使用 tidytext 删除包含停用词的 ngram

Question

更新：感谢您迄今为止的意见。我重写了问题并添加了一个更好的示例来突出我的第一个示例中未涵盖的隐含要求。

问题 我正在寻找一个通用的tidy 解决方案来删除包含停用词的 ngram。简而言之，ngram 是由空格分隔的单词字符串。一个unigram包含1个单词，一个bigram包含2个单词，依此类推。我的目标是在使用unnest_tokens() 后将其应用于数据框。该解决方案应适用于包含任意长度（uni、bi、tri..）或至少 bi & tri 及以上的 ngram 混合的数据帧。

• 有关 ngram 的更多信息，请参阅 wiki：https://en.wikipedia.org/wiki/N-gram
• 我知道这个问题：Remove ngrams with leading and trailing stopwords。但是，我正在寻找一个通用的解决方案，它不需要停用词在前导或尾随，并且也可以很好地扩展。
• 正如 cmets 中所指出的，这里记录了一个二元组的解决方案：https://www.tidytextmining.com/ngrams.html#counting-and-filtering-n-grams

新示例数据

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"
  )

^{由reprex package (v0.2.1) 于 2019 年 3 月 21 日创建}

期望的行为

• 应使用stopword_df 中word 列中的停用词将ngram_df 转换为desired_output。
• 应删除包含停用词的每一行
• 应遵守单词边界（即查找 is 不应删除 basis）

---

我第一次尝试下面的reprex：

示例数据

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

停用词示例列表

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

想要的输出

#> 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

^{由reprex package (v0.2.1) 于 2019 年 3 月 20 日创建}

(例句来自：https://en.wikipedia.org/wiki/Groundwater_remediation)

Answer 1

在这里，您还有另一种使用上一个答案中的“stopwords_collapsed”的方法：

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 

这里有一个比较两个系统的简单基准：

#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