Fast DocumentTermMatrix and TermDocumentMatrix Creation

Harness the power of 'quanteda', 'data.table' & 'stringi' to quickly generate 'tm' DocumentTermMatrix and TermDocumentMatrix data structures.


gofastr

Project Status: Active - The project has reached a stable, usablestate and is being activelydeveloped. BuildStatus CoverageStatus

gofastr is designed to do one thing really well...make a DocumentTermMatrix. It harnesses the power quanteda (which in turn wraps data.table, stringi, & Matrix) to quickly generate tm DocumentTermMatrix and TermDocumentMatrix data structures. There are two ways in which time is meaingingful to an analyst: (a) coding time, or the time spent writing code and (b) computational run time, or the time the computer takes to run the code. Ideally, we want to minimize both of these sources of time expenditures. The gofaster package is my attempt to reduce the time an analysts takes to turn raw text into an analysis ready data format and relies on quanteda to minimize the run time.

In my work I often get data in the form of large .csv files or SQL databases. Additionally, most of the higher level analysis of text I undertake utilizes a TermDocumentMatrix or DocumentTermMatrix as the input data. Generally, the tm package's Corpus structure is an unnecessary step in building a usable data structure that requires additional coding and run time. gofastr skips this step and uses quanteda to quickly make the DocumentTermMatrix or TermDocumentMatrix structures that are fast to code up and fast for the computer to build.

Table of Contents

Function Usage

Functions typically fall into the task category of matrix (1) creation & (2) manipulating. The main functions, task category, & descriptions are summarized in the table below:

Function Category Description
q_tdm & q_tdm_stem creation TermDocumentMatrix from string vector
q_dtm & q_dtm_stem creation DocumentTermMatrix from string vector
remove_stopwords manipulation Remove stopwords and minimal character words from TermDocumentMatrix/DocumentTermMatrix
filter_words manipulation Filter words from TermDocumentMatrix/DocumentTermMatrix
filter_tf_idf manipulation Filter low tf-idf words from TermDocumentMatrix/DocumentTermMatrix
filter_documents manipulation Filter documents from a TermDocumentMatrix/DocumentTermMatrix
select_documents manipulation Select documents from TermDocumentMatrix/DocumentTermMatrix
sub_in_na manipulation Sub missing (NA) for regex matches (default: non-content elements)

Installation

To download the development version of gofastr:

Download the zip ball or tar ball, decompress and run R CMD INSTALL on it, or use the pacman package to install the development version:

if (!require("pacman")) install.packages("pacman")
pacman::p_load_gh("trinker/gofastr")

Contact

You are welcome to:

Demonstration

Load Packages

if (!require("pacman")) install.packages("pacman")
pacman::p_load(gofastr, tm, magrittr)

DocumentTerm/TermDocument Matrices

(w <-with(presidential_debates_2012, q_dtm(dialogue, paste(time, tot, sep = "_"))))
## Non-/sparse entries: 42058/9791766
## Sparsity           : 100%
## Maximal term length: 16
## Weighting          : term frequency (tf)

(x <- with(presidential_debates_2012, q_tdm(dialogue, paste(time, tot, sep = "_"))))

## <<TermDocumentMatrix (terms: 3377, documents: 2912)>>
## Non-/sparse entries: 42058/9791766
## Sparsity           : 100%
## Maximal term length: 16
## Weighting          : term frequency (tf)

Stopwords

Stopwords are those words that we want to remove from the analysis because they give little information gain. These words occur so frequently in all documents or give very content information (i.e., function words) and thus are excluded. The remove_stopwords function allows the user to remove stopwords using three approaches/arguments:

  1. stopwords - A vector of common + resercher defined words (see lexicon package)
  2. min.char/max.char - Automatic removal of words less/greater than n characters in length
  3. denumber - Removal of words that are numbers

By default stopwords = tm::stopwords("english"), min.char = 3, and denumber =TRUE.

with(presidential_debates_2012, q_dtm(dialogue, paste(time, tot, sep = "_"))) %>%
    remove_stopwords()

## <<DocumentTermMatrix (documents: 2912, terms: 3180)>>
## Non-/sparse entries: 19014/9241146
## Sparsity           : 100%
## Maximal term length: 16
## Weighting          : term frequency (tf)

with(presidential_debates_2012, q_tdm(dialogue, paste(time, tot, sep = "_"))) %>%
    remove_stopwords()

## <<TermDocumentMatrix (terms: 3180, documents: 2912)>>
## Non-/sparse entries: 19014/9241146
## Sparsity           : 100%
## Maximal term length: 16
## Weighting          : term frequency (tf)

Weighting

As the output from gofastr matrix create functions is a true tm object, weighting is done in the standard way using tm's built in weighting functions. This is done post-hoc of creation.

with(presidential_debates_2012, q_dtm(dialogue, paste(time, tot, sep = "_"))) %>%
    tm::weightTfIdf()

## <<DocumentTermMatrix (documents: 2912, terms: 3377)>>
## Non-/sparse entries: 42058/9791766
## Sparsity           : 100%
## Maximal term length: 16
## Weighting          : term frequency - inverse document frequency (normalized) (tf-idf)

Stemming

To stem words utilize q_dtm_stem and q_tdm_stem which utilize SnowballC's stemmer under the hood.

with(presidential_debates_2012, q_dtm_stem(dialogue, paste(time, tot, sep = "_"))) %>%
    remove_stopwords()

## <<DocumentTermMatrix (documents: 2912, terms: 2261)>>
## Non-/sparse entries: 19557/6564475
## Sparsity           : 100%
## Maximal term length: 16
## Weighting          : term frequency (tf)

Manipulating via Words

Filter Out Low Occurring Words

To filter out words with counts below a threshold we use filter_words.

with(presidential_debates_2012, q_dtm(dialogue, paste(time, person, sep = "_"))) %>%
    filter_words(5)

## <<DocumentTermMatrix (documents: 10, terms: 967)>>
## Non-/sparse entries: 5021/4649
## Sparsity           : 48%
## Maximal term length: 14
## Weighting          : term frequency (tf)

Filter Out High/Low Frequency (low information) Words

To filter out words with high/low frequency in all documents (thus low information) use filter_tf_idf. The default min uses the tf-idf's median per Grüen & Hornik's (2011) demonstration.

with(presidential_debates_2012, q_dtm(dialogue, paste(time, person, sep = "_"))) %>%
    filter_tf_idf()

## <<DocumentTermMatrix (documents: 10, terms: 1689)>>
## Non-/sparse entries: 4024/12866
## Sparsity           : 76%
## Maximal term length: 16
## Weighting          : term frequency (tf)

*Grüen, B. & Hornik, K. (2011). topicmodels: An R Package for Fitting Topic Models. Journal of Statistical Software, 40(13), 1-30. http://www.jstatsoft.org/article/view/v040i13/v40i13.pdf

Manipulating via Documents

Filter Out Low Occurring Documents

To filter out documents with word counts below a threshold use filter_documents. Remember the warning from above:

Here we use filter_documents' default (a document must have a row/column sum greater than 1) to eliminate the warning:

with(presidential_debates_2012, q_dtm(dialogue, paste(time, tot, sep = "_"))) %>%
    filter_documents() %>%
    tm::weightTfIdf()

## <<DocumentTermMatrix (documents: 2912, terms: 3377)>>
## Non-/sparse entries: 42058/9791766
## Sparsity           : 100%
## Maximal term length: 16
## Weighting          : term frequency - inverse document frequency (normalized) (tf-idf)

Selecting Documents

To select only documents matching a regex use the select_documents function. This is useful for selecting only particular documents within the corpus.

with(presidential_debates_2012, q_dtm(dialogue, paste(time, person, sep = "_"))) %>%
    select_documents('romney', ignore.case=TRUE)

## <<DocumentTermMatrix (documents: 3, terms: 3377)>>
## Non-/sparse entries: 3404/6727
## Sparsity           : 66%
## Maximal term length: 16
## Weighting          : term frequency (tf)

with(presidential_debates_2012, q_dtm(dialogue, paste(time, person, sep = "_"))) %>%
    select_documents('^(?!.*romney).*$', ignore.case = TRUE)

## <<DocumentTermMatrix (documents: 7, terms: 3377)>>
## Non-/sparse entries: 4960/18679
## Sparsity           : 79%
## Maximal term length: 16
## Weighting          : term frequency (tf)

Putting It Together

Of course we can chain matrix creation functions with several of the manipulation function to quickly prepare data for analysis. Here I demonstrate preparing data for a topic model using gofastr and then the analysis. Finally, I plot the results and use the LDAvis package to interact with the results. Note that this is meant to demonstrate the types of analysis that gofastr may be of use to; the methods and parameters/hyper-parameters are selected with little regard to analysis.

pacman::p_load(tm, topicmodels, dplyr, tidyr, gofastr, devtools, LDAvis, ggplot2)

## Source topicmodels2LDAvis function
devtools::source_url("https://gist.githubusercontent.com/trinker/477d7ae65ff6ca73cace/raw/79dbc9d64b17c3c8befde2436fdeb8ec2124b07b/topicmodels2LDAvis")

## SHA-1 hash of file is f9a066b61c9f992daff3991a3293e18897268598

data(presidential_debates_2012)

## Generate Stopwords 
stops <- c(
        tm::stopwords("english"),
        "governor", "president", "mister", "obama","romney"
    ) %>%
    prep_stopwords() 

## Create the DocumentTermMatrix
doc_term_mat <- presidential_debates_2012 %>%
    with(q_dtm_stem(dialogue, paste(person, time, sep = "_"))) %>%           
    remove_stopwords(stops) %>%                                                    
    filter_tf_idf() %>%
    filter_words(4) %>%                       
    filter_documents() 

## Run the Model
lda_model <- topicmodels::LDA(doc_term_mat, 10, control = list(seed=100))

## Plot the Topics Per Person_Time
topics <- posterior(lda_model, doc_term_mat)$topics
topic_dat <- tibble::rownames_to_column(as.data.frame(topics), "Person_Time")
colnames(topic_dat)[-1] <- apply(terms(lda_model, 10), 2, paste, collapse = ", ")

gather(topic_dat, Topic, Proportion, -c(Person_Time)) %>%
    separate(Person_Time, c("Person", "Time"), sep = "_") %>%
    mutate(Person = factor(Person, 
        levels = c("OBAMA", "ROMNEY", "LEHRER", "SCHIEFFER", "CROWLEY", "QUESTION" ))
    ) %>%
    ggplot(aes(weight=Proportion, x=Topic, fill=Topic)) +
        geom_bar() +
        coord_flip() +
        facet_grid(Person~Time) +
        guides(fill=FALSE) +
        xlab("Proportion")

LDAvis of Model

The output from LDAvis is not easily embedded within an R markdown document, thus the reader will need to run the code below to interact with the results.

lda_model %>%
    topicmodels2LDAvis() %>%
    LDAvis::serVis()

Comparing Timings

On a smaller 2912 rows these are the time comparisons between gofastr and tm using Sys.time. Notice the gofaster runs faster (the creation of a corpus is expensive) and requires significantly less code.

pacman::p_load(gofastr, tm)
pd <- as.data.frame(presidential_debates_2012, stringsAsFactors = FALSE)

## tm Timing
tic <- Sys.time()
rownames(pd) <- paste("docs", 1:nrow(pd))
pd[['groups']] <- with(pd, paste(time, tot, sep = "_"))
pd <- Corpus(DataframeSource(setNames(pd[, 5:6, drop=FALSE], c('text', 'doc_id'))))

(out <- DocumentTermMatrix(pd,
    control = list(
        tokenize=scan_tokenizer,
        stopwords=TRUE,
        removeNumbers = TRUE,
        removePunctuation = TRUE,
        wordLengths=c(3, Inf)
    )
) )

## <<DocumentTermMatrix (documents: 2912, terms: 3141)>>
## Non-/sparse entries: 19349/9127243
## Sparsity           : 100%
## Maximal term length: 16
## Weighting          : term frequency (tf)

difftime(Sys.time(), tic)

## Time difference of 0.06804895 secs

## gofastr Timing
tic <- Sys.time()
x <-with(presidential_debates_2012, q_dtm(dialogue, paste(time, tot, sep = "_")))
remove_stopwords(x)

## <<DocumentTermMatrix (documents: 2912, terms: 3180)>>
## Non-/sparse entries: 19014/9241146
## Sparsity           : 100%
## Maximal term length: 16
## Weighting          : term frequency (tf)

difftime(Sys.time(), tic)

## Time difference of 0.195137 secs

With Stemming

pacman::p_load(gofastr, tm)
pd <- as.data.frame(presidential_debates_2012, stringsAsFactors = FALSE)

## tm Timing
tic <- Sys.time()
rownames(pd) <- paste("docs", 1:nrow(pd))
pd[['groups']] <- with(pd, paste(time, tot, sep = "_"))
pd <- Corpus(DataframeSource(setNames(pd[, 5:6, drop=FALSE], c('text', 'doc_id'))))
pd <- tm_map(pd, stemDocument)

(out <- DocumentTermMatrix(pd,
    control = list(
        tokenize=scan_tokenizer,
        stopwords=TRUE,
        removeNumbers = TRUE,
        removePunctuation = TRUE,
        wordLengths=c(3, Inf)
    )
) )

## <<DocumentTermMatrix (documents: 2912, terms: 2855)>>
## Non-/sparse entries: 19468/8294292
## Sparsity           : 100%
## Maximal term length: 16
## Weighting          : term frequency (tf)

difftime(Sys.time(), tic)

## Time difference of 0.1631322 secs

## gofastr Timing
tic <- Sys.time()
x <-with(presidential_debates_2012, q_dtm_stem(dialogue, paste(time, tot, sep = "_")))
remove_stopwords(x, stem=TRUE)

## <<DocumentTermMatrix (documents: 2912, terms: 2249)>>
## Non-/sparse entries: 19776/6529312
## Sparsity           : 100%
## Maximal term length: 16
## Weighting          : term frequency (tf)

difftime(Sys.time(), tic)

## Time difference of 0.170115 secs

News

NEWS

Versioning

Releases will be numbered with the following semantic versioning format:

..

And constructed with the following guidelines:

  • Breaking backward compatibility bumps the major (and resets the minor and patch)
  • New additions without breaking backward compatibility bumps the minor (and resets the patch)
  • Bug fixes and misc changes bumps the patch

gofastr 0.3.0

BUG FIXES

  • Changed the quanteda argument calls to reflect the changes in quanteda to use snake case rather than camel case arguments.

gofastr 0.2.0 - 0.2.1

NEW FEATURES

  • ngrams argument added to q_dtm and q_tdm to allow ngrams in the matrix terms (see isue #3).

  • select_document picks up an invert argument to pick documents not containing the regex match.

  • as_dtm and as_tdm added to coerce objects to tm DocumentTermMatrix and TermDocumentMatrix data structures. This will allow other packages to easily import this function for their objects.

gofastr 0.1.0 - 0.1.1

NEW FEATURES

  • partial_republican_debates_2015 data set added.

gofastr 0.0.1

gofastr is designed to do one thing really well...It harnesses the power of data.table and stringi to quickly generate tm DocumentTermMatrix and TermDocumentMatrix data structures.

Reference manual

It appears you don't have a PDF plugin for this browser. You can click here to download the reference manual.

install.packages("gofastr")

0.3.0 by Tyler Rinker, 5 months ago


http://github.com/trinker/gofastr


Report a bug at http://github.com/trinker/gofastr/issues


Browse source code at https://github.com/cran/gofastr


Authors: Tyler Rinker [aut, cre]


Documentation:   PDF Manual  


GPL-2 license


Imports data.table, quanteda, slam, SnowballC, stats, tm

Suggests testthat


See at CRAN