Ingredients Extraction and Analysis

From Open Food Facts wiki
Revision as of 14:16, 10 September 2019 by Stephane (talk | contribs)
Jump to navigation Jump to search

This page describes how the ingredients list extraction and ingredients analysis is done on Open Food Facts and points to ressources that could be used to improve it.

Objectives

Ingredients list extraction

The goal of ingredients list extraction is to get the text of the ingredients list of each product in exactly the same form as it appears on the product package and label.

Ingredients analysis

Once the ingredients list is available, we need to analyze it to recognize the actual ingredients and indications of quantity, labels, processing etc. There is a lot of variety in how ingredients are listed on products, with many different synonyms, ways to indicate sub-ingredients etc.

The analysis needs to work for ingredients lists written in many different languages.

The output is structured data that links to our multilingual ingredients taxonomy.

Why it's important

Ingredients list extraction and analysis is necessary for many tasks:

  • Detecting food additives and allergens
  • Determining the degree of processing of food products (NOVA classification)
  • Identifying food products that can be or cannot be eaten by people following specific diets:
    • Vegetarian, vegan
    • Casher, Halal
    • Palm oil
  • Estimating the carbon impact of ingredients
  • Translating ingredients lists

Ingredients list extraction

Data sources for ingredients lists

The possible input sources for the ingredients lists are:

  • Ingredients lists typed in by users
    • Time consuming and not pleasant task, especially on mobile
    • Can contain typos, but usually typed ingredients lists are very close to what is written on the product
  • Ingredients lists given by manufacturers in data files
    • Usually of very good quality, but depending on manufacturers, can contain typos and sometimes formatting errors
  • Photos of product labels
    • Photo quality varies a lot
      • Some products are hard to photograph (round cans and bottles, foil bags etc.)
      • Sometimes very poor lighting, orientation, camera, focus etc.
  • High resolution images or PDFs of the printable package
    • Available for a few producers
    • Perfect quality
    • Needs cropping and/or rotation to select the ingredients list

Steps for ingredients lists

Picture taking

  • Taken with mobile app, uploaded to OFF server

Ingredients list cropping

  • Done on mobile app just after picture taking
    • Cropping may be very inaccurate
  • Or done on web site at a later time, possibly by another user
    • Cropping slightly easier than on mobile

OCR

  • Launched after cropping, done by the server which calls Google Cloud Vision
  • Cloud Vision returns a JSON object which is stored on the server

Ingredients list cutting

  • The image sent by the OCR can also contain other text content
    • Things that are not ingredients
    • Ingredients in other languages
    • The word "Ingredients:"
  • Current solution
    • Hardcoded regular expressions
  • Other possible solutions
    • Language identification to remove other languages
  • Metrics
    • False negatives (words before or after the ingredients list that should have been removed)
    • False positives (words that were removed but are part of the ingredients and should have been kept)
      • It is very important to have as few false positives as possible as it destructs data
  • Test and training sets
    • Only a few adhoc tests run during builds
    • Test sets needs to be created

Validation and/or correction by users

  • Current solution:
    • Users on the app or the web site are shown the OCR result
    • OCR result is not applied if not validated by the user
    • but users tend to validate lists without changes even if there are errors, especially on mobile
  • Other possible solutions
    • Use the result of ingredient analysis to show users ingredients that were not recognized
    • Show spell suggestions

Spell correction

  • Current solution:
    • Currently only done during ingredients analysis, not during ingredients extraction
    • Very simple (and slow) implementation of Peter Norvig algorithm
  • Other possible solutions
    • Spell checkers trained on ingredients
      • Elastic search spellchecker
      • Simspell
  • Metrics
    • Recall and precision
  • Test and training sets
    • Language models can be build with lists of ingredients from OFF
      • e.g. including only ingredients lists from producers, or lists for which we have a very high ingredients recognition rate
    • Test sets need to be created
      • Run spellcheckers on actual ingredients lists from OFF, review corrections


Ingredients analysis

Steps for ingredients analysis

Ingredients pre-parsing

  • The ingredients list is transformed to make parsing easier
    • Remove / normalize strange characters
    • De-abbreviate abbreviations
    • Split enumerations
      • e.g. "Vitamins A, B et C" -> Vitamine A, Vitamine B, Vitamine C
    • Additives E-numbers normalization (E330, e330, e-330, INS 330, SIN330 etc.)
    • Additives classes + additive splits
      • e.g. "Colour caramel" -> Colour: Caramel
    • Split some "A of B, C and D" (but not all...)
      • e.g. "Huile de palme, colza et tournesol" -> Huile de palme, huile de colza, huile de tournesol
  • Current solution
    • Perl code and regular expressions
      • lib/ProductOpener/Ingredients.pm - preparse_ingredients_text()