DoubleShapExpresso Baseline and Resulting Shapes

Description This dataset accompanies the paper *"Cross-Syntax Shape Generation for Knowledge Graphs: A PairedBenchmark and Agentic LLM Workflow."* It contains **SHACL and ShEx shapes for DBpedia andYAGO 4.5**, produced by four families of methods, plus the ground-truth benchmark they are evaluatedagainst: 1. **Ground truth** — manually curated shapes produced via SPARQL profiling, two-round crowd annotation, and expert consolidation, released as paired SHACL/ShEx shapes for DBpedia (20 classes, in an `aligned` and a `complete` variant) and YAGO (37 classes). 2. **Statistical baselines** — shapes induced automatically by two systems: [sheXer](https://github.com/DaniFdezAlvarez/shexer) (frequency-based induction) and RDFShapeInduction (four ML classifiers — decision tree, gradient boosting, MLP, random forest — plus a best-performing "best model" selection for DBpedia).3. **LLM few-shot shapes** — shapes generated by five LLMs (GPT-4o mini, DeepSeek-V3, and Qwen3 at 8B/14B/32B) under three few-shot prompting strategies (`local`, `global`, `triples`), in both SHACL and ShEx syntax.4. **LLM agentic-workflow shapes** — shapes produced by an evidence-driven agentic pipeline that queries the KG predicate-by-predicate, then assembles and reviews the schema. Provided both without (`global_skills`) and with (`global_skills_review`) the review/repair step, together with the full per-class process trace (evidence gathered, chat history, review verdict, repair diff, structured output) for reproducibility and error analysis. The exact prompts sent to each LLM for the few-shot runs are included under `prompts/`. This is an extension of the DBpedia/YAGO benchmark introduced in the EMNLP 2025 paper["Schema Generation for Large Knowledge Graphs Using Large Language Models"](https://arxiv.org/abs/2506.04512),whose original Wikidata/WES ground-truth ShEx shapes are also included (`ground_truth/ShEx/wes/`) forreference. Note that the WES shapes are **ShEx-only** (no SHACL counterpart exists), so they are notpart of the paired DBpedia/YAGO benchmark and are not otherwise reused in this dataset. Dataset Structure ```cross-syntax-shape-generation-kg-dataset/├── ground_truth/│ ├── dbpedia.csv, yago.csv, wes.csv # class URI, label, dataset, instance count│ ├── SHACL/│ │ ├── dbpedia-aligned/ # 20 shapes — DBpedia ground truth, ShEx-aligned class selection│ │ ├── dbpedia-complete/ # 20 shapes — DBpedia ground truth, full property coverage│ │ └── yago/ # 37 shapes — YAGO ground truth│ └── ShEx/│ ├── dbpedia-aligned/ # 20 shapes, paired 1:1 with SHACL/dbpedia-aligned│ ├── yago/ # 37 shapes, paired 1:1 with SHACL/yago│ └── wes/ # 53 shapes — legacy Wikidata/WES ground truth (EMNLP 2025 paper);│ # ShEx only, not part of the DBpedia/YAGO paired benchmark│├── baselines/│ ├── shexer/│ │ ├── dbpedia/ # 20 shapes│ │ └── yago/ # 36 shapes│ └── rdfshapeinduction/│ ├── dbpedia/│ │ ├── best_model/ # 19 shapes — best classifier per class│ │ ├── decision_tree/ # 20 shapes│ │ ├── gradient_boosting/ # 20 shapes│ │ ├── mlp/ # 20 shapes│ │ └── random_forest/ # 20 shapes│ └── yago/│ ├── decision_tree/ # 35 shapes│ ├── gradient_boosting/ # 35 shapes│ ├── mlp/ # 35 shapes│ └── random_forest/ # 35 shapes│├── llm_fewshot/ # few-shot prompting, one call per class│ ├── SHACL/{global,local,triples}/{model}/{dbpedia,yago}/predicate_count/5/│ └── ShEx/{global,local,triples}/{model}/dbpedia/predicate_count/5/│ # {model} ∈ deepseek-chat, gpt-4o-mini, qwen3-8b, qwen3-14b, qwen3-32b│ # each leaf folder holds one {ClassName}.ttl or .shex shape plus the│ # structured .json record of the generation call│ # (ShEx few-shot runs exist for DBpedia only)│├── llm_agentic/ # evidence-driven agentic pipeline│ ├── dbpedia/│ │ ├── evidences/ # shared per-class predicate evidence, reused across models│ │ ├── SHACL/{global_skills,global_skills_review}/{model}/│ │ └── ShEx/{global_skills,global_skills_review}/{model}/│ │ # global_skills = skill-injected agent, no review/repair│ │ # global_skills_review = + review, repair and consolidation step│ │ # each leaf folder holds, per class: {Class}.ttl (or .shex),│ │ # .evidence.json, .chat_history.json, .review.json, .repair.json,│ │ # .structured_output.json, .trace.json│ └── yago/│ └── SHACL/{global_skills,global_skills_review}/{model}/│ # (agentic ShEx runs and evidence cache exist for DBpedia only)│└── prompts/ # exact prompts sent to each LLM for the few-shot runs ├── SHACL/{global,local,triples}/{dbpedia,yago}/predicate_count/5/ └── ShEx/{global,local,triples}/dbpedia/predicate_count/5/``` **Known gaps** (kept as-is for transparency, not filled in): `llm_fewshot` and `llm_agentic` modelcoverage is not perfectly uniform across classes (a handful of classes have no output for a givenmodel where generation failed, e.g. `qwen3-14b`/`qwen3-32b` on YAGO local few-shot); `prompts/ShEx` and`llm_fewshot/ShEx` only cover DBpedia; `llm_agentic` ShEx and the shared evidence cache only coverDBpedia; `baselines/rdfshapeinduction` has no YAGO `best_model` selection. File formats - `.ttl` — SHACL shapes, Turtle syntax.- `.shex` — ShEx shapes, ShEx compact syntax.- `.json` alongside a shape — structured metadata about that generation call (varies by method; see above for the agentic-workflow fields).- `.csv` — one row per KG class: `class_uri, class_label, dataset, count` (instance count in the KG).