Under the hood - tokenlist

library(textrecipes)
#> Loading required package: recipes
#> Loading required package: dplyr
#> 
#> Attaching package: 'dplyr'
#> The following objects are masked from 'package:stats':
#> 
#>     filter, lag
#> The following objects are masked from 'package:base':
#> 
#>     intersect, setdiff, setequal, union
#> 
#> Attaching package: 'recipes'
#> The following object is masked from 'package:stats':
#> 
#>     step

textrecipes has been using lists of character vectors to carry around the tokens. A simple S3 vector class has been implemented with the vctrs package to handle that list of tokens, henceforth to be known as a tokenlist.

If you are only using this package for preprocessing then you most likely won’t even notice that this change has happened. However if you are thinking of contributing to textrecipes then knowing about tokenlists will be essential.

A tokenlist is based around a simple list of character vectors, and has 3 attributes, lemma, pos and tokens.

tokens attribute

The tokens attribute is a vector of the unique tokens contained in the data list. This attribute is calculated automatically when using tokenlist(). If a function is applied to the tokenlist where the resulting unique tokens can be derived then new_tokenlist() can be used to create a tokenlist with known tokens attribute.

lemma and pos attributes

Both the lemma and pos attribute are used in the same way. They default to NULL but can be filled depending on which engine is being used in step_tokenize(). The attribute is a list of characters in the exact shape and size as the tokenlist and should have a one-to-one relationship.

If a specific element is removed from the tokenlist then the corresponding element in lemma and pos should be removed.