In this vignette we show how the quanteda package can be used to replicate the analysis from Matthew Jockers' book Text Analysis with R for Students of Literature (London: Springer, 2014). Most of the Jockers book consists of loading, transforming, and analyzing quantities derived from text and data from text. Because quanteda has built in most of the code to perform these data transformations and analyses, it makes it possible to replicate the results from the book with far less code. Throughout this vignette, we name objects based on Jockers' book, but follow the quanteda style guide.
In what follows, each section corresponds to the respective chapter in the book.
Our closest equivalent is simply:
install.packages("quanteda") install.packages("readtext")
But if you are reading this vignette, than chances are that you have already completed this step.
library("quanteda")
We can load the text from Moby Dick using the readtext package, directly from the Project Gutenberg website.
data_char_mobydick <- as.character(readtext::readtext("http://www.gutenberg.org/cache/epub/2701/pg2701.txt")) names(data_char_mobydick) <- "Moby Dick"
load("data_char_mobydick.rda") quanteda_options(threads = 1) RcppParallel::setThreadOptions(numThreads = 1)
The readtext()
function from the readtext package loads the text files into a data.frame
object. We can access the text from a data.frame
object (and also, as we will see, a corpus
class object). Here we will display just the first 75 characters, to prevent a massive dump of the text of the entire novel. We do this using the stri_sub()
function from the stringi package, which shows the 1st through the 75th characters of the texts of our new object data_char_mobydick
. Because we have not assigned the return from this command to any object, it invokes a print method for character objects, and is displayed on the screen.
library("stringi") stri_sub(data_char_mobydick, 1, 75)
The Gutenburg edition of the text contains some metadata before and after the text of the novel. The code below uses the regexec
and substring
functions to separate this from the text.
# extract the header information (start_v <- stri_locate_first_fixed(data_char_mobydick, "CHAPTER 1. Loomings.")[1]) (end_v <- stri_locate_last_fixed(data_char_mobydick, "orphan.")[1])
Here, we found the character index of the beginning and end of the novel, rather than counting the lines as in the book, but the result will be very similar. If we want to verify that "orphan." is the end of the novel, we can use the kwic()
function:
# verify that "orphan" is the end of the novel kwic(tokens(data_char_mobydick), "orphan")
If we want to count the number of lines, we can do so by counting the newlines in the text.
stri_count_fixed(data_char_mobydick, "\n")
To measure just the number lines in the novel itself, without the metadata, we can subset the text from the start and end of the novel part, as identified above.
stri_sub(data_char_mobydick, from = start_v, to = end_v) |> stri_count_fixed("\n")
To trim the non-book content, we use stri_sub()
to extract the text between the beginning and ending indexes found above:
novel_v <- stri_sub(data_char_mobydick, start_v, end_v) length(novel_v) stri_sub(novel_v, 1, 94) |> cat()
We begin processing the text by converting to lower case. quanteda's *_tolower()
functions work like the built-in tolower()
, with an extra option to preserve upper-case acronyms when detected. To work with the novel efficiently, however, we will first tokenise it. Then, we can manipulate it using functions such as tokens_tolower()
.
novel_v_toks <- tokens(novel_v) # lowercase text novel_v_toks_lower <- tokens_tolower(novel_v_toks)
quanteda's tokens()
function splits the text into words, with many options available for which characters should be preserved, and which should be used to define word boundaries. The default behaviour works similarly to splitting on the regular expression for non-word characters (\W
as in the book), but it much smarter. For instance, it does not treat apostrophes as word boundaries, meaning that 's
and 't
are not treated as whole words from possessive forms and contractions.
To remove punctuation, we can re-process the existing tokens:
moby_word_v <- tokens(novel_v_toks_lower, remove_punct = TRUE) (total_length <- ntoken(moby_word_v)) moby_word_v[["text1"]][1:10] moby_word_v[["text1"]][99986] moby_word_v[["text1"]][c(4, 5, 6)] # check positions of "whale" which(moby_word_v[["text1"]] == "whale") |> head()
The code below uses the tokenized text to the occurrence of the word whale. To include the possessive form whale's, we may sum the counts of both forms, count the keyword-in-context matches by regular expression or glob. A glob is a simple wildcard matching pattern common on Unix systems -- asterisks match zero or more characters.
Note that the counts below do not match those in the book, due to differences in how the book has split on any non-word character, while quanteda's tokenizer splits on a more comprehensive set of "word boundaries". quanteda's tokens()
function by default does not remove punctuation or numbers (both defined as "non-word" characters) by default. To more closely match the counts in the book, we have removed punctuation.
lengths(tokens_select(moby_word_v, "whale")) # total occurrences of "whale" including possessive lengths(tokens_select(moby_word_v, c("whale", "whale's"))) # same thing using kwic() nrow(kwic(novel_v_toks_lower, pattern = "whale")) nrow(kwic(novel_v_toks_lower, pattern = "whale*")) # includes words like "whalemen" (total_whale_hits <- nrow(kwic(novel_v_toks_lower, pattern = "^whale('s){0,1}$", valuetype = "regex")))
What fraction of the total words (excluding punctuation) in the novel are "whale"?
total_whale_hits / ntoken(novel_v_toks_lower, remove_punct = TRUE)
With ntype()
we can calculate the size of the vocabulary -- includes possessive forms, but excludes punctuation, symbols and numbers.
# total unique words length(unique(moby_word_v)) ntype(novel_v_toks_lower, remove_punct = TRUE)
To quickly sort the word types by their frequency, we can use the dfm()
command to create a matrix of counts of each word type -- a document-frequency matrix. In this case there is only one document, the entire book.
# ten most frequent words moby_dfm <- dfm(moby_word_v) moby_dfm
Getting the list of the most frequent 10 terms is easy, using textstat_frequency()
.
library("quanteda.textstats") textstat_frequency(moby_dfm, n = 10)
Finally, if we wish to plot the most frequent (50) terms, we can supply the results of textstat_frequency()
to ggplot()
to plot their frequency by their rank:
# plot frequency of 50 most frequent terms library("ggplot2") theme_set(theme_minimal()) textstat_frequency(moby_dfm, n = 50) |> ggplot(aes(x = rank, y = frequency)) + geom_point() + labs(x = "Frequency rank", y = "Term frequency")
For direct comparison with the next chapter, we also create the sorted list of the most frequently found words using this:
sorted_moby_freqs_t <- topfeatures(moby_dfm, n = nfeat(moby_dfm))
We can query the document-frequency matrix to retrieve word frequencies, as with a normal matrix:
# frequencies of "he" and "she" - these are matrixes, not numerics sorted_moby_freqs_t[c("he", "she", "him", "her")] # another method: indexing the dfm moby_dfm[, c("he", "she", "him", "her")] sorted_moby_freqs_t[1] sorted_moby_freqs_t["the"] # term frequency ratios sorted_moby_freqs_t["him"] / sorted_moby_freqs_t["her"] sorted_moby_freqs_t["he"] / sorted_moby_freqs_t["she"]
Total number of tokens:
ntoken(moby_dfm) sum(sorted_moby_freqs_t)
Relative term frequencies:
sorted_moby_rel_freqs_t <- sorted_moby_freqs_t / sum(sorted_moby_freqs_t) * 100 sorted_moby_rel_freqs_t["the"] # by weighting the dfm directly moby_dfm_pct <- dfm_weight(moby_dfm, scheme = "prop") * 100 dfm_select(moby_dfm_pct, pattern = "the")
Plotting the most frequent terms, replicating the plot from the book:
plot(sorted_moby_rel_freqs_t[1:10], type = "b", xlab = "Top Ten Words", ylab = "Percentage of Full Text", xaxt = "n") axis(1,1:10, labels = names(sorted_moby_rel_freqs_t[1:10]))
Plotting the most frequent terms using ggplot2:
textstat_frequency(moby_dfm_pct, n = 10) |> ggplot(aes(x = reorder(feature, -rank), y = frequency)) + geom_bar(stat = "identity") + coord_flip() + labs(x = "", y = "Term Frequency as a Percentage")
A dispersion plot allows us to visualize the occurrences of particular terms throughout the text. The object returned by the kwic
function can be plotted to display a dispersion plot. The quanteda textplot_
objects are based on ggplot2, so you can easily change the plot, for example by adding custom title.
# using words from tokenized corpus for dispersion library("quanteda.textplots") textplot_xray(kwic(novel_v_toks, pattern = "whale")) + ggtitle("Lexical dispersion")
To produce multiple dispersion plots for comparison, you can simply send more than one kwic()
output to textplot_xray()
:
textplot_xray( kwic(novel_v_toks, pattern = "whale"), kwic(novel_v_toks, pattern = "Ahab")) + ggtitle("Lexical dispersion")
# identify the chapter break locations chap_positions_v <- kwic(novel_v_toks, phrase(c("CHAPTER \\d")), valuetype = "regex")$from head(chap_positions_v)
Splitting the text into chapters means that we will have a collection of documents, which makes this a good time to make a corpus
object to hold the texts. Initially, we make a single-document corpus, and then use the corpus_segment()
function to split this by the string which specifies the chapter breaks.
Because of the header information, however, we want to discard the first part. We can do this by segmenting the text according to the first chapter, "CHAPTER 1. Loomings.", which is preceded by 5 newlines.
chapters_char <- data_char_mobydick |> char_segment(pattern = "\\n{5}CHAPTER 1\\. Loomings\\.\\n", valuetype = "regex", remove_pattern = FALSE) sapply(chapters_char, substring, 1, 100) # remove header segment chapters_char <- chapters_char[-1] cat(substring(chapters_char, 1, 200))
Now we can segment the text based on the chapter titles. These titles are automatically extracted into the pattern
document variables, and the text of each chapter becomes the text of each new document unit. To tidy this up, we can remove the trailing \n
character, using stri_trim_both()
, since the \n
is a member of the "whitespace" group.
chapters_corp <- chapters_char |> corpus() |> corpus_segment(pattern = "CHAPTER\\s\\d+.*\\n\\n", valuetype = "regex") chapters_corp$pattern <- stringi::stri_trim_both(chapters_corp$pattern) chapters_corp <- corpus_subset(chapters_corp, chapters_corp != "") summary(chapters_corp, 10)
For better reference, let's also rename the document labels with these chapter headings:
docnames(chapters_corp) <- chapters_corp$pattern
With the corpus split into chapters, we can use the dfm()
function to create a matrix of counts of each word in each chapter -- a document-frequency matrix.
# create a dfm chap_dfm <- tokens(chapters_corp) |> dfm() # extract row with count for "whale"/"ahab" in each chapter # and convert to data frame for plotting whales_ahabs_df <- chap_dfm |> dfm_keep(pattern = c("whale", "ahab")) |> convert(to = "data.frame") whales_ahabs_df$chapter <- 1:nrow(whales_ahabs_df) ggplot(data = whales_ahabs_df, aes(x = chapter, y = whale)) + geom_bar(stat = "identity") + labs(x = "Chapter", y = "Frequency", title = 'Occurrence of "whale"') ggplot(data = whales_ahabs_df, aes(x = chapter, y = ahab)) + geom_bar(stat = "identity") + labs(x = "Chapter", y = "Frequency", title = 'Occurrence of "ahab"')
The above plots are raw frequency plots. For relative frequency plots, (word count divided by the length of the chapter) we can weight the document-frequency matrix. To obtain expected word frequency per 100 words, we multiply by 100. To get a feel for what the resulting weighted dfm (document-feature matrix) looks like, you can inspect it with the head
function, which prints the first few rows and columns.
rel_dfm <- dfm_weight(chap_dfm, scheme = "prop") * 100 head(rel_dfm) # subset dfm and convert to data.frame object rel_chap_freq <- rel_dfm |> dfm_keep(pattern = c("whale", "ahab")) |> convert(to = "data.frame") rel_chap_freq$chapter <- 1:nrow(rel_chap_freq) ggplot(data = rel_chap_freq, aes(x = chapter, y = whale)) + geom_bar(stat = "identity") + labs(x = "Chapter", y = "Relative frequency", title = 'Occurrence of "whale"') ggplot(data = rel_chap_freq, aes(x = chapter, y = ahab)) + geom_bar(stat = "identity") + labs(x = "Chapter", y = "Relative frequency", title = 'Occurrence of "ahab"')
Correlation analysis (and many other similarity measures) can be constructed using fast, sparse means through the textstat_simil()
function. Here, we select feature comparisons for just "whale" and "ahab", and convert this into a matrix as in the book. Because correlations are sensitive to document length, we first convert this into a relative frequency using dfm_weight()
.
dfm_weight(chap_dfm, scheme = "prop") |> textstat_simil(y = chap_dfm[, c("whale", "ahab")], method = "correlation", margin = "features") |> as.matrix() |> head(2)
With the ahab frequency and whale frequency vectors extracted from the dfm, it is easy to calculate the significance of the correlation.
cor_data_df <- dfm_weight(chap_dfm, scheme = "prop") |> dfm_keep(pattern = c("ahab", "whale")) |> convert(to = "data.frame") # sample 1000 replicates and create data frame n <- 1000 samples <- data.frame( cor_sample = replicate(n, cor(sample(cor_data_df$whale), cor_data_df$ahab)), id_sample = 1:n ) # plot distribution of resampled correlations ggplot(data = samples, aes(x = cor_sample, y = after_stat(density))) + geom_histogram(colour = "black", binwidth = 0.01) + geom_density(colour = "red") + labs(x = "Correlation Coefficient", y = NULL, title = "Histogram of Random Correlation Coefficients with Normal Curve")
# length of the book in chapters ndoc(chapters_corp) # chapter names docnames(chapters_corp) |> head()
Calculating the mean word frequencies is easy:
# for first few chapters ntoken(chapters_corp) |> head() # average (ntoken(chapters_corp) / ntype(chapters_corp)) |> head()
Since the quotient of the number of tokens and number of types is a vector, we
can simply feed this to plot()
using the pipe operator:
(ntoken(chapters_corp) / ntype(chapters_corp)) |> plot(type = "h", ylab = "Mean word frequency")
For the scaled plot:
(ntoken(chapters_corp) / ntype(chapters_corp)) |> scale() |> plot(type = "h", ylab = "Scaled mean word frequency")
mean_word_use_m <- (ntoken(chapters_corp) / ntype(chapters_corp)) sort(mean_word_use_m, decreasing = TRUE) |> head()
Measures of lexical diversity can be estimated using textstat_lexdiv()
. The TTR (Type-Token Ratio), a measure used in section 6.5, can be calculated for each document of the dfm
.
tokens(chapters_corp) |> dfm() |> textstat_lexdiv(measure = "TTR") |> head(n = 10)
Another measure of lexical diversity is Hapax richness, defined as the number of words that occur only once divided by the total number of words. We can calculate Hapax richness very simply by using a logical operation on the document-feature matrix, to return a logical value for each term that occurs once, and then sum these to get a count.
# hapaxes per document rowSums(chap_dfm == 1) |> head() # as a proportion hapax_proportion <- rowSums(chap_dfm == 1) / ntoken(chap_dfm) head(hapax_proportion)
To plot this:
barplot(hapax_proportion, beside = TRUE, col = "grey", names.arg = seq_len(ndoc(chap_dfm)))
For this, and the next chapter, we simply use quanteda's excellent kwic()
function. To find the indexes of the token positions for "gutenberg", for instance, we use the following, which returns a data.frame with the name from
indicating the index position of the start of the token match:
data_tokens_mobydick <- tokens(data_char_mobydick) gutenberg_kwic <- kwic(data_tokens_mobydick, pattern = "gutenberg") head(gutenberg_kwic$from, 10)
This is going to be super easy since we don't need to reinvent the wheel here, since kwic()
already does all that we need.
Let's create a corpus containing Moby Dick but also Jane Austen's Sense and Sensibility.
data_char_senseandsensibility <- as.character(readtext::readtext("http://www.gutenberg.org/files/161/161-0.txt")) names(data_char_senseandsensibility) <- "Sense and Sensibility" litcorpus <- corpus(c(data_char_mobydick, data_char_senseandsensibility))
load("data_char_mobydick.rda") load("data_char_senseandsensibility.rda") litcorpus <- corpus(c(data_char_mobydick, data_char_senseandsensibility))
Now we can use kwic()
to find out where in each novel this occurred:
(dogkwic <- kwic(tokens(litcorpus), pattern = "dog"))
We can plot this easily too, as a lexical dispersion plot. By specifying the scale as "absolute", we are looking at absolute token index position rather than relative position, and therefore we see that Moby Dick is nearly twice as long as Sense and Sensibility.
textplot_xray(dogkwic, scale = "absolute")
Chapter 11 describes how to detect clusters in a corpus. While the book uses the XMLAuthorCorpus
, we describe clustering using U.S. State of the Union addresses included in the quanteda.corpora package. We trim the corpus with dfm_trim()
by keeping only those words that occur at least five times in the corpus and in at least three speeches.
library(quanteda.corpora) pres_dfm <- tokens(corpus_subset(data_corpus_sotu, Date > "1980-01-01"), remove_punct = TRUE) |> tokens_wordstem("en") |> tokens_remove(stopwords("en")) |> dfm() |> dfm_trim(min_termfreq = 5, min_docfreq = 3) # hierarchical clustering - get distances on normalized dfm pres_dist_mat <- dfm_weight(pres_dfm, scheme = "prop") |> textstat_dist(method = "euclidean") |> as.dist() # hiarchical clustering the distance object pres_cluster <- hclust(pres_dist_mat) # label with document names pres_cluster$labels <- docnames(pres_dfm) # plot as a dendrogram plot(pres_cluster, xlab = "", sub = "", main = "Euclidean Distance on Normalized Token Frequency")
Finally, Jockers' book introduces topic modelling of a corpus and the visualisation through wordclouds. We can easily apply functions from the topicmodels package by using quanteda's convert()
function. In our example, we use the Irish budget speeches from 2010 (data_corpus_irishbudget2010
) and classify 20 topics using Latent Dirichlet Allocation.
data(data_corpus_irishbudget2010, package = "quanteda.textmodels") dfm_speeches <- tokens(data_corpus_irishbudget2010, remove_punct = TRUE, remove_numbers = TRUE) |> tokens_remove(stopwords("en")) |> dfm() |> dfm_trim(min_termfreq = 4, max_docfreq = 10) library(topicmodels) LDA_fit_20 <- convert(dfm_speeches, to = "topicmodels") |> LDA(k = 20) # get top five terms per topic get_terms(LDA_fit_20, 5)
Add the following code to your website.
For more information on customizing the embed code, read Embedding Snippets.