Training data

Data sources

There are three sources of data which are used to train the model, each with their own advantages and disadvantages.

New York Times

The New York Times released a dataset of labelled ingredients in their Ingredient Phrase Tagger repository, which had the same goal as this.

• The dataset is has each sentence labelled, but the labelling is inconsistent.

• The dataset primarily uses imperial/US customary units

• The dataset is large, roughly 178,000 entries

Cookstr

The Cookstr dataset is derived from 7,918 recipes scraped from cookstr.com (no longer available) between 2017-06 and 2017-07. The scraped data can be found at https://archive.org/details/recipes-en-201706.

• The dataset is unlabelled and will need labelling manually.

• The dataset primarily uses imperial/US customary units, although many ingredients give the quantity in multiple units

• The dataset is medium sized, roughly 40,000 entries

BBC Food

The Cookstr dataset is derived from 10,599 recipes scraped from bbc.co.uk/food between 2017-06 and 2017-07. The scraped data can be found at https://archive.org/details/recipes-en-201706.

• The dataset is unlabelled and will need labelling manually.

• The dataset primarily uses metric units, although many ingredients give the quantity in multiple units

• The dataset is medium sized, roughly 63,000 entries

The three datasets have different advantages and disadvantages, therefore combining the two should yield an improvement over using any on their own.

Labelling the data

Note

The details described in this section also apply to how the labelling was performed for the Cookstr and BBC Food datasets.

The New York Times dataset has gone through, and continues to go through, the very manual process of labelling the training data. This process is there to ensure that the labels assigned to each token in each ingredient sentence are correct and consistent across the dataset. In general, the idea is to avoid modifying the input sentence and only correct the labels for each, although entries have been removed where there is too much missing information or the entry is not actually an ingredient sentence (a few recipe instructions have been found mixed into the data).

The model is currently trained using the first 30,000 entries of the New York Times dataset, so the labelling efforts have primarily been focussed on that subset.

Tip

The impact of the consistent labelling can be seen by training the model using the full New York Times dataset, where the majority of the data has not been consistently labelled. The model performance drops significantly.

The following operations were done to clean up the labelling (note that this is not exhaustive, the git history for the dataset will give the full details).

• Convert all numbers in the labels to decimal

  This includes numbers represented by fractions in the input e.g. 1 1/2 becomes 1.5

• Convert all ranges to a standard format of X-Y

  This includes ranges represented textually, e.g. 1 to 2, 3 or 4 become 1-2, 3-4 respectively

• Entries where the quantities and units were originally consolidated should be unconsolidated

  There were many examples where the input would say

  1/2 cup, plus 1 tablespoon …

  with the quantity set as “9” and the unit “tablespoon”. The model will not do maths for us, nor will it understand have to convert between units. In this example, the correct labelling is a quantity of “0.5”, a unit of “cup”, and a comment of “plus 1 tablespoon”.

• Adjectives that are a fundamental part of the ingredient identity should be part of the name

  This was mostly an inconsistency across the data, for example if the entry contained “red onion”, sometimes this was labelled with a name of “red onion” and sometimes with a name of “onion” and a comment of “red”.

  Three general rules were applied:

  1. If the adjective changes the ingredient in a way that the chef cannot, it should be part of the name.

  2. If the adjective changes the item you would purchase in a shop, it should be part of the name.

  3. If the adjectve changes the item in a way that the chef would not expect to do as part of the recipe, it should be part of the name.

  It is recognised that this can be subjective. Universal correctness is not the main goal of this, only consistency.

  Examples of this:

  • red/white/yellow/green/Spanish onion

  • granulated/brown/confectioners’ sugar

  • soy/coconut/skim/whole milk

  • ground spices

  • extra-virgin olive oil

  • fresh x/y/z

  • ice water

  • cooked chicken

• All units should be made singular

  This is to reduce the amount the model needs to learn. “teaspoon” and “teaspoons” are fundamentally the same unit, but because they are different words, the model could learn different associations.

• Where alternative ingredients are given in the sentence, these should be part of the name if the alternative is in the same quantity, or the comment if it is a different quantity.

  For example:

  • 3 tablespoons butter or olive oil, or a mixture should have the name as butter or olive oil

  however

  • 4 shoots spring shallots or 4 shallots, minced should have the name as spring shallots and the comment as or 4 shallots, minced because there are different quantities of spring shallots to shallots.

Warning

The labelling processing is very manual and as such has not been completed on all of the available data. The labelling has been completed for the following subsets of the datasets:

• The first 30,000 sentences of the New York Times dataset

• The first 15,000 sentences of the Cookstr dataset

• The first 15,000 sentences of the BBC Food dataset

Data storage

The labelled training data is stored in an sqlite3 database at train/data/training.sqlite3. The database contains a single table, eng, with the following fields:

Field	Description
id	Unique ID for the sentence
source	The source dataset the sentence is from
sentence	The ingredient sentence
tokens	List of tokens from the sentence
labels	List of token labels

It is the data in this database that is used to train the models.

CSV files of the full datasets are in the train/data/<dataset> directories. These csv files contain the full set of ingredient sentences, including those not properly labelled. The csv files are kept aligned with the database using the following command.

$ python train/data/db_to_csv.py