Problem
Evotor makes online cash registers, and every receipt line holds a barcode and the item name exactly as a store employee typed it. One product has a name per store, with typos, abbreviations and gaps. To count sales of a brand in such data you would have to list every spelling of it. The catalog (barcode as ID, with category, brand, weight, flavor, manufacturer) used to be filled in by hand, which is slow.
Example from the article: barcode 460134567893, receipt text "chips chipi chapa sour cream and onion 70 g ridged" → category chips, brand Chipi Chapa, weight 70 g, flavor "sour cream and onion", ridged: yes.
Solution in brief
Each property gets its own extractor: a classifier for category, a fine-tuned generative T5 model (pseudo-NER) for brand and flavor, regular expressions for weight and similar properties, and a barcode-prefix lookup for manufacturer. Names of one barcode are combined, brand spellings are merged by Levenshtein similarity after transliteration, and annotators validate predictions.
Questions this case answers
- How to build a clean entity catalog keyed by an identifier from many noisy free-text names written by different sources
- How to extract open-vocabulary attributes from short noisy text with a generative model and closed-set attributes with rules, while cutting manual labeling
- How to extract brand, flavor and weight from noisy product names in receipts
- How to build a product catalog keyed by barcode from many store-specific item names
Tools and methods
- rubert-tiny classifier — category, with outlier removal and active learning against noisy labels
- FRED-T5-large-ods-ner-2023 — generates "type;brand" and, after fine-tuning, "type;brand;flavor"
- Levenshtein aggregation with transliteration — merges spellings of one brand
- spaCy, spacy-llm, pymorphy2, Claude — LLM-labeled NER data, flavor normalization
- Regular expressions — weight, fat content, alcohol strength and other category-specific properties
- Barcode prefix lookup — manufacturer
Google or a one-time email code. No password.