step_mutate with textrecipes tokenlists

user_15793856 0
Viewed 110

I'm doing NLP with the tidymodels framework, taking advantage of the textrecipes package, which has recipe steps for text preprocessing. Here, step_tokenize takes a character vector as input and returns a tokenlist object. Now, I want to perform spell checking on the new tokenized variable with a custom function for correct spelling, using functions from the hunspell package, but I get the following error (link to the spell check blog post):

Error: Problem with `mutate()` column `desc`.
i `desc = correct_spelling(desc)`.
x is.character(words) is not TRUE

Apparently, tokenlists don't parse easily to character vectors. I've noticed the existence of step_untokenize, but simply disolves the tokenlist by pasting and collapsing and that's not what I need.

REPREX

library(tidyverse)
library(tidymodels)
library(textrecipes)
library(hunspell)

product_descriptions <- tibble(
  desc = c("goood product", "not sou good", "vad produkt"),
  price = c(1000, 700, 250)
)

correct_spelling <- function(input) {
  output <- case_when(
    # check and (if required) correct spelling
    !hunspell_check(input, dictionary('en_US')) ~
      hunspell_suggest(input, dictionary('en_US')) %>%
      # get first suggestion, or NA if suggestions list is empty
      map(1, .default = NA) %>%
      unlist(),
    TRUE ~ input # if word is correct
  )
  # if input incorrectly spelled but no suggestions, return input word
  ifelse(is.na(output), input, output)
}

product_recipe <- recipe(desc ~ price, data = product_descriptions) %>% 
  step_tokenize(desc) %>% 
  step_mutate(desc = correct_spelling(desc))

product_recipe %>% prep()

WHAT I WANT, BUT WITHOUT RECIPES

product_descriptions %>% 
  unnest_tokens(word, desc) %>% 
  mutate(word = correct_spelling(word))
r-recipes spell-checking nlp r
2 Answers

There isn't a canonical way to do this using {textrecipes} yet. We need 2 things, a function that takes a vector of tokens and returns spell-checked tokens (you provided that) and a way to apply that function to each element of the tokenlist. For now, there isn't a general step that lets you do that, but you can cheat it by passing the function to custom_stemmer in step_stem(). Giving you the results you want

library(tidyverse)
library(tidymodels)
#> Registered S3 method overwritten by 'tune':
#>   method                   from   
#>   required_pkgs.model_spec parsnip
library(textrecipes)
library(hunspell)

product_descriptions <- tibble(
  desc = c("goood product", "not sou good", "vad produkt"),
  price = c(1000, 700, 250)
)

correct_spelling <- function(input) {
  output <- case_when(
    # check and (if required) correct spelling
    !hunspell_check(input, dictionary('en_US')) ~
      hunspell_suggest(input, dictionary('en_US')) %>%
      # get first suggestion, or NA if suggestions list is empty
      map(1, .default = NA) %>%
      unlist(),
    TRUE ~ input # if word is correct
  )
  # if input incorrectly spelled but no suggestions, return input word
  ifelse(is.na(output), input, output)
}

product_recipe <- recipe(desc ~ price, data = product_descriptions) %>% 
  step_tokenize(desc) %>% 
  step_stem(desc, custom_stemmer = correct_spelling) %>%
  step_tf(desc)

product_recipe %>% 
  prep() %>%
  bake(new_data = NULL)
#> # A tibble: 3 × 6
#>   price tf_desc_cad tf_desc_good tf_desc_not tf_desc_product tf_desc_sou
#>   <dbl>       <dbl>        <dbl>       <dbl>           <dbl>       <dbl>
#> 1  1000           0            1           0               1           0
#> 2   700           0            1           1               0           1
#> 3   250           1            0           0               1           0
user_49120800

Not nearly as short, but this should work:

library(tidyverse)
library(hunspell)

product_descriptions <- tibble(
  desc = c("goood product", "not sou good", "vad produkt"),
  price = c(1000, 700, 250)
)

correct_spelling <- function(input) {
  output <- case_when(
    # check and (if required) correct spelling
    !hunspell_check(input, dictionary('en_US')) ~
     hunspell_suggest(input, dictionary('en_US')) %>%
     # get first suggestion, or NA if suggestions list is empty
     map(1, .default = NA) %>%
     unlist(),
    TRUE ~ input # if word is correct
    )
    # if input incorrectly spelled but no suggestions, return input word
    ifelse(is.na(output), input, output)
}

my_stopwords <- c("sou")

product_descriptions %>% 
  #create a row identifier
  mutate(id = row_number()) %>%
  #separate all `desc` into separate words (by space) into separate rows
  separate_rows(desc, sep = " ") %>% 
  #helper for naming later on
  mutate(word_id = "word") %>%
  #word identifier
  group_by(id) %>% 
  mutate(word = row_number()) %>% 
  ungroup() %>% 
  #exclude stopwords as defined above
  filter(!desc %in% my_stopwords) %>% 
  #add spellchecker
  mutate(desc = correct_spelling(desc)) %>% 
  #make tibble wide again
  pivot_wider(names_from = c(word_id, word), values_from = desc) %>% 
  #unite all strings that were put into separate columns
  unite(desc, starts_with("word_"), remove = FALSE, sep = " ", na.rm = TRUE) %>% 
  #omit all helper columns
  select(-c(id, starts_with("word_"))) %>% 
  #clean up column ordering 
  relocate(desc, price)

In this case "sou" is deleted as a stopword and "produkt" is getting corrected to "product". The spellcheck function changes "cad" to "vad" instead of "bad", though.

user_60190280
Related