The correctness of knowledge graphs is crucial for their successful application. However, knowledge graphs often contain errors, making knowledge cleaning, i.e., the detection and correction of errors, essential. This survey presents a structured review of approaches for cleaning assertion errors in knowledge graphs based on the Resource Description Framework (RDF) data model. Its main contribution is a taxonomy that classifies cleaning approaches for the three most common assertion types: instance, property-value, and equality assertions. The taxonomy organizes approaches into 10 cleaning techniques and associates them with 6 background knowledge sources. Background knowledge may originate from the knowledge graph itself, including graph structure and contextual triples and values, or from external sources such as curated data, formal specifications, human expertise, and structured or unstructured sources. Together with its three facets and evaluation measures, the taxonomy provides a framework for understanding and comparing cleaning approaches based on knowledge graph characteristics, the available background knowledge, the targeted assertion type, and the cleaning process. The comparative analysis reveals that the diversity of knowledge graph cleaning approaches reflects the task’s inherent complexity, with different techniques leveraging complementary background knowledge sources. It also identifies open research challenges, including uneven support across assertion types and cleaning processes. Together, these insights provide practitioners with a systematic basis for selecting suitable cleaning approaches and researchers with a structured foundation for future work.
Knowledge graphs (KGs) have emerged as foundational infrastructures in modern natural language processing (NLP) and recommender systems. However, long-tailed relational distributions and structural sparsity pose significant challenges to few-shot knowledge graph completion (FKGC). Under extremely low-resource scenarios, conventional neural architectures are prone to overfitting, leading to suboptimal generalization capabilities. To address these issues, we propose RANS-KAN-DRIA, a dynamic meta-learning framework incorporating three core components: (1) a Relation-Aware Neighbor Selector (RANS), which adaptively integrates relation-relevant neighborhood context with support instances to strengthen relational reasoning; (2) a lightweight KAN-based Relational Meta-Learner, which operates on low-dimensional entity-pair representations from the few-shot support set and constructs nonlinear and topology-aware relational features through Kolmogorov–Arnold decomposition; and (3) a Diffusion-Regularized Inner-loop Adapter (DRIA), which enables rapid parameter adaptation via gradient updates and stabilizes model performance using diffusion-generated hard negatives. Experimental results on two benchmark datasets demonstrate that RANS-KAN-DRIA achieves strong performance in few-shot scenarios. The proposed framework contributes a task-specific integration of relation-aware neighborhood selection, lightweight KAN-based relational meta-learning, and diffusion-regularized inner-loop adaptation for robust FKGC under extreme sparsity.
While existing knowledge bases maintain factual world knowledge, we view the social network of Twitter (now, X) as a complementary source of social world knowledge. Treating popular Twitter accounts as entities of general interest, our aim is to elicit useful semantic information about these entities that may be used by downstream applications. We first address the task of entity typing, such as determining whether an account belongs to a politician or a musical artist. To obtain labeled data, we align a subset of social entities with DBpedia and Wikidata, yielding an aligned dataset of over 20K Twitter entities annotated with 136 fine-grained semantic types. In learning models of type prediction, we process network and content-based evidence as entity embeddings, where we finetune the textual encoder on the semantic typing task. Our best classifier yields weighted F1 performance of 0.68 on set-aside labeled examples; manual inspection indicates that many apparent mismatches correspond to semantically plausible alternative types. Applying classification at large-scale, we obtain semantic embeddings and types for as many as 200K social entities, many of which are not covered by existing factual knowledge bases. Our study offers insights into the distribution of entity types across the ‘Twitter sphere’. Additionally, we show that content-based embeddings encode fine-grained entity semantics, as demonstrated on the key task of entity similarity assessment. We believe that the inferred semantic entity types and embeddings can serve various applications, such as social conversation and recommendation tasks using LLM agents, which involve social entities. We make all artifacts of this work available to the research community.
The Internet of Things enables continuous acquisition of physiological data through wearable and environmental sensors, creating new opportunities for early risk detection and timely clinical intervention. However, most IoT-based healthcare monitoring systems lack real-time semantic interpretation, limiting the transformation of raw sensor measurements into clinically meaningful and actionable information.This paper proposes an ontology-driven real-time architecture for semantic interpretation of IoT sensor data in healthcare monitoring scenarios. The system integrates streaming data ingestion, in-stream semantic classification, composite health condition inference, and scalable persistence within a unified pipeline. Physiological parameters such as heart rate, blood oxygen saturation, respiratory rate, heart rate variability, blood pressure, sleep duration, and physical activity are classified using guideline-based thresholds, while higher-level composite conditions are inferred from co-occurring semantic statuses.The approach is experimentally validated using a hybrid dataset composed of real-world wearable data and simulated high-throughput sensor streams. Results show that the complete pipeline processes 1000 observations with an average end-to-end latency of approximately 195 ms, enabling near real-time and deterministic semantic interpretation and visualization. A semantic correctness analysis based on predefined clinical scenarios confirms consistent and reproducible mapping between physiological measurements and ontology-defined health statuses.These findings demonstrate that ontology-based semantic interpretation can be operationalized within real-time IoT healthcare pipelines, significantly improving interpretability, responsiveness, and clinical relevance. The proposed architecture provides a practical and scalable foundation for semantically enriched patient health monitoring and supports future integration with standardized clinical data ecosystems.
The rapid growth of open and structured RDF data on the Web has promoted the development of dataset search as an important research topic. The core function of existing systems is ad hoc dataset retrieval (AHDR) based on the metadata of datasets, which contains limited information and often suffers from quality issues. To overcome the limitations, in this article, we systematically investigate content-based AHDR to exploit the actual RDF data in datasets. We address three main tasks of content-based AHDR with novel methods for handling the large size and complex structure of RDF data to facilitate dataset retrieval, deduplication, and snippet extraction. These methods are integrated into an online and open-source prototype called Caddie. The effectiveness and practicability of its components are evaluated on a public test collection and by a user study.
