Remove ngrams with leading and trailing stopwords

I want to identify major n-grams in a bunch of academic papers, including n-grams with nested stopwords, but not n-grams with leading or trailing stopwords.

I have about 100 pdf files. I converted them to plain-text files through an Adobe batch command and collected them within a single directory. From there I use R. (It's a patchwork of code because I'm just getting started with text mining.)

My code:

# Make path for sub-dir which contains corpus files 
path <- file.path(getwd(), "txt")
# Load corpus files
docs <- Corpus(DirSource(path), readerControl=list(reader=readPlain, language="en"))

docs <- tm_map(docs, tolower)
docs <- tm_map(docs, stripWhitespace)
docs <- tm_map(docs, removeNumbers)
docs <- tm_map(docs, removePunctuation)

# Merge corpus (Corpus class to character vector)
txt <- c(docs, recursive=T)

# Find trigrams (but I might look for other ngrams as well)
myDfm <- dfm(txt, ngrams = 3)
# Remove sparse features
myDfm <- dfm_trim(myDfm, min_count = 5)
# Display top features
#                  as_well_as             of_the_ecosystem                  in_order_to         a_business_ecosystem       the_business_ecosystem strategic_management_journal 
#603                          543                          458                          431                          431                          359 
#in_the_ecosystem        academy_of_management                  the_role_of                the_number_of 
#336                          311                          289                          276

For example, in the top ngrams sample provided here, I'd want to keep "academy of management", but not "as well as", nor "the_role_of". I'd like the code to work for any n-gram (preferably including less than 3-grams, although I understand it's simpler in this case to just remove stopwords first).

2 answers

  • answered 2017-10-11 12:20 Patrick Perry

    Using the corpus R package, with The Wizard of Oz as an example (Project Gutenberg ID#55):

    library(Matrix) # needed for sparse matrix operations
    # download the corpus
    corpus <- gutenberg_corpus(55)
    # set the preprocessing options
    text_filter(corpus) <- text_filter(drop_punct = TRUE, drop_number = TRUE)
    # compute trigram statistics for terms appearing at least 5 times;
    # specify `types = TRUE` to report component types as well 
    stats <- term_stats(corpus, ngrams = 3, min_count = 5, types = TRUE)
    # discard trigrams starting or ending with a stopword
    stats2 <- subset(stats, !type1 %in% stopwords_en & !type3 %in% stopwords_en)
    # print first five results:
    print(stats2, 5)
    ##    term               type1 type2 type3     count support
    ## 4  said the scarecrow said  the   scarecrow    36       1
    ## 7  back to kansas     back  to    kansas       28       1
    ## 16 said the lion      said  the   lion         19       1
    ## 17 said the tin       said  the   tin          19       1
    ## 48 road of yellow     road  of    yellow       12       1
    ## ⋮  (35 rows total)
    # form a document-by-term count matrix for these terms
    x <- term_matrix(corpus, select = stats2$term)

    In your case, you can convert from the tm Corpus object with

    corpus <- as_corpus_frame(docs)

  • answered 2017-10-11 16:56 Ken Benoit

    Here's how in quanteda: use dfm_remove(), where the pattern you want to remove is the stopword list followed by the concatenator character, for the beginning and end of the expression. (Note here that for reproducibility, I have used a built-in text object.)

    # remove for your own txt
    txt <- data_char_ukimmig2010
    (myDfm <- dfm(txt, remove_numbers = TRUE, remove_punct = TRUE, ngrams = 3))
    ## Document-feature matrix of: 9 documents, 5,518 features (88.5% sparse).
    (myDfm2 <- dfm_remove(myDfm, 
                         pattern = c(paste0("^", stopwords("english"), "_"), 
                                     paste0("_", stopwords("english"), "$")), 
                         valuetype = "regex"))
    ## Document-feature matrix of: 9 documents, 1,763 features (88.6% sparse).
    ## [1] "immigration_an_unparalleled" "bnp_can_solve"               "solve_at_current"           
    ## [4] "immigration_and_birth"       "birth_rates_indigenous"      "rates_indigenous_british" 

    Bonus answer:

    You can read your pdfs using the readtext package, which also works just fine with quanteda using the above code.

    txt <- readtext("yourpdfolder/*.pdf") %>% corpus()