✍️ Ingredients Spellcheck

Most ingredient lists are extracted from packaging images using Optical Character Recognition (OCR) techniques. However, due to factors such as poor image quality, product curvature, or light reflections, the OCR process may incorrectly recognize text characters.

These errors affect the Ingredients Extraction and Analysis process, impacting the overall quality of the database.

Recent advancements in Machine Learning, particularly with the development of Large Language Models, offer new opportunities to address this issue, where previous solutions have yielded limited success.

Picture of a product list of ingredients on its packaging.

⭐ Guidelines

The impact of the Ingredient-Spellcheck needs to be controlled to avoid altering contributions and/or add new errors to "healthy" ingredient names. The text needs to be kept as unchanged as possible. From the different types of errors observed across products, we came up with these spellcheck guidelines:

• Whitespaces between words and percentages shouldn't be corrected.
  • Example: "salt 0,2 %"
• If a percentage is ambiguous, it shouldn't be corrected.
  • Examples: "15 0%" ; "396"
• A whitespace involving a percentage shouldn't be modified, except if the digit is stuck in the previous word.
  • Example: "cheese1.9%" → "cheese 1.9%"
• Some ingredients are enclosed in '_', such as "_milk_" or "_cacahuetes_", to detect allergens. Should remain unchanged. However, in the case it is not an ingredient, such as "_Cacahuetes_ con cáscara tostado. _Trazas de frutos de cáscara_.", it needs to be modified into "_Cacahuetes_ con cáscara tostado. Trazas de frutos de cáscara.";
• Some percentages were badly parsed by the OCR. Since we cannot be sure about what is the right value, it is preferable to keep it as it is.
• In Romanian, the characters 'ş' (U+015F) and 'ţ' (U+0163) should be retrieved by the Spellcheck when necessary,
• Uppercase letters shouldn't be modified in order to add/remove accents.
  • Examples: "ECOSSE" ; "ÉCOSSE"
• Lowercase letters should be modified in order to add missing accents.
• '*' shouldn't be removed.
  • Example: "Schweinefleisch*"
• Whitespaces shouldn't been modified, except if two words are stuck to each other.
  • Example: "rizbrun" → "riz brun"
• Regarding uppercases and lowercases, since the spellcheck should modify at least as possible lists of ingredient, we don't modify uppercases or lowercases except for the following reasons:
  • after a period: "orange.trace de..." → "orange. Trace de..."
  • if it's a proper noun: "france" → "France"
• In French, the character "oe" or 'œ' should remain unchanged after correction. If it is missing, it should be replaced by default by 'œ'.
  • Examples: "œuf" ; "bœuf"
• Commas and other word separators (, -, .) should be added to distinct ingredients.
  • Example: "citric acid electrolytes (salt, magnesium and calcium chlorides, mono-potassion phosphate)" → "citric acid, electrolytes (salt, magnesium and calcium chlorides, mono-potassion phosphate)"
• We don't add a period or modify the existing punctuation at the end of the list of ingredients.
• A ':' should be added if it is missing between the category and name of an additive.
  • Example: "conservateur nitrite de sodium" → "conservateur: nitrite de sodium"