Knowledge Graph (KG) completion aims to infer new facts from existing knowledge. While recent efforts have explored leveraging large language models (LLMs) for this task, it remains unclear whether LLMs truly understand KG facts or how they utilize such knowledge in reasoning. In this work, we investigate these questions by proposing ProbeKGC, a benchmark dataset that reformulates KG completion as multiple-choice question answering with systematically controlled option difficulties. Empirical results show that LLMs often produce inconsistent answers when the same question is presented with varying distractor difficulty, suggesting a reliance on shallow reasoning such as elimination rather than genuine knowledge recall. To better quantify model confidence and knowledge grasp, we introduce Normalized Knowledge Divergence (NKD), a novel metric that complements accuracy by capturing distributional confidence in answer selection. We further analyze the influence of selection biases on LLM predictions and highlight that LLMs do not always fully exploit their stored knowledge. Finally, we evaluate three enhancement strategies and provide insights into potential directions for improving KG completion. Resource Type: Benchmark and Evaluation License: GPL-3.0 License GitHub Repository: https://github.com/nju-websoft/ProbeKGC
The rising prevalence and complexity of online harms, particularly those disproportionately affecting women, demand urgent, interdisciplinary, and socio-technical responses. Despite increasing awareness and policy action, current responses remain fragmented across disciplinary silos, limiting the development of cohesive and effective interventions. This paper presents our efforts to co-create a comprehensive Gender-aware Ontology of Online Harms as a shared knowledge structure to bridge disciplinary perspectives and inform practice across sectors, including policing, law, behavioural science, and technology. Our ontology aims to capture the full spectrum of gender-based online harms, their sociotechnical enablers and inhibitors, and their manifestations in online contexts. The development process employs a co-creation approach grounded in collaborative ontology engineering and iterative stakeholder engagement. It has been driven by discipline-specific Personas and Competency Questions to ensure relevance, usability, and impact across diverse domains. We argue that this work represents a crucial step toward formalising a shared understanding of online harms to support policy reform, technological innovation, and survivor support.
In adaptive query processing (AQP), the query plan is adjusted based on actual execution conditions. AQP has proven effective in dynamic querying environments, such as knowledge graphs (KGs) on the web. The technique known as eddies enables tuple-wise adaptivity by dynamically reordering query operators at runtime. Eddies operate under a predefined symbolic routing policy, which determines the next operator to process each tuple. Although various routing policies have been proposed, their effectiveness varies across queries, and choosing a suboptimal policy can significantly degrade performance. To address this challenge, we propose a neuro-symbolic AQP approach that combines representation learning and supervised learning to predict the optimal routing policy for a given query. Experimental results on synthetic and real-world KGs demonstrate that our method achieves high precision in predicting optimal policies, is efficient to train and use at inference time, and generalizes well to queries with constants not seen during training.
Description Logic (DL) ontologies underpin many Semantic Web applications. Consequence-based reasoning, which integrates techniques from hypertableau and resolution, has proved effective for tasks such as consistency checking and classification in both lightweight and expressive DLs. However, existing reasoners often fall short when applied to large, complex ontologies commonly found in domains such as healthcare and industry. In this paper, we extend the state-of-the-art consequence-based reasoner Sequoia [13] to support parallel reasoning, improving its scalability by leveraging system architectures with multiple cores. We explore and evaluate two parallelisation strategies for consequence-based reasoners: message passing and thread pools, and demonstrate their application within the Sequoia reasoner. Our extensive empirical evaluation shows that thread pool-based implementations achieve superior performance and resource efficiency, offering up to 2.62x speedup over the baseline on hard ontologies. We also explore the effect of increasing the number of available cores or restricting the expressivity of the ontology in the performance of our implementations.
