Information Processing & Management最新文献

Person Re-identification (ReID) could search for the same pedestrian from non-overlapping cameras, which completes the pedestrian location and search purpose. However, the process contains much sensitive pedestrian information and raises serious privacy problems. Conventional methods mainly remove identity-related features from pedestrian images to alleviate the privacy issue. Unfortunately, these strategies cause pedestrian information loss and poor data utility. In the paper, we propose a novel Model-Aware Privacy-Preserving with Start Trigger (MPST) method, which not only prevents personal identity for third parties but also achieves accurate pedestrian location for authorized organizations. The core idea is that authorized organizations obtain the start trigger to activate the ReID model that has the ability to search for target pedestrians, while third parties (i.e., hackers) cannot employ the ReID model to complete the pedestrian matching task without the start trigger. To this end, we develop a universal adversarial algorithm to learn an ingenious start trigger for the person ReID system. Afterwards, we further design a model-aware training strategy to facilitate our deep model to perceive issued instructions by synthetically utilizing the start trigger and original pedestrian samples. As a result, we successfully install an activate button to change the ReID model state for deciding whether the deep model has the ability to search for pedestrians or not. Abundant experiments demonstrate that the proposed MPST is effective for pedestrian identity anonymization. Our study achieves superior performance for authorized organizations and completes the privacy protection goal.

Medical Visual Question Answering (MedVQA) is designed to answer natural language questions related to medical images. Existing methods largely adopting the cross-modal pre-training and fine-tuning paradigm, face limitations in accuracy due to data scarcity and insufficient incorporation of extensive medical knowledge. Drawing inspiration from the Knowledge-Based Visual Question Answering (KB-VQA) domain, which leverages Large Language Models (LLMs) and external knowledge bases, we introduce the Candidate-Heuristic In-Context Learning (CH-ICL) framework, a novel approach that leverages LLMs augmented with external knowledge to directly enhance existing MedVQA models. Specifically, we collect a pathology terminology dictionary from a public digital pathology library as an external knowledge base and use it to train a knowledge scope discriminator, which helps identify the knowledge scope required to answer a question. Then, we employ existing MedVQA models to provide reliable answer candidates along with their confidence scores. Finally, the knowledge scope and candidates, combined with retrieved in-context exemplars, are aggregated into prompts for heuristically guiding LLMs in answer generation. Experimental results on the PathVQA, VQA-RAD, and SLAKE public benchmarks show state-of-the-art performance, with improvements of 1.91%, 1.88%, and 2.17% respectively over the baseline. Code and dataset are available at https://github.com/ecoxial2007/CH-ICL.

In weakly supervised learning, labeling rules can automatically label data to train models. However, due to insufficient prior knowledge, rule discovery often suffers from semantic drift. Since misclassified rules are generated from wrongly matched sentences, the sentences matched by rules shift from the target labels to other labels. It is worth noting that rules do not exist in isolation. The multi-dimensional semantic associations among rules can impose semantic constraints for rule generation, as well as enrich the semantic information of rules for rule matching. Therefore, we propose a Knowledge-Graph-based RulE Discovery method (KGRED), which can leverage the multi-dimensional semantic associations among rules to alleviate semantic drift in rule discovery. Specifically, to decrease misclassified rules, we design a label-aware rule generation approach to attentively propagate prior knowledge from seed rules to candidate rules based on rule KG. To reduce wrongly-matched sentences, we present a cross-attention-based semantic matching mechanism to refine the semantic information of sentences while enriching that of rules. Moreover, we propose an inconsistency-directed active learning strategy to verify rules that perform inconsistently in rule generation and matching. Experiments on two public datasets prove that KGRED can achieve at least 5.1 % gain in F1 score compared to state-of-the-art methods.

Existing benchmarks for Large Language Models (LLMs) mostly focus on general or specific domain capabilities, overlooking structured output capabilities. We introduce SoEval, a benchmark for assessing LLMs’ ability to generate structured outputs like JSON, XML, and lists. SoEval contains 3.7K entries in Chinese and English, covering 13 types of structured output tasks across 20 subjects. In experiments, we found that while current mainstream LLMs have deficiencies in structured output, GPT-4 outperforms them in this aspect. GPT-4 achieved an average score of 0.4 on SoEval, representing a 24% enhancement over the next best-performing model. At the same time, the performance of current mainstream models on English tasks is also better than on Chinese tasks. We also report the performance of mainstream large models on different structured output types and task subjects. The benchmark construction code and SoEval dataset are open-sourced at https://github.com/MoranCoder95/SoEval.

Measuring semantic similarity between two pieces of text is a widely known problem in Natural language processing(NLP). It has many applications, such as finding similar medical notes of patients to accelerate the diagnosis process, plagiarism detection, and document clustering. Most state-of-the-art models are based on machine/deep learning and lack sufficient explanations for their results, limiting their adoption in critical domains like healthcare. This paper presents a hybrid framework SUMEX (Semantic textUal siMilarity and EXplanation generation) that uniquely combines ontology with a state-of-the-art embedding-based model for semantic textual similarity. The primary strength of the framework is that it explains its results in human-understandable natural language, which is vital in critical domains such as healthcare. Experiments have been conducted on two datasets of clinical notes using four embeddings: ScispaCy, BioWord2Vec, ClinicalBERT, and a customized Word2Vec trained on clinical notes. The SUMEX framework outperforms the embedding-based model on the benchmark datasets of ClinicalSTS by improving average precision scores by 7 % and reducing the false-positives-rate by 23 %. On the Patients Similarity Dataset, the average top-five and top-three precision scores were improved by 14% and 10%, respectively, using SUMEX. The SUMEX also generates explanations for its results in natural language. The domain experts evaluated the quality of the explanations. The results show that the generated explanations are of significantly good quality, with a score of 90 % and 93 % for measures of Completeness and Correctness, respectively. In addition, ChatGPT was also used for similarity score and generating explanations. The experiments show that the SUMEX framework performed better than the ChatGPT.

Growing multiple-affiliation collaboration in Artificial Intelligence (AI) can help solve complex integrated problems, but will it trigger the disruption in AI? Scholars have discussed the related topics in other fields. However, these studies did not specifically target the field of AI and primarily relied on correlation methods, which may not lead to a causal conclusion. Analyzing around 0.6 million AI collaborative papers between 1950 and 2019 with 872,727 authors and 9,258 affiliations, this study tests the causal effect of multiple-affiliation collaboration on the disruption in AI by using descriptive analysis and a causal inference method, i.e., the Propensity Score Matching (PSM). We propose an improved affiliation diversity indicator to measure the distribution of affiliation differences in multiple-affiliation collaboration by taking disparity into account. Our results show that affiliation diversity exerts a negative causal effect on the disruption of papers in AI: (a) The average level of AI papers with diverse affiliation types or affiliation countries of authors is less disruptive than those with a single type or country. (b) Affiliation diversity will causally reduce the disruption of papers in AI by 2.006%∼5.891%. That indicates that AI papers with high affiliation diversity are significantly less disruptive, ranging from 2.006% to 5.891%, compared to those without. We cross-validate the findings by using five comparison experiments and five other matching methods. This study provides a comprehensive understanding of multiple-affiliation collaboration on AI disruption.

Fake news has become a significant concern in recent times, particularly during the COVID-19 pandemic, as spreading false information can pose significant public health risks. Although many models have been suggested to detect fake news, they are often limited in their ability to extend to new emerging domains since they are designed for a single domain. Previous studies on multidomain fake news detection have focused on developing models that can perform well on multiple domains, but they often lack the ability to generalize to new unseen domains, which limits their effectiveness. To overcome this limitation, in this paper, we propose the Entity-centric Multi-domain Transformer (EMT) model. EMT uses entities in the news as key components in learning domain-invariant and domain-specific news representations, which addresses the challenges of domain shift and incomplete domain labeling in multidomain fake news detection. It incorporates entity background information from external knowledge sources to enhance fine-grained news domain representation. EMT consists of a Domain-Invariant (DI) encoder, a Domain-Specific (DS) encoder, and a Cross-Domain Transformer (CT) that facilitates investigation of domain relationships and knowledge interaction with input news, enabling effective generalization. We evaluate the EMT's performance in multi-domain fake news detection across three settings: supervised multi-domain, zero-shot setting on new unseen domain, and limited samples from new domain. EMT demonstrates greater stability than state-of-the-art models when dealing with domain changes and varying training data. Specifically, in the zero-shot setting on new unseen domains, EMT achieves a good F1 score of approximately 72 %. The results highlight the effectiveness of EMT's entity-centric approach and its potential for real-world applications, as it demonstrates the ability to adapt to various training settings and outperform existing models in handling limited label data and adapting to previously unseen domains.

In classification tasks with noisy labels, eliminating the interference of noisy label samples in the dataset is the key to improving network performance. However, the distribution between some noise and clean samples is overlapping, so it is a great challenge to distinguish them. Clean label samples within the overlapping region often contain highly representative feature information, which is extremely valuable for deep learning. We propose a new method called twin binary classification-mixed input (TBC-MI) to tackle this challenge. Specifically, TBC-MI utilizes the twin classification network to partition the sample and converts the original complex classification problem into a binary classification. It filters clean label samples from hard label regions using a simple multilayer binary classification network. TBC-MI uses noise from the dataset in the dividing process to better reflect real-world scenarios. After maximizing the clean label samples, TBC-MI adopts a hybrid online and offline input method to expand the subsequent input form of the samples. The proposed method is verified on CIFAR-10 and CIFAR-100 datasets containing artificially synthesized noise and Clothing1M ANIMAL-10N, CIFAR-10N, and CHAOYANG datasets with real-world noise. Extensive experiments show that our method achieves the best test accuracy on most datasets, with the best improvement of 2% compared to previous learning methods with noisy labels.

Knowledge graph representation learning entails transforming entities and relationships within a knowledge graph into vectors to enhance downstream tasks. The rise of pre-trained language models has recently promoted text-based approaches for knowledge graph representation learning. However, these methods often need more structural information on knowledge graphs, prompting the challenge of integrating graph structure knowledge into text-based methodologies. To tackle this issue, we introduce a text-enhanced model with local structure (TEGS) that embeds local graph structure details from the knowledge graph into the text encoder. TEGS integrates k-hop neighbor entity information into the text encoder and employs a decoupled attention mechanism to blend relative position encoding and text semantics. This strategy augments learnable content through graph structure information and mitigates the impact of semantic ambiguity via the decoupled attention mechanism. Experimental findings demonstrate TEGS’s effectiveness at fusing graph structure information, resulting in state-of-the-art performance across three datasets in link prediction tasks. In terms of Hit@1, when compared to the previous text-based models, our model demonstrated improvements of 2.1% on WN18RR, 2.4% on FB15k-237, and 2.7% on the NELL-One dataset. Our code is made publicly available on https://github.com/HubuKG/TEGS.

Structured knowledge bases are critical for the interpretability of AI techniques. RDF KBs, which are the dominant representation of structured knowledge, are expanding extremely fast to increase their knowledge coverage, enhancing the capability of knowledge reasoning while bringing heavy burdens to downstream applications. Recent studies employ semantic compression to detect and remove knowledge redundancies via semantic models and use the induced model for further applications, such as knowledge completion and error detection. However, semantic models that are sufficiently expressive for semantic compression cannot be efficiently induced, especially for large-scale KBs, due to the hardness of logic induction. In this article, we present estimation-based optimizations for the semantic compression of RDF KBs from the perspectives of input and intermediate data involved in the induction of first-order logic rules. The negative sampling technique selects a representative subset of all negative tuples with respect to the closed-world assumption, reducing the cost of evaluating the quality of a logic rule used for knowledge inference. The number of logic inference operations used during a compression procedure is reduced by a statistical estimation technique that prunes logic rules of low quality. The evaluation results show that the two techniques are feasible for the purpose of semantic compression and accelerate the compression algorithm by up to 47x compared to the state-of-the-art system.