This R package provides a set of functions for measuring similarity among documents and detecting passages which have been reused. It implements shingled n-gram, skip n-gram, and other tokenizers; similarity/dissimilarity functions; pairwise comparisons; minhash and locality sensitive hashing algorithms; and a version of the Smith-Waterman local alignment algorithm suitable for natural language. It is broadly useful for, for example, detecting duplicate documents in a corpus prior to text analysis, or for identifying borrowed passages between texts. The classes provides by this package follow the model of other natural language processing packages for R, especially the NLP and tm packages. (However, this package has no dependency on Java, which should make it easier to install.)


To install this package from CRAN:


To install the development version from GitHub, use devtools.

# install.packages("devtools")
devtools::install_github("ropensci/textreuse", build_vignettes = TRUE)


There are three main approaches that one may take when using this package: pairwise comparions, minhashing/locality sensitive hashing, and extracting matching passages through text alignment.

See the introductory vignette for a description of the classes provided by this package.

vignette("textreuse-introduction", package = "textreuse")

Pairwise comparisons

In this example we will load a tiny corpus of three documents. These documents are drawn from Kellen Funk’s research into the propagation of legal codes of civil procedure in the nineteenth-century United States.

dir <- system.file("extdata/legal", package = "textreuse")
corpus <- TextReuseCorpus(dir = dir, meta = list(title = "Civil procedure"),
                          tokenizer = tokenize_ngrams, n = 7)

We have loaded the three documents into a corpus, which involves tokenizing the text and hashing the tokens. We can inspect the corpus as a whole or the individual documents that make it up.

Now we can compare each of the documents to one another. The pairwise_compare() function applies a comparison function (in this case, jaccard_similarity()) to every pair of documents. The result is a matrix of scores. As we would expect, some documents are similar and others are not.

We can convert that matrix to a data frame of pairs and scores if we prefer.

See the pairwise vignette for a fuller description.

vignette("textreuse-pairwise", package = "textreuse")

Minhashing and locality sensitive hashing

Pairwise comparisons can be very time-consuming because they grow geometrically with the size of the corpus. (A corpus with 10 documents would require at least 45 comparisons; a corpus with 100 documents would require 4,950 comparisons; a corpus with 1,000 documents would require 499,500 comparisons.) That’s why this package implements the minhash and locality sensitive hashing algorithms, which can detect candidate pairs much faster than pairwise comparisons in corpora of any significant size.

For this example we will load a small corpus of ten documents published by the American Tract Society. We will also create a minhash function, which represents an entire document (regardless of length) by a fixed number of integer hashes. When we create the corpus, the documents will each have a minhash signature.

Now we can calculate potential matches, extract the candidates, and apply a comparison function to just those candidates.

For details, see the minhash vignette.

vignette("textreuse-minhash", package = "textreuse")

Text alignment

We can also extract the optimal alignment between to documents with a version of the Smith-Waterman algorithm, used for protein sequence alignment, adapted for natural language. The longest matching substring according to scoring values will be extracted, and variations in the alignment will be marked.

For details, see the text alignment vignette.

vignette("textreuse-alignment", package = "textreuse")

Parallel processing

Loading the corpus and creating tokens benefit from using multiple cores, if available. (This works only on non-Windows machines.) To use multiple cores, set options("mc.cores" = 4L), where the number is how many cores you wish to use.

Contributing and acknowledgements

Please note that this project is released with a Contributor Code of Conduct. By participating in this project you agree to abide by its terms.

Thanks to Noam Ross for his thorough peer review of this package for rOpenSci.

rOpenSCi logo