Removing ngrams containing stopwords using tidytext

Question

UPDATE: Thanks for the input so far. I rewritten the question and added a better example to highlight the implicit requirements that were not covered in my first example.

Question I am looking for a general tidy solution to removing ngrams containing stopwords. In short, ngrams are strings of words separated by a space. A unigram contains 1 word, a bigram 2 words, and so on. My goal would be to apply this on a data frame after using unnest_tokens(). The solution should work with a data frame containing a mix of ngrams of any length (uni, bi, tri..), or at least bi & tri and above.

• For more information on ngrams, see wiki: https://en.wikipedia.org/wiki/N-gram
• I am aware of this question: Remove ngrams with leading and trailing stopwords. However, I am looking for a general solution, that does not require the stopword to be leading or trailing and that would also scale somewhat nicely.
• as pointed out in the comments, there is a solution for bigrams documented here: https://www.tidytextmining.com/ngrams.html#counting-and-filtering-n-grams

New example data

ngram_df <- tibble::tribble(
  ~Document,                   ~ngram,
          1,                    "the",
          1,              "the basis",
          1,                  "basis",
          1,       "basis of culture",
          1,                "culture",
          1,        "is ground water",
          1,           "ground water",
          1, "ground water treatment"
  )
stopword_df <- tibble::tribble(
  ~word, ~lexicon,
  "the", "custom",
   "of", "custom",
   "is", "custom"
  )
desired_output <- tibble::tribble(
  ~Document,                   ~ngram,
          1,                  "basis",
          1,                "culture",
          1,           "ground water",
          1, "ground water treatment"
  )

^{Created on 2019-03-21 by the reprex package (v0.2.1)}

Desired behaviour

• the ngram_df should be transformed into the desired_output, using the stopwords from the word column in the stopword_df.
• every row containing a stopword should be removed
• word boundaries should be respected (i.e. looking for is should not remove basis)

---

my first attempt for a reprex below:

example data

library(tidyverse)
library(tidytext)
df <- "Groundwater remediation is the process that is used to treat polluted groundwater by removing the pollutants or converting them into harmless products." %>% 
  enframe() %>% 
  unnest_tokens(ngrams, value, "ngrams", n = 2)
#apply magic here

df
#> # A tibble: 21 x 2
#>     name ngrams                 
#>    <int> <chr>                  
#>  1     1 groundwater remediation
#>  2     1 remediation is         
#>  3     1 is the                 
#>  4     1 the process            
#>  5     1 process that           
#>  6     1 that is                
#>  7     1 is used                
#>  8     1 used to                
#>  9     1 to treat               
#> 10     1 treat polluted         
#> # ... with 11 more rows

example list of stopwords

stopwords <- c("is", "the", "that", "to")

desired output

#> Source: local data frame [9 x 2]
#> Groups: <by row>
#> 
#> # A tibble: 9 x 2
#>    name ngrams                 
#>   <int> <chr>                  
#> 1     1 groundwater remediation
#> 2     1 treat polluted         
#> 3     1 polluted groundwater   
#> 4     1 groundwater by         
#> 5     1 by removing            
#> 6     1 pollutants or          
#> 7     1 or converting          
#> 8     1 them into              
#> 9     1 harmless products

^{Created on 2019-03-20 by the reprex package (v0.2.1)}

(example sentence from: https://en.wikipedia.org/wiki/Groundwater_remediation)

Answer 1

Here you have another way using the "stopwords_collapsed" from the previous answer:

swc <- paste(stopwords, collapse = "|")
df <- df[str_detect(df$ngrams, swc) == FALSE, ] #select rows without stopwords

df
# A tibble: 8 x 2
   name ngrams                 
  <int> <chr>                  
1     1 groundwater remediation
2     1 treat polluted         
3     1 polluted groundwater   
4     1 groundwater by         
5     1 by removing            
6     1 pollutants or          
7     1 or converting          
8     1 harmless products 

Here you have a simple benchmark comparing both systems:

#benchmark
txtexp <- rep(txt,1000000)
dfexp <- txtexp %>% 
    enframe() %>% 
    unnest_tokens(ngrams, value, "ngrams", n = 2)

benchmark("mutate+filter (small text)" = {df1 <- df %>%
        mutate(
            has_stop_word = str_detect(ngrams, stopwords_collapsed)
        ) %>%
        filter(!has_stop_word)},
          "[] row selection (small text)" = {df2 <- df[str_detect(df$ngrams, stopwords_collapsed) == FALSE, ]},
        "mutate+filter (large text)" = {df3 <- dfexp %>%
            mutate(
                has_stop_word = str_detect(ngrams, stopwords_collapsed)
            ) %>%
            filter(!has_stop_word)},
        "[] row selection (large text)" = {df4 <- dfexp[str_detect(dfexp$ngrams, stopwords_collapsed) == FALSE, ]},
          replications = 5,
          columns = c("test", "replications", "elapsed")
)

                           test replications elapsed
4 [] row selection (large text)            5   30.03
2 [] row selection (small text)            5    0.00
3    mutate+filter (large text)            5   30.64
1    mutate+filter (small text)            5    0.00