Evaluation Dataset for TEXTER: A Hybrid LLM Approach for Constructing Extended Entity-Relationship Models from Natural Language

File formats and encoding. All text and JSON files are UTF-8 encoded. The input texts (verbal_descriptions/NN.txt) are plain text. The reference models (reference_models/NN.txt) and the TEXTER outputs (results_hybrid/NN/*.txt) are JSON, despite the .txt extension. The classification and metrics are in an Excel workbook (results CPCIM.xlsx). No special software is required: a text editor, a JSON viewer, and a spreadsheet application are sufficient. Case correspondence. The case number NN (01 to 14) is consistent across all folders, so verbal_descriptions/NN.txt, reference_models/NN.txt, and results_hybrid/NN/ refer to the same case study. Model JSON schema (model_schemas/model.schema.json). Each model has: entities: name, attributes (name, type, isIdentifier), the weak flag isWeak with ownerEntity, and isAssociative; relationships: name, sourceEntity, destinationEntity, cardinality and inverseCardinality (each lowerBound and upperBound, where upperBound = -1 denotes many / N), and isIdentifying; generalizations: superEntity and subEntities.The TEXTER output files additionally include per-element provenance (evidence, confidence, reasoning); these fields document the pipeline and are not part of the reference models. suggested-model-result.schema.json is the schema of the extraction sub-runs. results_hybrid/NN contents. run_1.txt is the final scored model; run_1_consensus.txt is the majority-vote consensus model before review; run_1_suggested.txt is the permissive union of the three sub-runs; run_1_suggested_sub_1.txt, _sub_2.txt, _sub_3.txt are the three independent extraction sub-runs. Reproducing the metrics. The precision and recall in the paper are computed from the per-case counts in results CPCIM.xlsx. For each concept type, with C, PC, I, M counts, produced = C + PC + I and reference = C + PC + M; strict precision = C / produced, strict recall = C / reference, lenient precision = (C + PC) / produced, lenient recall = (C + PC) / reference. The headline figures average entities, attributes, and relationships. Generation settings. The models were produced with gpt-5.4-mini at temperature 0. LLM outputs are not bit-identical across invocations, so re-running the pipeline may produce small differences; the files provided are the exact models used in the paper.