It is of great significance to effectively utilize the knowledge data in Intelligent Transportation System (ITS) and extract valuable information from it, which has certain reference value for the subsequent refined prediction of vehicle behavior.

Cross-validation (CV) is a widely adopted approach for selecting the optimal model. However, the computation of empirical cross-validation error (CVE) has high complexity due to multiple times of learner training. In this paper, we develop a novel approximation theory of CVE and present an approximate approach to CV based on the Bouligand influence function (BIF) for kernel-based algorithms. We first represent the BIF and higher order BIFs in Taylor expansions, and approximate CV via the Taylor expansions. We then derive an upper bound of the discrepancy between the original and approximate CV. Furthermore, we provide a novel computing method to calculate the BIF for general distribution, and evaluate BIF criterion for sample distribution to approximate CV. The proposed approximate CV requires training on the full data set only once and is suitable for a wide variety of kernel-based algorithms. Experimental results demonstrate that the proposed approximate CV is sound and effective.

Cross-validation (CV) is the most widely adopted approach for selecting the optimal model. However, the computation of CV has high complexity due to multiple times of learner training, making it disabled for large scale model selection. In this paper, we present an approximate approach to CV based on the theoretical notion of Bouligand influence function (BIF) and the Nystr\"{o}m method for kernel methods. We first establish the relationship between the theoretical notion of BIF and CV, and propose a method to approximate the CV via the Taylor expansion of BIF. Then, we provide a novel computing method to calculate the BIF for general distribution, and evaluate BIF for sample distribution. Finally, we use the Nystr\"{o}m method to accelerate the computation of the BIF matrix for giving the finally approximate CV criterion. The proposed approximate CV requires training only once and is suitable for a wide variety of kernel methods. Experimental results on lots of datasets how that our approximate CV has no statistical discrepancy with the original CV, but can significantly improve the efficiency.

A knowledge graph is a graph with entities of different types as nodes and various relations among them as edges. The construction of knowledge graphs in the past decades facilitates many applications, such as link prediction, web search analysis, question answering, and so on. Knowledge graph embedding aims to represent entities and relations in a large-scale knowledge graph as elements in a continuous vector space. Existing methods, for example, TransE, TransH, and TransR, learn the embedding representation by defining a global margin-based loss function over the data. However, the loss function is determined during experiments whose parameters are examined among a closed set of candidates. Moreover, embeddings over two knowledge graphs with different entities and relations share the same set of candidates, ignoring the locality of both graphs. This leads to the limited performance of embedding related applications. In this article, a locally adaptive translation method for knowledge graph embedding, called TransA, is proposed to find the loss function by adaptively determining its margin over different knowledge graphs. Then the convergence of TransA is verified from the aspect of its uniform stability. To make the embedding methods up-to-date when new vertices and edges are added into the knowledge graph, the incremental algorithm for TransA, called iTransA, is proposed by adaptively adjusting the optimal margin over time. Experiments on four benchmark data sets demonstrate the superiority of the proposed method, as compared to the state-of-the-art ones.

The study on generalization performance of ranking algorithms is one of the fundamental issues in ranking learning theory. Although several generalization bounds have been proposed based on different measures, the convergence rates of the existing bounds are usually at most O(√1/n), where n is the size of data set. In this paper, we derive novel generalization bounds for the regularized ranking in reproducing kernel Hilbert space via integral operator of kernel function. We prove that the rates of our bounds are much faster than (√1/n). Specifically, we first introduce a notion of local Rademacher complexity for ranking, called local ranking Rademacher complexity, which is used to measure the complexity of the space of loss functions of the ranking. Then, we use the local ranking Rademacher complexity to obtain a basic generalization bound. Finally, we establish the relationship between the local Rademacher complexity and the eigenvalues of integral operator, and further derive sharp generalization bounds of faster convergence rate.

Kernel learning is a fundamental problem both in recent research and application of kernel methods. Existing kernel learning methods commonly use some measures of generalization errors to learn the optimal kernel in a convex (or conic) combination of prescribed basic kernels. However, the generalization bounds derived by these measures usually have slow convergence rates, and the basic kernels are finite and should be specified in advance. In this paper, we propose a new kernel learning method based on a novel measure of generalization error, called principal eigenvalue proportion (PEP), which can learn the optimal kernel with sharp generalization bounds over the convex hull of a possibly infinite set of basic kernels. We first derive sharp generalization bounds based on the PEP measure. Then we design two kernel learning algorithms for finite kernels and infinite kernels respectively, in which the derived sharp generalization bounds are exploited to guarantee faster convergence rates, moreover, basic kernels can be learned automatically for infinite kernel learning instead of being prescribed in advance. Theoretical analysis and empirical results demonstrate that the proposed kernel learning method outperforms the state-of-the-art kernel learning methods.

Kernel selection is a fundamental problem of kernel methods. Existing measures for kernel selection either provide less theoretical guarantee or have high computational complexity. In this paper, we propose a novel kernel selection criterion based on a newly defined spectral measure of a kernel matrix, with sound theoretical foundation and high computational efficiency. We first show that the spectral measure can be used to derive generalization bounds for some kernel-based algorithms. By minimizing the derived generalization bounds, we propose the kernel selection criterion with spectral measure. Moreover, we demonstrate that the popular minimum graph cut and maximum mean discrepancy are two special cases of the proposed criterion. Experimental results on lots of data sets show that our proposed criterion can not only give the comparable results as the state-of-the-art criterion, but also significantly improve the efficiency.

Knowledge graph embedding aims to represent entities and relations in a large-scale knowledge graph as elements in a continuous vector space. Existing methods, e.g., TransE and TransH, learn embedding representation by defining a global margin-based loss function over the data. However, the optimal loss function is determined during experiments whose parameters are examined among a closed set of candidates. Moreover, embeddings over two knowledge graphs with different entities and relations share the same set of candidate loss functions, ignoring the locality of both graphs. This leads to the limited performance of embedding related applications. In this paper, we propose a locally adaptive translation method for knowledge graph embedding, called TransA, to find the optimal loss function by adaptively determining its margin over different knowledge graphs. Experiments on two benchmark data sets demonstrate the superiority of the proposed method, as compared to the-state-of-the-art ones.

Populating a knowledge base with new entity mentions extracted from unstructured text can help enhance its coverage and freshness. It naturally consists of two subtasks, namely, fine-grained entity classification and entity linking. Existing studies often focus on one of these two subtasks and they usually populate entity mentions in the same text by implicitly assuming that they are independent. However, these entity mentions are often semantically related to each other and it would be better to populate them into the knowledge base collectively. For solving these problems, in this paper we propose an interdependence graph based and unified collective inference approach, called CIIGA, to populating a knowledge base with collective entities, which can jointly determine the proper locations of all entity mentions in the same text by exploiting their interdependence relationships. Experimental results show that this approach can achieve significant accuracy improvement, as compared to the baseline approach, APOLLO, on the task of knowledge base population with multiple entities.