GastroCorp: A Bilingual Gastronomic NER Corpus (Spanish–English)

GastroCorp is a bilingual (Spanish–English) gastronomic corpus for Named Entity Recognition (NER), constructed through human annotation by 71 university students at Universidad Tecnológica de Bolívar (Cartagena, Colombia) using Label Studio. What's new in v2 This version adds the complete-context corpus variant (corpus_with_o/), which reconstructs full token sequences including O-labeled context tokens from the original annotation export. This variant is required to reproduce the NER training pipeline described in the associated thesis. The v1 annotated files (HuggingFace JSONL, CoNLL, NMT pairs) are unchanged. Corpus statistics have also been corrected to reflect the final published counts after IOB2 validation and stratified splitting. Corpus Statistics 20,616 published sequences (15,818 menus · 4,798 recipes) 86,320 annotated entity spans across four types: DISH, BEVERAGE, INGREDIENT, BRAND Split: 70% train (14,428) / 15% dev (3,085) / 15% test (3,103) — stratified by entity type Inter-annotator agreement: Cohen's κ = 0.9031 (Almost perfect) · Pairwise F1 = 0.9727 (Brandsen et al., 2020) Annotation period: October–December 2025 Entity Types DISH: complete dish names in menus (e.g., "Lomo saltado", "Grilled salmon") BEVERAGE: drinks listed as final products (e.g., "Mojito", "Café americano") INGREDIENT: individual components in menu descriptions and recipes BRAND: commercial names and registered trademarks Each entity is annotated with semantic attributes (cooking method, preparation state, culinary origin, language, beverage type, etc.). Files menu_{train,dev,test}.jsonl — Menu annotations in HuggingFace Datasets format recipe_{train,dev,test}.jsonl — Recipe annotations in HuggingFace Datasets format All files follow the HuggingFace NER JSONL format with aligned token/tag sequences and full entity metadata. Labels use English names (DISH, BEVERAGE, INGREDIENT, BRAND) for international interoperability. Community Benchmark A shared task based on this corpus was hosted on Codabench (competition ID: 15173) as the GastroCorp NER Task 2026. The evaluation set (dev/test without labels) is available via the shared task platform. Data Sources Menus: provided by CLUVI (Colombia) under academic use agreement Recipes: RecipeNLG (Bień et al., 2020), HumbleIntelligence Food NER 1M (HuggingFace), TASTEset (Wróblewska et al., 2021) Context This corpus was developed as part of a Master's thesis in Engineering at Universidad Tecnológica de Bolívar, investigating the integration of NER with Neural Machine Translation (NMT) to improve translation quality of gastronomic texts between Spanish and English. License and Use Restrictions Released under CC BY-NC 4.0 (Attribution-NonCommercial). Academic and research use is permitted with proper attribution. Commercial use is not permitted. The menu data component is derived from a source provided under an academic-only agreement; this restriction applies to the full dataset.