KG4FUN: A benchmark for Knowledge Graph Funneling

KG4FUN KG4FUN is a general-purpose pipeline for extracting datasets to tackle the Knowledge Graph Funneling (KGF) task. It can be applied to any input Linked Opena Data (LOD) source to generate a supervised dataset for KGF. In this work, we instantiate the pipeline using the Wikidata Thematic Subgraph Selection Dataset. Available Datasets Dataset VI EI RI VS ES RS VO EO RI Dataset 1 304982 299193 46295 5030 58868 2 233013 249186 29434 Dataset 2 82044 59393 6568 1279 8606 2 72915 54920 4303 Each dataset folder contains: Node Artifacts File Description nodes.json All nodes in KG with their supervision label (cls_idx). A value of -1 means the node is discarded; any other value maps the node to a target node class. node_types.json Node type definitions, each grouping a seed node with its positive and negative examples. Used to derive the schema node types in SG. node_info.json Human-readable metadata (label, comment, description) for each node in nodes.json. node_type_info.json Human-readable metadata for schema-level nodes (seeds, positives, negatives). node_types/summarization/qwen3-8b/node_type_info.json LLM-generated labels and descriptions for each node type. Edge Artifacts File Description edges/partition/part_idx.json Partitioned edge storage for scalable memory management. Each entry encodes an edge with its predicted type, head/tail node QIDs, and supervision signal. edges/edge_info.json Human-readable metadata (label, comment, description) for each relation type appearing in KG. edges/edge_types.json Supervision signals for each detected edge type, including head/tail node class indices and target label. edges/connected_components.json Connected components of the relational graph, encoding groupings of semantically related relations used to derive abstract edge types. edges/edge_component_mapping.json Compact mapping from each edge type to its corresponding connected component ID. edge_types/summarization/qwen3-8b/edge_type_info.json LLM-generated labels and descriptions for each abstract edge type.