June 13, 2017
Textual data and natural language processing are still a niche domain within the R ecosytstem. The NLP task view gives an overview of existing work however a lot of basic infrastructure is still missing. At the rOpenSci text workshop in April we discussed many ideas for improving text processing in R which revealed several core areas that need improvement:
Participants also had many good suggestions for C/C++ libraries that text researchers in R might benefit from. Over the past weeks I was able to look into these suggestions and work on a few packages for reading and analyzing text. Below is an update on new and improved rOpenSci tools for text processsing in R!
cld3 are wrappers C++ libraries by Google for language identification. CLD2 is a Naïve Bayesian classifier, whereas CLD3 uses a neural network model. I found
cld2 to give better results for short text.
# Get the latest versions install.packages(c("cld2", "cld3")) # Vectorized function text <- c("À chaque fou plaît sa marotte.", "猿も木から落ちる", "Алты́нного во́ра ве́шают, а полти́нного че́ствуют.", "Nou breekt mijn klomp!") cld2::detect_language(text) #  "fr" "ja" "ru" "nl"
Maëlle has written a cool post comparing language classification methods using 18000
"#RolandGarros2017" tweets and Thomas reminds us that algorithms can easily be fooled. Still I found the accuracy on real text quite astonishing given the relatively small size of these libraries.
Note that the algorithm for CLD3 is still under development and the engineers at Google have recently opened their Github issues page for feedback.
Many archived documents are only available in legacy formats such as
.rtf. The only tools available for extracting text from these documents were difficult to install and could not be imported from packages and scripts.
# Get the latest versions install.packages(c("antiword", "unrtf")) # Extract text from 'rtf' file text <- unrtf::unrtf("https://jeroen.github.io/files/sample.rtf", format = "text") cat(text) ### Lots of text... # Extract text from 'doc' file text <- antiword::antiword("https://jeroen.github.io/files/UDHR-english.doc") cat(text) ### Lots of text...
Also have a look at meta packages
textreadr which wrap these and other packages for automatically reading text in many different formats.
Our pdftools package now supports reading pdf (extracting text or metadata) and rendering pdf to png, jpeg, tiff, or raw vectors on all platforms (incl. Windows).
# Read some text text <- pdftools::pdf_text('https://cran.r-project.org/doc/manuals/r-release/R-intro.pdf') cat(text) # An Introduction to R # Notes on R: A Programming Environment for Data Analysis and Graphics # Version 3.4.0 (2017-04-21) # W. N. Venables, D. M. Smith # and the R Core Team # Read meta data pdftools::pdf_info('https://cran.r-project.org/doc/manuals/r-release/R-intro.pdf') # $version #  "1.5" # # $pages #  105 # # .... much more :)
You can use either render a page into a raw bitmap array, or directly to an image format such as png, jpeg or tiff.
files <- pdftools::pdf_convert('https://cran.r-project.org/doc/manuals/r-release/R-intro.pdf', format = "png", pages = 1:5) # Converting page 1 to R-intro_1.png... done! # Converting page 2 to R-intro_2.png... done! # Converting page 3 to R-intro_3.png... done! # Converting page 4 to R-intro_4.png... done! # Converting page 5 to R-intro_5.png... done!
To extract text from scanned images, also check out our tesseract package which wraps Google’s powerful OCR engine.
# Extract incorrect from a piece of text bad <- hunspell("spell checkers are not neccessairy for langauge ninja's") print(bad[]) #  "neccessairy" "langauge" hunspell_suggest(bad[]) # [] #  "necessary" "necessarily" "necessaries" "recessionary" "accessory" "incarcerate" # # [] #  "language" "Langeland" "Lagrange" "Lange" "gaugeable" "linkage" "Langland"
The package is also used by
devtools to spell-check manual pages in R packages:
devtools::spell_check() # WORD FOUND IN # alltogether pdftools.Rd:36 # cairo pdf_render_page.Rd:42 # jpeg pdf_render_page.Rd:40 # libpoppler pdf_render_page.Rd:42, pdftools.Rd:30, description:1 # png pdf_render_page.Rd:40 # Poppler pdftools.Rd:34
Finally hunspell also exposes the underlying methods needed for spell checking such as stemming words:
# Find possible stems for each word words <- c("loving", "loved", "lover", "lovely", "love") hunspell_analyze(words) # [] #  " st:loving" " st:love fl:G" # # [] #  " st:loved" " st:love fl:D" # # [] #  " st:lover" " st:love fl:R" # # [] #  " st:lovely" " st:love fl:Y" # # [] #  " st:love"
Hunspell also suppors tokenizing words from html, latex, man, or plain text. For more advanced word extraction, check out the rOpenSci tokenizers package.