Knowledge graphs (KGs) such as Wikidata store large collections of structured facts but remain inherently incomplete. Link prediction aims to identify missing relations between entities and plays a key role in keeping KGs up to date. However, predicted links are not always reliable and must be supported by clear evidence and efficiently validated by humans before they can be added to the KG. We present an interactive demonstration of our WikiRAG (a framework that combines automatic link prediction with retrieval augmented generation) designed for Wikidata link prediction and integrating evidence-based human-in-the-loop validation. Given a head entity and a relation, the system generates candidate tail entities using knowledge graph embeddings, retrieves relevant Wikipedia passages, and applies a large language model to assess each candidate based on the retrieved evidence. The interface enables users to inspect supporting passages, validate suggested links, and export confirmed triples in batch for upload to Wikidata. Our demonstration integrates automated link prediction with evidence-driven reasoning and human-in-the-loop validation, providing a practical workflow for reliable knowledge graph completion. The system can be seen in action at https://youtu.be/UlnaxCRxlp8, with a live version available at https://wikirag.com.

Incomplete knowledge graphs present a fundamental challenge for reliable multi-hop knowledge graph question answering (KGQA), causing reasoning failures when key factual triples are missing. While large language models (LLMs) offer strong reasoning capabilities for KGQA, they are prone to hallucinations and often assume complete knowledge graphs (KGs). This research identifies key bottlenecks in current LLM-KGQA pipelines caused by KG incompleteness. We propose targeted remedies, centered on integrating link prediction tools, to enhance performance in sparse KGs. We explore two main directions: (1) improving the robustness of LLM-KGQA methods under KG sparsity, and (2) leveraging advanced link prediction techniques to recover missing graph connections. Preliminary experiments on benchmark datasets demonstrate significant improvements in both answer accuracy and link prediction performance under simulated KG sparsity. These results bridge the gap between LLM-based reasoning and incomplete KGs, laying the foundation for more faithful and interpretable KGQA systems.

Knowledge Graph Question Answering (KGQA) is a task that aims to answer natural language queries by extracting facts from a knowledge graph. Current state-of-the-art techniques for KGQA rely on text-based information from graph entity and relations labels, as well as external textual corpora. By reasoning over multiple edges in the graph, these can accurately rank and return the most relevant entities. However, one of the limitations of these methods is that they cannot handle the inherent incompleteness of real-world knowledge graphs and may lead to inaccurate answers due to missing edges. To address this issue, recent advances in graph representation learning have led to the development of systems that can use link prediction techniques to handle missing edges probabilistically, allowing the system to reason with incomplete information. However, existing KGQA frameworks that use such techniques often depend on learning a transformation from the query representation to the graph embedding space, which requires access to a large training dataset. We present BeamQA, an approach that overcomes these limitations by combining a sequence-to-sequence prediction model with beam search execution in the embedding space. Our model uses a pre-trained large language model and synthetic question generation. Our experiments demonstrate the effectiveness of BeamQA when compared to other KGQA methods on two knowledge graph question-answering datasets.

Knowledge Graph Question Answering (KGQA) models enable users to acquire entity-based answers from a Knowledge Graph by asking natural language questions (NLQs) without the need to learn a specialized graph query language or knowing the underlying schema of the knowledge graph. This work investigates hyperbolic graph representation learning methods to effectively and efficiently represent knowledge base items and natural questions. Our system, HyperKGQA, proposes a technique that embeds the knowledge graph in a hyperbolic manifold, then learns an adaptive transformation of pre-trained sentence representations into the space of entities and relations. Finally, a post-processing step refines the ranking of the candidate answers by computing the relevance score of the set of relations and the question. An extensive set of experiments conducted on two datasets shows that our method outperforms the current state-of-the-art models when reasoning over sparse graphs to answer multi-hop questions.