It offers functions for splitting, parsing, tokenizing and creating a vocabulary for big text data files. Moreover, it includes functions for building a document-term matrix and extracting information from those (term-associations, most frequent terms). It also embodies functions for calculating token statistics (collocations, look-up tables, string dissimilarities) and functions to work with sparse matrices. Lastly, it includes functions for Word Vector Representations (i.e. 'GloVe', 'fasttext') and incorporates functions for the calculation of (pairwise) text document dissimilarities. The source code is based on 'C++11' and exported in R through the 'Rcpp', 'RcppArmadillo' and 'BH' packages.
The textTinyR package consists of text processing functions for small or big data files. More details on the functionality of textTinyR can be found in the package Documentation and Vignettes. The R package can be installed, in the following Operating Systems: Linux, Mac and Windows. However, there is one limitation : chinese, japanese, korean, thai or languages with ambiguous word boundaries are not supported.
UPDATE 01-04-2018 : boost-locale is no longer a system requirement for the textTinyR package.
To install the package from CRAN use,
and to download the latest version from Github use the install_github function of the devtools package,
devtools::install_github(repo = 'mlampros/textTinyR')
I added the global_term_weights() method in the sparse_term_matrix R6 class
I removed the threads parameter from the term_associations method of the sparse_term_matrix R6-class. I modified the OpenMP clauses of the .cpp files to address the ASAN errors.
I added the triplet_data() method in the sparse_term_matrix R6 class
I removed the ngram_sequential and ngram_overlap stemmers from the vocabulary_parser function. I fixed a bug in the char_n_grams of the token_stats.h source file.
I removed the ngram_sequential and ngram_overlap stemmers from the sparse_term_matrix and tokenize_transform_vec_docs functions. I overlooked the fact that the n-gram stemming is based on the whole corpus and not on each vector of the document(s), which is the case for the sparse_term_matrix and tokenize_transform_vec_docs functions. I added a zzz.R file with a packageStartupMessage to inform the users about the previous change in n-gram stemming. I also updated the package documentation and Vignette. I modified the secondary_n_grams of the tokenization.h source file due to a bug. I've used the enc2utf8 function) to encode (utf-8) the terms of the sparse matrix.
I modified the res_token_vector(), res_token_list() [ export_all_funcs.cpp file ] and append_2file() [ tokenization.h file ] functions, because the tokenize_transform_vec_docs() function returned an incorrect output in case that the path_2folder parameter was not the empty string.
I corrected the UBSAN-memory errors, which occured in the adj_Sparsity() function of the term_matrix.h header file (the errors happen, when passing empty vectors to the armadillo batch_insertion() function)
I included detailed installation instructions for the Macintosh OSx I modified the source code to correct the boost-locale errors, which occurred during testing on Macintosh OSx
I added the following system-flag in the Makevars.in file to avoid linking errors for the Mac OS: -lboost_system I modified the term_associations and Term_Matrix_Adjust methods to avoid indexing errors I corrected mistakes in the Vignette