c operation on tokens and tokens_xptr objects significantly faster.dfm_lookup() that leads to wrong feature names when exclusive = TRUE (#2424).Improved the handling of invisible control characters causing some tokens operations to crash (#2407).
Addressed #2358 and #2359 more thoroughly, at the C++ level, to convert list(integer(), integer(),...) to std::vector<std::vector
Removed RcppArmadillo as a dependency in an effort to avoid UBSAN warnings in #2417.
Added tokens_trim() function similar to dfm_trim() (#2419).
Added keep_unigrams argument to tokens_compound(), to keep in the returned object the unigrams that are to be compounded (#2399).
print.tokens() now allows passing arguments to base print() via ..., providing for instance the ability to print tokens without surround quotes using quote = FALSE (#2381).
Introduces the tokens_xptr objects that extend the tokens objects with external pointers for a greater efficiency. Once tokens objects are converted to tokens_xptr objects using as.tokens_xptr(), tokens_*.tokens_xptr() methods are called automatically.
Improved C++ functions to allow the users to change the number of threads for parallel computing in more flexible manner using quanteda_options(). The value of threads can be changed in the middle of analysis pipeline.
Makes "word4" the default (word) tokeniser, with improved efficiency, language handling, and customisation options.
Replaced all occurrences of the magrittr %>% pipe with the R pipe |> introduced in R 4.1, although the %>% pipe is still re-exported and therefore available to all users of quanteda without loading any additional packages.
Added min_ntoken and max_ntoken to tokens_subset() and dfm_subset() to extract documents based on number of tokens easily. It is equivalent to selecting documents using ntoken().
Added a new argument apply_if that allows a tokens-based operation to apply only to documents that meet a logical condition. This argument has been added to tokens_select(), tokens_compound(), tokens_replace(), tokens_split(), and tokens_lookup(). This is similar to applying purrr::map_if() to a tokens object, but is implemented within the function so that it can be performed efficiently in C++.
Added new arguments append_key, separator and concatenator to tokens_lookup(). These allow tokens matched by dictionary values to be retained with their keys appended to them, separated by separator. The addition of the concatenator argument allows additional control at the lookup stage for tokens that will be concatenated from having matched multi-word dictionary values. (#2324)
Added a new argument remove_padding to ntoken() and ntype() that allows for not counting padding that might have been left over from tokens_remove(x, padding = TRUE). This changes the previous number of types from ntype() when pads exist, by counting pads by default. (#2336)
Removed dependency on RcppParallel to improve the stability of the C++ code. This change requires the users of Linux-like OS to install the Intel TBB library manually to enable parallel computing.
bootstrap_dfm() was removed for character and corpus objects. The correct way to bootstrap sentences is not to tokenize them as sentences and then bootstrap them from the dfm. This is consistent with requiring the user to tokenise objects prior to forming dfms or other "downstream" objects.
dfm() no longer works on character or corpus objects, only on tokens or other dfm objects. This was deprecated in v3 and removed in v4.
Very old arguments to dfm() options that were not visible but worked with warnings (such as stem = TRUE) are removed.
Deprecated or renamed arguments formerly passed in tokens() that formerly mapped to the v3 arguments with a warning are removed.
Methods for readtext objects are removed, since these are data.frame objects that are straightforward to convert into a corpus object.
topfeatures() no longer works on an fcm object. (#2141)
Some on-the-fly calculations applied to character or corpus objects that require a temporary tokenisation are now deprecated. This includes:
nsentence() -- use lengths(tokens(x, what = "sentence")) instead;ntype() -- use ntype(tokens(x)) instead; and.ntoken() -- use ntoken(tokens(x)) instead.char_ngrams() -- use tokens_ngrams(tokens(x)) instead.corpus.kwic() is deprecated, with the suggestion to form a corpus from using tokens_select(x, window = ...) instead.
tokens_group() works efficiently even when the number of documents and groups are very large.Fixed a potential crash when calling tokens_compound() with patterns containing paddings (#2254).
Updated for compatibility with (forthcoming) Matrix 1.5.5 handling of dimnames() for empty dimensions.
restores readtext object class method extensions, to work better with the readtext package.
Removes some unused internal methods, such as docvars.kwic() that were not exported despite matching exported generics.
Implements a "word4" tokeniser that is based on new RBBI (RuleBasedBreakIterator) rules, implemented in a new .yml file that can be edited and changed by users, but whose defaults represent a significant improvement in pattern handling for words, sentences, and other forms of patterns. These rules are customised from the ICU rules for breaks, with the standard and customised rules found now in the breakrules/ system folder, so that they could, in principle, be modified by the user.
Other minor changes:
"word2":
\\p{M}).preserve_special() that rejoined splits created by the default stringi tokeniser machinery.dfm_group() now works correctly with an empty dfm (#2225).convert(x, to = "stm") no longer vulnerable to large numbers of removed features as in #2189.segid() is added to extract document serial numbers from corpus, tokens or dfm objects.Fixes test failures caused by recent changes to Matrix package behaviours on some operating systems.
fcm() to prevent some (chance) errors downstream in LSX. (#2181)fcm() computes the marginal frequency of upper-case tokens correctly (#2176).tokens_chunk() keeps all the docid, including those of empty documents, in the original object.tokens_select() recycles values when the length of startpos or endpos is less than ndoc(x).tokens_lookup() and dfm_lookup() can apply very large dictionaries (more than 100,000 keys).dfm_lookup() ignores matches of multiple dictionary values in the same key in a similar way as tokens_lookup() (#2159).split_tags argument has been added to tokens(), to provide the user with an option not to preserve social media tags (addresses #2156).dfm() returns a dfm with the identical column order even if tokens_compound() or tokens_ngrams() is used in the upstream (#2100).dfm_group() with NA values in a grouping variable now drops those, similar to the behaviour of tokens_group() and corpus_group() (#2134).char_wordstem() now has a a new argument check_whitespace, which will not throw an error when lower-casing text containing a whitespace character.dfm_remove() now has a new argument padding = FALSE that when TRUE, collects counts of the removed features in the first column. This produces results consistent with what is compiled as a dfm built from tokens where some have been removed with padding = TRUE (#2152).rbind.dfm() now preserves docvars (#2109).data_corpus_inaugural is now consistent with all other documents.phrase() now has a separator argument (#2124).phrase() methods for tokens, collocations, and lists are deprecated in favour of as.phrase() (#2129).quanteda 3.0 is a major release that improves functionality, completes the modularisation of the package begun in v2.0, further improves function consistency by removing previously deprecated functions, and enhances workflow stability and consistency by deprecating some shortcut steps built into some functions.
Modularisation: We have now separated the textplot_*() functions from the main package into a separate package quanteda.textplots, and the textstat_*() functions from the main package into a separate package quanteda.textstats. This completes the modularisation begun in v2 with the move of the textmodel_*() functions to the separate package quanteda.textmodels. quanteda now consists of core functions for textual data processing and management.
The package dependency structure is now greatly reduced, by eliminating some unnecessary package dependencies, through modularisation, and by addressing complex downstream dependencies in packages such as stopwords. v3 should serve as a more lightweight and more consistent platform for other text analysis packages to build on.
We have added non-standard evaluation for by and groups arguments to access object docvars:
*_sample() functions' argument by, and groups in the *_group() functions, now take unquoted document variable (docvar) names directly, similar to the way the subset argument works in the *_subset() functions.by = "document" formerly sampled from docid(x), but this functionality is now removed. Instead, use by = docid(x) to replicate this functionality.groups, the default is now docid(x), which is now documented more completely. See ?groups and ?docid.dfm() has a new argument, remove_padding, for removing the "pads" left behind after removing tokens with padding = TRUE. (For other extensive changes to dfm(), see "Deprecated" below.)
tokens_group(), formerly internal-only, is now exported.
corpus_sample(), dfm_sample(), and tokens_sample() now work consistently (#2023).
The kwic() return object structure has been redefined, and built with an option to use a new function index() that returns token spans following a pattern search. (#2045 and #2065)
The punctuation regular expression and that for matching social media usernames has now been redefined so that the valid Twitter username @_ is now counted as a "tag" rather than as "punctuation". (#2049)
The data object data_corpus_inaugural has been updated to include the Biden 2021 inaugural address.
A new system of validators for input types now provides better argument type and value checking, with more consistent error messages for invalid types or values.
Upon startup, we now message the console with the Unicode and ICU version information. Because we removed our redefinition of View() (see below), the former conflict warning is now gone.
as.character.corpus() now has a use.names = TRUE argument, similar to as.character.tokens() (but with a different default value).
The main potentially breaking changes in version 3 relate to the deprecation or elimination of shortcut steps that allowed functions that required tokens inputs to skip the tokens creation step. We did this to require users to take more direct control of tokenization options, or to substitute the alternative tokeniser of their choice (and then coercing it to tokens via [as.tokens()]). This also allows our function behaviour to be more consistent, with each function performing a single task, rather than combining functions (such as tokenisation and constructing a matrix).
The most common example involves constructing a dfm directly from a character
or corpus object. Formerly, this would construct a tokens object internally
before creating the dfm, and allowed passing arguments to tokens() via ....
This is now deprecated, although still functional with a warning.
We strongly encourage either creating a tokens object first, or piping the
tokens return to dfm() using %>%. (See examples below.)
We have also deprecated direct character or corpus inputs to [kwic()], since this also requires a tokenised input.
The full listing of deprecations is:
dfm.character() and dfm.corpus() are deprecated. Users should create a tokens object first, and input that to dfm().
dfm(): As of version 3, only tokens objects are supported as inputs to dfm(). Calling dfm() for character or corpus objects is still functional, but issues a warning. Convenience passing of arguments to tokens() via ... for dfm() is also deprecated, but undocumented, and functions only with a warning. Users should now create a tokens object (using tokens() from character or corpus inputs before calling dfm().
kwic(): As of version 3, only tokens objects are supported as inputs to kwic(). Calling kwic() for character or corpus objects is still functional, but issues a warning. Passing arguments to tokens() via ... in kwic() is now disabled. Users should now create a tokens object (using tokens() from character or corpus inputs before calling kwic().
Shortcut arguments to dfm() are now deprecated. These are still active, with a warning, although they are no longer documented. These are:
stem -- use tokens_wordstem() or dfm_wordstem() instead.select, remove -- use tokens_select() / dfm_select() or tokens_remove() / dfm_remove() instead.dictionary, thesaurus -- use tokens_lookup() or dfm_lookup() instead.valuetype, case_insensitive -- these are disabled; for the deprecated arguments that take these qualifiers, they are fixed to the defaults "glob" and TRUE.groups -- use tokens_group() or dfm_group() instead.texts() and texts<- are deprecated.
as.character.corpus() to turn a corpus into a simple named character vector.corpus_group() instead of texts(x, groups = ...) to aggregate texts by a grouping variable.[<- instead of texts()<- for replacing texts in a corpus object.See note above under "Changes" about the textplot_*() and textstat_*() functions.
The following functions have been removed:
corpuszip objects.View() functionsas.wfm() and as.DocumentTermMatrix() (the same functionality is available via convert())metadoc() and metacorpus()corpus_trimsentences() (replaced by corpus_trim())tortl functionsdfm objects can no longer be used as a pattern in dfm_select() (formerly deprecated).
dfm_sample():
margin argument. Instead, dfm_sample() now samples only on documents, the same as corpus_sample() and tokens_sample(); andby = "document" -- use by = docid(x) instead.dictionary_edit(), char_edit(), and list_edit() are removed.
dfm_weight() - formerly deprecated "scheme" options are now removed.
tokens() - formerly deprecated options remove_hyphens and remove_twitter are now removed. (Use split_hyphens instead, and the default tokenizer always now preserves Twitter and other social media tags.)
Special versions of head() and tail() for corpus, dfm, and fcm objects are now removed, since the base methods work fine for these objects. The main consequence was the removal of the nf option from the methods for dfm and fcm objects, which limited the number of features. This can be accomplished using the index operator [ instead, or for printing, by specifying print(x, max_nfeat = 6L) (for instance).
Fixed a bug causing topfeatures(x, group = something) to fail with weighted dfms (#2032).
kwic() is more stable and does not crash when a vector is supplied as the window argument (#2008).
Allow use of multi-threading with more than two threads by fixing quanteda_options().
Mentions of the now-removed ngrams option in dfm(x, ...) has now been removed from the dfm documentation. (#1990)
Handling for some early-cycle v2 dfm object is improved, to ensure that they are updated to the latest object format. (#2097)
textstat_keyness() performance is now improved through implementation in (multi-threaded) C++.corpus_reshape() now allows reshaping back to documents even when segmented texts were of zero length. (#1978)summary.corpus()/textstat_summary().block_size to quanteda_options() to control the number of documents in blocked tokenization.print.dictionary2() to control the printing of nested levels with max_nkey (#1967)textstat_summary() to provide detailed information about dfm, tokens and corpus objects. It will replace summary() in future versions.what = "word") corpora with large numbers of documents that contain social media tags and URLs that needed to be preserved (such a large corpus of Tweets).quanteda_options(). The following are now preserved: "#政治" as well as Weibo-style hashtags such as "#英国首相#".convert(x, to = "data.frame") now outputs the first column as "doc_id" rather than "document" since "document" is a commonly occurring term in many texts. (#1918)char_select(), char_keep(), and char_remove() for easy manipulation of character vectors.dictionary_edit() for easy, interactive editing of dictionaries, plus the functions char_edit() and list_edit() for editing character and list of character objects.textplot_wordcloud() that plots objects from textstat_keyness(), to visualize keywords either by comparison or for the target category only.kwic() (#1840).logsmooth scheme to dfm_weight().textstat_summary() method, which returns summary information about the tokens/types/features etc in an object. It also caches summary information so that this can be retrieved on subsequent calls, rather than re-computed.NA for non-existent features when n > nfeat(x) in textstat_frequency(x, n). (#1929)dfm_lookup() and tokens_lookup() in which an error was caused when no dictionary key returned a single match (#1946).textstat_simil/dist object converted to a data.frame to drop its document2 labels (#1939).dfm_match() to fail on a dfm that included "pads" (""). (#1960)data_dfm_lbgexample object using more modern dfm internals.textstat_readability(), textstat_lexdiv(), and nscrabble() so that empty texts are not dropped in the result. (#1976)data_corpus_irishbudget2010 and data_corpus_dailnoconf1991 to the quanteda.textmodels package.stringsAsFactors = FALSE for data.frame objects.tokens_replace() when the pattern was not matched (#1895)quanteda 2.0 introduces some major changes, detailed here.
New corpus object structure.
The internals of the corpus object have been redesigned, and now are based around a character vector with meta- and system-data in attributes. These are all updated to work with the existing extractor and replacement functions. If you were using these before, then you should not even notice the change. Docvars are now handled separately from the texts, in the same way that docvars are handled for tokens objects.
New metadata handling.
Corpus-level metadata is now inserted in a user metadata list via meta() and meta<-(). metacorpus() is kept as a synonym for meta(), for backwards compatibility. Additional system-level corpus information is also recorded, but automatically when an object is created.
Document-level metadata is deprecated, and now all document-level information is simply a "docvar". For backward compatibility, metadoc() is kept and will insert document variables (docvars) with the name prefixed by an underscore.
Corpus objects now store default summary statistics for efficiency. When these are present, summary.corpus() retrieves them rather than computing them on the fly.
New index operators for core objects. The main change here is to redefine the $ operator for corpus, tokens, and dfm objects (all objects that retain docvars) to allow this operator to access single docvars by name. Some other index operators have been redefined as well, such as [.corpus returning a slice of a corpus, and [[.corpus returning the texts from a corpus.
See the full details at https://github.com/quanteda/quanteda/wiki/indexing_core_objects.
*_subset() functions.
The subset argument now must be logical, and the select argument has been removed. (This is part of base::subset() but has never made sense, either in quanteda or base.)
Return format from textstat_simil() and textstat_dist().
Now defaults to a sparse matrix from the Matrix package, but coercion methods are provided for as.data.frame(), to make these functions return a data.frame just like the other textstat functions. Additional coercion methods are provided for as.dist(), as.simil(), and as.matrix().
settings functions (and related slots and object attributes) are gone. These are now replaced by a new meta(x, type = "object") that records object-specific meta-data, including settings such as the n for tokens (to record the ngrams).
All included data objects are upgraded to the new formats. This includes the three corpus objects, the single dfm data object, and the LSD 2015 dictionary object.
New print methods for core objects (corpus, tokens, dfm, dictionary) now exist, each with new global options to control the number of documents shown, as well as the length of a text snippet (corpus), the tokens (tokens), dfm cells (dfm), or keys and values (dictionary). Similar to the extended printing options for dfm objects, printing of corpus objects now allows for brief summaries of the texts to be printed, and for the number of documents and the length of the previews to be controlled by new global options.
All textmodels and related functions have been moved to a new package quanteda.textmodels. This makes them easier to maintain and update, and keeps the size of the core package down.
quanteda v2 implements major changes to the tokens() constructor. These are designed to simplify the code and its maintenance in quanteda, to allow users to work with other (external) tokenizers, and to improve consistency across the tokens processing options. Changes include:
A new method tokens.list(x, ...) constructs a tokens object from named list of characters, allowing users to tokenize texts using some other function (or package) such as tokenize_words(), tokenize_sentences(), or tokenize_tweets() from the tokenizers package, or the list returned by spacyr::spacy_tokenize(). This allows users to use their choice of tokenizer, as long as it returns a named list of characters. With tokens.list(), all tokens processing (remove_*) options can be applied, or the list can be converted directly to a tokens object without processing using as.tokens.list().
All tokens options are now intervention options, to split or remove things that by default are not split or removed. All remove_* options to tokens() now remove them from tokens objects by calling tokens.tokens(), after constructing the object. "Pre-processing" is now actually post-processing using tokens_*() methods internally, after a conservative tokenization on token boundaries. This both improves performance and improves consistency in handling special characters (e.g. Twitter characters) across different tokenizer engines. (#1503, #1446, #1801)
Note that tokens.tokens() will remove what is found, but cannot "undo" a removal -- for instance it cannot replace missing punctuation characters if these have already been removed.
The option remove_hyphens is removed and deprecated, but replaced by split_hyphens. This preserves infix (internal) hyphens rather than splitting them. This behaviour is implemented in both the what = "word" and what = "word2" tokenizer options. This option is FALSE by default.
The option remove_twitter has been removed. The new what = "word" is a smarter tokenizer that preserves social media tags, URLs, and email-addresses. "Tags" are defined as valid social media hashtags and usernames (using Twitter rules for validity) rather than removing the # and @ punctuation characters, even if remove_punct = TRUE.
size argument in dfm_sample() to the number of features, not the number of documents. (#1643)startpos and endpos arguments to tokens_select(), for selecting on token positions relative to the start or end of the tokens in each document. (#1475)convert() method for corpus objects, to convert them into data.frame or json formats.spacy_tokenize() method for corpus objects, to provide direct access via the spacyr package.force = TRUE option and error checking for the situations of applying dfm_weight() or dfm_group() to a dfm that has already been weighted. (#1545) The function textstat_frequency() now allows passing this argument to dfm_group() via .... (#1646)textstat_frequency() now has a new argument for resolving ties when ranking term frequencies, defaulting to the "min" method. (#1634)$. (See Index Operators for Core Objects above.)textstat_entropy() now produces a data.frame that is more consistent with other textstat methods. (#1690)tokens_group() and dfm_group() are more robust to using multiple grouping variables, and preserve these correctly as docvars in the new dfm. (#1809)textstat_lexdiv().featfreq() to compute the overall feature frequencies from a dfm.tokens_lookup() when exclusive = FALSE and the tokens object has paddings. (#1743)tokens_replace() (#1765).omit_empty as an argument to convert(), to allow the user to control whether empty documents are excluded from converted dfm objects for certain formats. (#1660)textstat_dist() and textstat_simil(). (#1730)textstat_dist() and textstat_simil() class symmetric matrices.flatten and levels arguments to as.list.dictionary2() to enable more flexible conversion of dictionary objects. (#1661)corpus_sample(), the size now works with the by argument, to control the size of units sampled from each group.textstat_dist() and textstat_simil(), see below.tokens(). (#1713)textstat_dist() and textstat_simil() now return sparse symmetric matrix objects using classes from the Matrix package. This replaces the former structure based on the dist class. Computation of these classes is now also based on the fast implementation in the proxyC package. When computing similarities, the new min_simil argument allows a user to ignore certain values below a specified similarity threshold. A new coercion method as.data.frame.textstat_simildist() now exists for converting these returns into a data.frame of pairwise comparisons. Existing methods such as as.matrix(), as.dist(), and as.list() work as they did before.textstat_dist() and textstat_simil() because these were either not symmetric or not invariant to document or feature ordering. Finally, the selection argument has been deprecated in favour of a new y argument.textstat_readability() now defaults to measure = "Flesch" if no measure is supplied. This makes it consistent with textstat_lexdiv() that also takes a default measure ("TTR") if none is supplied. (#1715)max_nchar and min_nchar in tokens_select() are now NULL, meaning they are not applied if the user does not supply values. Fixes #1713.kwic.corpus() and kwic.tokens() behaviour now aligned, meaning that dictionaries are correctly faceted by key instead of by value. (#1684)tokens() verbose output. (#1683)textstat_readability(). (#1701)docvars<-.corpus() in a way that solves #1603 (reassignment of docvar names).dfm_compress() and dfm_group() that changed or deleted docvars attributes of dfm objects (#1506).textplot_xray() that caused incorrect facet labels when a pattern contained multiple list elements or values (#1514).kwic() now correctly returns the pattern associated with each match as the "keywords" attribute, for all pattern types (#1515)textstat_simil() and textstat_dist().textstat_lexdiv() now works on tokens objects, not just dfm objects. New methods of lexical diversity now include MATTR (the Moving-Average Type-Token Ratio, Covington & McFall 2010) and MSTTR (Mean Segmental Type-Token Ratio).tokens_split() allows splitting single into multiple tokens based on a pattern match. (#1500)tokens_chunk() allows splitting tokens into new documents of equally-sized "chunks". (#1520)textstat_entropy() now computes entropy for a dfm across feature or document margins.textstat_readability() is vastly improved, now providing detailing all formulas and providing full references.dfm_match() allows a user to specify the features in a dfm according to a fixed vector of feature names, including those of another dfm. Replaces dfm_select(x, pattern) where pattern was a dfm.vertex_labelsize added to textplot_network() to allow more precise control of label sizes, either globally or individually.tokens.tokens(x, remove_hyphens = TRUE) where x was generated with remove_hyphens = FALSE now behaves similarly to how the same tokens would be handled had this option been called on character input as tokens.character(x, remove_hyphens = TRUE). (#1498)textstat_keyness() (#1482).textstat_simil() return object coerced to matrix now default to 1.0, rather than 0.0 (#1494).textstat_lexdiv(): Yule's K, Simpson's D, and Herdan's Vm.fcm(x, ordered = TRUE). (#1413) Also set the condition that window can be of size 1 (formerly the limit was 2 or greater).tokens(x, what = "fasterword", remove_separators = TRUE) so that it correctly splits words separated by \n and \t characters. (#1420)textstat_readability(), fixed a bug in Dale-Chall-based measures and in the Spache word list measure. These were caused by an incorrect lookup mechanism but also by limited implementation of the wordlists. The new wordlists include all of the variations called for in the original measures, but using fast fixed matching. (#1410)rowMeans(), rowSums(), colMeans(), colSums()) caused by not having access to the Matrix package methods. (#1428)textplot_scale1d() when input a predicted wordscores object with se.fit = TRUE (#1440).textplot_network(). (#1460)intermediate to textstat_readability(x, measure, intermediate = FALSE), which if TRUE returns intermediate quantities used in the computation of readability statistics. Useful for verification or direct use of the intermediate quantities.separator argument to kwic() to allow a user to define which characters will be added between tokens returned from a keywords in context search. (#1449)textstat_dist() and textstat_simil() in C++ for enhanced performance. (#1210)tokens_sample() function (#1478).textstat_dist() (#1443), based on the reasoning in #1442.textstat_simil(). (#1442)predict.textmodel_worscores() when training and test feature sets are difference (#1380).char_segment() and corpus_segment() are more robust to whitespace characters preceding a pattern (#1394).tokens_ngrams() is more robust to handling large numbers of documents (#1395).corpus.data.frame() is now robust to handling data.frame inputs with improper or missing variable names (#1388).as.igraph.fcm() method for converting an fcm object into an igraph graph object.case_insensitive argument to char_segment() and corpus_segment().to = "tripletlist" output type for convert(), to convert a dfm into a simple triplet list. (#1321)tokens_tortl() and char_tortl() to add markers for right-to-left language tokens and character objects. (#1322)corpus.kwic() by adding new arguments split_context and extract_keyword.dfm_remove(x, selection = anydfm) is now equivalent to dfm_remove(x, selection = featnames(anydfm)). (#1320)predict.textmodel_nb() returns, and added type = argument. (#1329)textmodel_affinity() that caused failure when the input dfm had been compiled with tolower = FALSE. (#1338)tokens_lookup() and dfm_lookup() when nomatch is used. (#1347)"NA" (#1372)nsentence() method for spacyr parsed objects. (#1289)nsyllable() that incorrectly handled cased words, and returned wrong names with use.names = TRUE. (#1282)summary.character() caused by previous import of the network package namespace. (#1285)dfm_smooth() now correctly sets the smooth value in the dfm (#1274). Arithmetic operations on dfm objects are now much more consistent and do not drop attributes of the dfm, as sometimes happened with earlier versions.tokens_toupper() and tokens_tolower() no longer remove unused token types. Solves #1278.dfm_trim() now takes more options, and these are implemented more consistently. min_termfreq and max_termfreq have replaced min_count and max_count, and these can be modified using a termfreq_type argument. (Similar options are implemented for docfreq_type.) Solves #1253, #1254.textstat_simil() and textstat_dist() now take valid dfm indexes for the relevant margin for the selection argument. Previously, this could also be a direct vector or matrix for comparison, but this is no longer allowed. Solves #1266.dfm_group() (#1295).as.dfm() methods for tm DocumentTermMatrix and TermDocumentMatrix objects. (#1222)predict.textmodel_wordscores() now includes an include_reftexts argument to exclude training texts from the predicted model object (#1229). The default behaviour is include_reftexts = TRUE, producing the same behaviour as existed before the introduction of this argument. This allows rescaling based on the reference documents (since rescaling requires prediction on the reference documents) but provides an easy way to exclude the reference documents from the predicted quantities.textplot_wordcloud() now uses code entirely internal to quanteda, instead of using the wordcloud package.textplot_scale1d() by adjusting the refscores for data_corpus_irishbudget2010.dfm_trim() and dfm_weight() for previously weighted dfm objects and when supplied thresholds are proportions instead of counts. (#1237)summary.corpus(x, n = 101) when ndoc(x) > 100 (#1242).predict.textmodel_wordscores(x, rescaling = "mv") that always reset the reference values for rescaling to the first and second documents (#1251).textplot_keyness() are now resolved (#1233, #1233).textmodel_wordfish() to sparse = FALSE, in response to #1216.dfm_group() now preserves docvars that are constant for the group aggregation (#1228).quanteda_options(threads = ...).vertex_labelfont to textplot_network().textmodel_lsa() for Latent Semantic Analysis models.textmodel_affinity() for the Perry and Benoit (2017) class affinity scaling model.textplot_network() function.stopwords() function and the associated internal data object data_char_stopwords have been removed from quanteda, and replaced by equivalent functionality in the stopwords package.tokens_subset(), now consistent with other *_subset() functions (#1149).fcm() and for textmodel_wordfish().dfm() now correctly passes through all ... arguments to tokens(). (#1121)dfm_*() functions now work correctly with empty dfm objects. (#1133)dfm_weight() for named weight vectors (#1150)textplot_influence() from working (#1116).convert() are simplified and no longer exported. To convert a dfm, convert() is now the only official function.nfeat() replaces nfeature(), which is now deprecated. (#1134)textmodel_wordshoal() has been removed, and relocated to a new package (wordshoal).textmodel(), which used to be a gateway to specific textmodel_*() functions, has been removed.textmodel_*() have been reimplemented to make their behaviour consistent with the lm/glm() families of models, including especially how the predict, summary, and coef methods work (#1007, #108).tokens_segment() has a new window argument, permitting selection within an asymmetric window around the pattern of selection. (#521)tokens_replace() now allows token types to be substituted directly and quickly.textmodel_affinity() now adds functionality to fit the Perry and Benoit (2017) class affinity model.spacy_parse method for corpus objects. Also restored quanteda methods for spacyr spacy_parsed objects.textmodel_nb() (#1010), and made output quantities from the fitted NB model regular matrix objects instead of Matrix classes.tokens_group() is now significantly faster.tokenize() function and all methods associated with the tokenizedTexts object types have been removed.tokens_keep(), dfm_keep(), and fcm_keep(). (#1037)textmodel_NB() has been replaced by textmodel_nb().textmodel_lsa() for Latent Semantic Analysis.tokens_lookup(..., exclusive = FALSE).tokens_segment(), which works on tokens objects in the same way as corpus_segment() does on corpus objects (#902).%>% can now be used with quanteda without needing to attach magrittr (or, as many users apparently believe, the entire tidyverse.)corpus_segment() now behaves more logically and flexibly, and is clearly differentiated from corpus_reshape() in terms of its functionality. Its documentation is also vastly improved. (#908)data_dictionary_LSD2015, the Lexicoder Sentiment 2015 dictionary (#963).tokens_lookup() and dfm_lookup() (#960).head.corpus(), tail.corpus() provide fast subsetting of the first or last documents in a corpus. (#952)purrr::map() to dfm() (#928).regex2fixed() and associated functions.textstat_collocations.tokens() caused by "documents" containing only "" as tokens. (#940)cbind.dfm() when features shared a name starting with quanteda_options("base_featname") (#946)quanteda_options(). (#966)summary.corpus() now generates a special data.frame, which has its own print method, rather than requiring verbose = FALSE to suppress output (#926).textstat_collocations() is now multi-threaded.head.dfm(), tail.dfm() now behave consistently with base R methods for matrix, with the added argument nfeature. Previously, these methods printed the subset and invisibly returned it. Now, they simply return the subset. (#952)textstat_collocations(), which computes only the lambda method for now, but does so accurately and efficiently. (#753, #803). This function is still under development and likely to change further.quanteda_options that affect the maximum documents and features displayed by the dfm print method (#756).ngram formation is now significantly faster, including with skips (skipgrams).topfeatures():
phrase() converts whitespace-separated multi-word patterns into a list of patterns. This affects the feature/pattern matching in tokens/dfm_select/remove, tokens_compound, tokens/dfm_lookup, and kwic. phrase() and the associated changes also make the behaviour of using character vectors, lists of characters, dictionaries, and collocation objects for pattern matches far more consistent. (See #820, #787, #740, #837, #836, #838)corpus.Corpus() for creating a corpus from a tm Corpus now works with more complex objects that include document-level variables, such as data from the manifestoR package (#849).textplot_keyness() plots term "keyness", the association of words with contrasting classes as measured by textstat_keyness().tokens() that improve the consistency and efficiency of the tokenization.quanteda_options(): language_stemmer and language_stopwords, now used for default in *_wordstem functions and stopwords() for defaults, respectively. Also uses this option in dfm() when stem = TRUE, rather than hard-wiring in the "english" stemmer (#386).textstat_frequency() to compile feature frequencies, possibly by groups. (#825)nomatch option to tokens_lookup() and dfm_lookup(), to provide tokens or feature counts for categories not matched to any dictionary key. (#496)sequences() and collocations() have been removed and replaced by textstat_collocations().dfm objects with one or both dimensions having zero length, and empty kwic objects now display more appropriately in their print methods (per #811).*_select, *_remove, tokens_compound, features has been replaced by pattern, and in kwic, keywords has been replaced by pattern. These all behave consistently with respect to pattern, which now has a unified single help page and parameter description.(#839) See also above new features related to phrase().tokens_* functions using hashed tokens, making some of them 10x faster (#853).dfm_group() function now allow "empty" documents to be created using the fill = TRUE option, for making documents conform to a selection (similar to how dfm_select() works for features, when supplied a dfm as the pattern argument). The groups argument now behaves consistently across the functions where it is used. (#854)dictionary() now requires its main argument to be a list, not a series of elements that can be used to build a list.tokens() have improved the behaviour of remove_hyphens = FALSE, which now behaves more correctly regardless of the setting of remove_punct (#887).cbind.dfm() function allows cbinding vectors, matrixes, and (recyclable) scalars to dfm objects.textstat_collocations(), we corrected the word matching, and lambda and z calculation methods, which were slightly incorrect before. We also removed the chi2, G2, and pmi statistics, because these were incorrectly calculated for size > 2.textmodel_NB(x, y, distribution = "Bernoulli") was previously inactive even when this option was set. It has now been fully implemented and tested (#776, #780).remove_separators argument in tokens(). See #796.ntoken() and ntype(). (#795)quanteda_options() now does not throw an error when quanteda functions are called directly without attaching the package. In addition, quanteda options can be set now in .Rprofile and will not be overwritten when the options initialization takes place when attaching the package.textstat_readability() that wrongly computed the number of words with fewer than 3 syllables in a text; this affected the FOG.NRI and the Linsear.Write measures only."logave" and "inverseprob".quanteda_options() did not actually set the number of threads. In addition, we fixed a bug causing threading to be turned off on macOS (due to a check for a gcc version that is not used for compiling the macOS binaries) prevented multi-threading from being used at all on that platform.quanteda_options() are called without the namespace or package being attached or loaded (#864).corpus() now works for a tm::SimpleCorpus object. (#680)corpus_trim() and char_trim() functions for selecting documents or subsets of documents based on sentence, paragraph, or document lengths.$meta of the return object.dfm_group(x, groups = ) command, a convenience wrapper around dfm.dfm(x, groups = ) (#725).doc_id and text fields, which also provides interoperability with the readtext package. corpus construction methods are now more explicitly tailored to input object classes.dfm_lookup() behaves more robustly on different platforms, especially for keys whose values match no features (#704).textstat_simil() and textstat_dist() no longer take the n argument, as this was not sorting features in correct order.tokens(x, what = "character") when x included Twitter characters @ and # (#637).ntype.dfm() produced an incorrect result.textstat_readability() and textstat_lexdiv() for single-document returns when drop = TRUE.corpus_reshape().print, and head, and tail methods for dfm are more robust (#684).convert(x, to = "stm") caused by zero-count documents and zero-count features in a dfm (#699, #700, #701). This also removes docvar rows from $meta when this is passed through the dfm, for zero-count documents.dictionary(). (#722)dfm_compress now preserves a dfm's docvars if collapsing only on the features margin, which means that dfm_tolower() and dfm_toupper() no longer remove the docvars.fcm_compress() now retains the fcm class, and generates and error when an asymmetric compression is attempted (#728).textstat_collocations() now returns the collocations as character, not as a factor (#736)dfm_lookup(x, exclusive = FALSE) wherein an empty dfm ws returned with there was no no match (#116).dfm() to tokens() is now robust, and preserves variables defined in the calling environment (#721).str(), names(), or other indexing operations, which started happening on Linux and Windows platforms following the CRAN move to 3.4.0. (#744)dfm_weight() now print friendlier error messages when the weight vector contains features not found in the dfm. See this Stack Overflow question for the use case that sparked this improvement.corpus_reshape() can now go from sentences and paragraph units back to documents.by = argument to corpus_sample(), for use in bootstrap resampling of sub-document units.bootstrap_dfm() to generate a list of dimensionally-equivalent dfm objects based on sentence-level resampling of the original documents.tokens() and dfm() for passing docvars through to to tokens and dfm objects, and added docvars() and metadoc() methods for tokens and dfm class objects. Overall, the code for docvars and metadoc is now more robust and consistent.docvars() on eligible objects that contain no docvars now returns an empty 0 x 0 data.frame (in the spirit of #242).textmodel_scale1d now produces sorted and grouped document positions for fitted wordfish models, and produces a ggplot2 plot object.textmodel_wordfish() now preserves sparsity while processing the dfm, and uses a fast approximation to an SVD to get starting values. This also dramatically improves performance in computing this model. (#482, #124)kwic() is now dramatically improved, and also returns an indexed set of tokens that makes subsequent commands on a kwic class object much faster. (#603)quanteda_options().corpus_segment(). (#634)corpus_trimsentences() and char_trimsentences() to remove sentences from a corpus or character object, based on token length or pattern matching.textstat_readability(): min_sentence_length and max_sentence_length. (#632)[), or accessing values directly ([[). (#651)textstat_collocations(), which combines the existing collocations() and sequences() functions. (#434) Collocations now behave as sequences for other functions (such as tokens_compound()) and have a greatly improved performance for such uses.docvars() now permits direct access to "metadoc" fields (starting with _, e.g. _document)metadoc() now returns a vector instead of a data.frame for a single variable, similar to docvars()verbose options now take the default from getOption("verbose") rather than fixing the value in the function signatures. (#577)textstat_dist() and textstat_simil() now return a matrix if a selection argument is supplied, and coercion to a list produces a list of distances or similarities only for that selection.tokens(), the old arguments (e.g. removePunct) still produce the same behaviour but with a deprecation warning.n_target and n_reference columns to textstat_keyness() to return counts for each category being compared for keyness.str() on a corpus with no docvars (#571).removeURL in tokens() now removes URLs where the first part of the URL is a single letter (#587).dfm_select now works correctly for ngram features (#589).dfm_select(x, features) when features was a dfm, that failed to produce the intended featnames matches for the output dfm.corpus_segment(x, what = "tags") when a document contained a whitespace just before a tag, at the beginning of the file, or ended with a tag followed by no text (#618, #634).textstat_keyness() now returns a data.frame with p-values as well as the test statistic, and rownames containing the feature. This is more consistent with the other textstat functions.tokens_lookup() implements new rules for nested and linked sequences in dictionary values. See #502.tokens_compound() has a new join argument for better handling of nested and linked sequences. See #517.tokens are now significantly faster due to a reimplementation of the hash table functions in C++. (#510)dfm() now works with multi-word dictionaries and thesauruses, which previously worked only with tokens_lookup().fcm() is now parallelized for improved performance on multi-core systems.convert(x, to = "lsa") that transposed row and column names (#526)fcm() method for corpus objects (#538)dfm and tokens to break on > 10,000 documents. (#438)tokens(x, what = "character", removeSeparators = TRUE) that returned an empty string.corpus.VCorpus if the VCorpus contains a single document. (#445)dfm_compress in which the function failed on documents that contained zero feature counts. (#467)textmodel_NB that caused the class priors Pc to be refactored alphabetically instead of in the order of assignment (#471), also affecting predicted classes (#476).textstat_keyness() discovers words that occur at differential rates between partitions of a dfm (using chi-squared, Fisher's exact test, and the G^2 likelihood ratio test to measure the strength of associations).data_corpus_inaugual and data_char_inaugural).groups argument in texts() (and in dfm() that uses this function), which will now coerce to a factor rather than requiring one.sequences(): ordered and max_length, the latter to prevent memory leaks from extremely long sequences.dictionary() now accepts YAML as an input file format.dfm_lookup and tokens_lookup now accept a levels argument to determine which level of a hierarchical dictionary should be applied.min_nchar and max_nchar arguments to dfm_select.dictionary() can now be called on the argument of a list() without explicitly wrapping it in list().fcm now works directly on a dfm object when context = "documents".This release has some major changes to the API, described below.
new name | original name | notes
:--------|:------------- |:-----
data_char_sampletext | exampleString |
data_char_mobydick | mobydickText|
data_dfm_LBGexample | LBGexample |
data_char_sampletext | exampleString |
The following objects have been renamed, but will not affect user-level functionality because they are primarily internal. Their man pages have been moved to a common ?data-internal man page, hidden from the index, but linked from some of the functions that use them.
new name | original name | notes
:--------|:------------- |:-----
data_int_syllables | englishSyllables | (used by textcount_syllables())
data_char_wordlists | wordlists | (used by readability())
data_char_stopwords | .stopwords | (used by stopwords()
In v.0.9.9 the old names remain available, but are deprecated.
new name | original name | notes
:--------|:------------- |:-----
data_char_ukimmig2010 | ukimmigTexts |
data_corpus_irishbudget2010 | ie2010Corpus |
data_char_inaugural | inaugTexts |
data_corpus_inaugural | inaugCorpus |
The following functions will still work, but issue a deprecation warning:
new function | deprecated function | constructs:
:--------|:------------- |:-------
tokens | tokenize() | tokens class object
corpus_subset | subset.corpus | corpus class object
corpus_reshape | changeunits | corpus class object
corpus_sample | sample| corpus class object
corpus_segment | segment| corpus class object
dfm_compress | compress | dfm class object
dfm_lookup | applyDictionary | dfm class object
dfm_remove | removeFeatures.dfm | dfm class object
dfm_sample | sample.dfm | dfm class object
dfm_select | selectFeatures.dfm | dfm class object
dfm_smooth | smoother | dfm class object
dfm_sort | sort.dfm | dfm class object
dfm_trim | trim.dfm | dfm class object
dfm_weight | weight | dfm class object
textplot_wordcloud | plot.dfm | (plot)
textplot_xray | plot.kwic | (plot)
textstat_readability | readability | data.frame
textstat_lexdiv | lexdiv | data.frame
textstat_simil | similarity | dist
textstat_dist | similarity | dist
featnames | features | character
nsyllable | syllables | (named) integer
nscrabble | scrabble | (named) integer
tokens_ngrams | ngrams | tokens class object
tokens_skipgrams | skipgrams | tokens class object
tokens_toupper | toUpper.tokens, toUpper.tokenizedTexts | tokens, tokenizedTexts
tokens_tolower | toLower.tokens, toLower.tokenizedTexts | tokens, tokenizedTexts
char_toupper | toUpper.character, toUpper.character | character
char_tolower | toLower.character, toLower.character | character
tokens_compound | joinTokens, phrasetotoken | tokens class object
The following are new to v0.9.9 (and not associated with deprecated functions):
new function | description | output class
:--------|:------------- |:-------
fcm() | constructor for a feature co-occurrence matrix | fcm
fcm_select | selects features from an fcm | fcm
fcm_remove | removes features from an fcm | fcm
fcm_sort | sorts an fcm in alphabetical order of its features| fcm
fcm_compress | compacts an fcm | fcm
fcm_tolower | lowercases the features of an fcm and compacts | fcm
fcm_toupper | uppercases the features of an fcm and compacts | fcm
dfm_tolower | lowercases the features of a dfm and compacts | dfm
dfm_toupper | uppercases the features of a dfm and compacts | dfm
sequences | experimental collocation detection | sequences
new name | reason
:--------|:-------------
encodedTextFiles.zip | moved to the readtext package
describeTexts | deprecated several versions ago for summary.character
textfile | moved to package readtext
encodedTexts | moved to package readtext, as data_char_encodedtexts
findSequences | replaced by sequences
to = "lsa" functionality added to convert() (#414)valuetype matches work for many functions.View methods for kwic objects, based on Javascript Datatables.kwic is completely rewritten, now uses fast hashed index matching in C++ and fully implements vectorized matches (#306) and all valuetypes (#307).tokens_lookup, tokens_select, and tokens_remove are faster and use parallelization (based on the TBB library).textstat_dist and textstat_simil add fast, sparse, and parallel computation of many new distance and similarity matrices.textmodel_wordshoal fitting function.max_docfreq and min_docfreq arguments, and better verbose output, to dfm_trim (#383).tokens(), for more memory-efficient token hashing when dealing with very large numbers of documents.corpus() through the metacorpus list argument.[, [[, and $ for (hashed) tokens objects.collocations() and kwic().tokens_select() (formerly selectFeatures.tokens()).ngrams() and joinTokens() performance for hashed tokens class objects.dfm.character() by using new tokens() constructor to create hashed tokenized texts by default when creating a dfm, resulting in performance gains when constructing a dfm. Creating a dfm from a hashed tokens object is now 4-5 times faster than the older tokenizedTexts object.tokens class object.textmodel_wordscores objects.tokens_lookup() method (formerly applyDictionary()), that also works with
dictionaries that have multi-word keys. Addresses but does not entirely yet solve #188.sparsity() function to compute the sparsity of a dfm.fcm).selectFeatures.tokenizedTexts().rbind.dfm().textfile(). (#147)plot.kwic(). (#146)convert(x, to = "stm") for dfm export, including adding an argument for meta-data (docvars, in quanteda parlance). (#209)textfile(), now supports more file types, more wildcard patterns, and is far more robust generally.format keyword for loading dictionaries (#227)messages() to display messages rather than print or catpunctuation argument to collocations() to provide new options for handling collocations separated by punctuation characters (#220).fcm(x, tri = TRUE) temporarily created a dense logical matrix.fcm).selectFeatures.dfm() that ignored case_insensitive = TRUE settings (#251)
correct the documentation for this function.tf(x, scheme = "propmax") that returned a wrong computation;
correct the documentation for this function.phrasetotoken() where if pattern included a + for valuetype = c("glob", "fixed") it threw a regex error. #239textfile() where source is a remote .zip set. (#172)wordstem.dfm() that caused an error if supplied a dfm with a feature whose total frequency
count was zero, or with a feature whose total docfreq was zero. Fixes #181.wordstem.dfm(), introduced in fixing #181.toLower = argument in dfm.tokenizedTexts().textfile (#221).dictionary() now works correctly when reading LIWC dictionaries where all terms belong to one key (#229).warn = FALSE to the readLines() calls in textfile(), so that no warnings are issued when files are read that are missing a final EOL or that contain embedded nuls.trim() now prints an output message even when no features are removed (#223)Improved Naive Bayes model and prediction, textmodel(x, y, method = "NB"), now works correctly on k > 2.
Improved tag handling for segment(x, what = "tags")
Added valuetype argument to segment() methods, which allows faster and more robust segmentation on large texts.
corpus() now converts all hyphen-like characters to simple hyphen
segment.corpus() now preserves all existing docvars.
corpus documentation now removes the description of the corpus object's structure since too many users were accessing these internal elements directly, which is strongly discouraged, as we are likely to change the corpus internals (soon and often). Repeat after me: "encapsulation".
Improve robustness of corpus.VCorpus() for constructing a corpus from a tm Corpus object.
Add UTF-8 preservation to ngrams.cpp.
Fix encoding issues for textfile(), improve functionality.
Added two data objects: Moby Dick is now available as mobydickText, without needing to access a zipped text file; encodedTextFiles.zip is now a zipped archive of different encodings of (mainly) the UN Declaration of Human Rights, for testing conversions from 8-bit encodings in different (non-Roman) languages.
phrasetotoken() now has a method correctly defined for corpus class objects.
lexdiv() now works just like readability(), and is faster (based on data.table) and the code is simpler.
removed quanteda::df() as a synonym for docfreq(), as this conflicted with stats::df().
added version information when package is attached.
improved rbind() and cbind() methods for dfm. Both now take any length sequence of dfms and perform better type checking.
rbind.dfm() also knits together dfms with different features, which can be useful for information and retrieval purposes
or machine learning.
selectFeatures(x, anyDfm) (where the second argument is a dfm) now works with a selection = "remove" option.
tokenize.character adds a removeURL option.
added a corpus method for data.frame objects, so that a corpus can be constructed directly from a data.frame. Requires the addition of a textField argument (similar to textfile).
added compress.dfm() to combine identically named columns or rows. #123
Much better phrasetotoken(), with additional methods for all combinations of corpus/character v. dictionary/character/collocations.
Added a weight(x, type, ...) signature where the second argument can be a named numeric vector of weights, not just a label for a type of weight. Thanks https://stackoverflow.com/questions/36815926/assigning-weights-to-different-features-in-r/36823475#36823475.
as.data.frame for dfms now passes ... to as.data.frame.matrix.
Fixed bug in predict.fitted_textmodel_NB() that caused a failure with k > 2 classes (#129)
Improved dfm.tokenizedTexts() performance by taking care of zero-token documents more efficiently.
dictionary(file = "liwc_formatted_dict.dic", format = "LIWC") now handles poorly formatted dictionary files better, such as the Moral Foundations Dictionary in the examples for ?dictionary.
added as.tokenizedTexts to coerce any list of characters to a tokenizedTexts object.
Fix bug in phrasetotoken, signature 'corpus,ANY' that was causing an infinite loop.
Fixed bug introduced in commit b88287f (0.9.5-26) that caused a failure in dfm() with empty (zero-token) documents. Also fixes Issue #168.
Fixed bug that caused dfm() to break if no features or only one feature was found.
Fixed bug in predict.fitted_textmodel_NB() that caused a failure with k > 2 classes (#129)
Fixed a false-alarm warning message in textmodel_wordfish()
Argument defaults for readability.corpus() now same as readability.character(). Fixes #107.
Fixed a bug causing LIWC format dictionary imports to fail if extra characters followed the closing % in the file header.
Fixed a bug in applyDictionary(x, dictionary, exclusive = FALSE) when the dictionary produced no matches at all, caused by an attempt to negative index a NULL. #115
Fixed #117, a bug where wordstem.tokenizedTexts() removed attributes from the object, causing a failure of dfm.tokenizedTexts().
Fixed #119, a bug in selectFeatures.tokenizedTexts(x, features, selection = "remove") that returned a NULL for a document's tokens when no matching pattern for removal was found.
Improved the behaviour of the removeHyphens option to tokenize() when what = "fasterword" or what "fastestword".
readability() now returns measures in order called, not function definition order.
textmodel(x, model = "wordfish") now removes zero-frequency documents and words prior to calling Rcpp.
Fixed a bug in sample.corpus() that caused an error when no docvars existed. #128
Added presidents' first names to inaugCorpus
Added textmodel implementation of multinomial and Bernoulli Naive Bayes.
Improved documentation.
Added c.corpus() method for concatenating arbitrarily large sets of corpus objects.
Default for similarity() is now margin = "documents" -- prevents overly massive results if
selection = NULL.
Defined rowMeans() and colMeans() methods for dfm objects.
Enhancements to summary.character() and summary.corpus(): Added n = to summary.character(); added pass-through options to tokenize() in summary.corpus() and summary.character() methods; added toLower as an argument to both.
Enhancements to corpus object indexing, including [[ and [[<-.
Fixed a bug preventing smoother() from working.
Fixed a bug in segment.corpus(x, what = "tag") that was failing to recover the tag values after the first text.
Fix bug in plot.dfm(x, comparison = TRUE) method causing warning about rowMeans() failing.
Fixed an issue for mfdict <- dictionary(file = "http://ow.ly/VMRkL", format = "LIWC") causing it to fail
because of the irregular combination of tabs and spaces in the dictionary file.
Fixed an exception thrown by wordstem.character(x) if one element of x was NA.
dfm() on a text or tokenized text containing an NA element now returns a row with 0 feature counts. Previously it returned a count of 1 for an NA feature.
Fix issue #91 removeHyphens = FALSE not working in tokenise for some multiple intra-word hyphens, such as "one-of-a-kind"
Fixed a bug in as.matrix.similMatrix() that caused scrambled conversion when feature sets compared were unequal, which
normally occurs when setting similarity(x, n = <something>) when n < nfeature(x)
Fixed a bug in which a corpusSource object (from textfile()) with empty docvars prevented this argument
from being supplied to corpus(corpusSourceObject, docvars = something).
Fixed inaccurate documentation for weight(), which previously listed unavailable options.
More accurate and complete documentation for tokenize().
traps an exception when calling wordstem.tokenizedTexts(x) where x was not word tokenized.
Fixed a bug in textfile() that prevented passthrough arguments in ..., such as fileEncoding = or encoding =
Fixed a bug in textfile() that caused exceptions with input documents containing docvars when there was only a single column of docvars (such as .csv files)
added new methods for similarity(), including sparse matrix computation for method = "correlation" and "cosine". (More planned soon.) Also allows easy conversion to a matrix using as.matrix() on similarity lists.
more robust implementation of LIWC-formatted dictionary file imports
better implementation of tf-idf, and relative frequency weighting, especially for very large sparse matrix objects. tf(), idf(), and tfidf() now provide relative term frequency, inverse document frequency, and tf-idf directly.
textmodel_wordfish() now accepts an integer dispersionFloor argument to constrain the phi parameter
to a minimum value (of underdispersion).
textfile() now takes a vector of filenames, if you wish to construct these yourself. See ?textfile examples.
removeFeatures() and selectFeatures.collocations() now all use a consistent interface and same underlying code, with removeFeatures() acting as a wrapper to selectFeatures().
convert(x, to = "stm") now about 3-4x faster because it uses index positions from the dgCMatrix to convert to the sparse matrix format expected by stm.
Fixed a bug in textfile() preventing encodingFrom and encodingTo from working properly.
Fixed a nasty bug problem in convert(x, to = "stm") that mixed up the word indexes. Thanks Felix Haass for
spotting this!
Fixed a problem where wordstem was not working on ngram=1 tokenized objects
Fixed toLower(x, keepAcronyms = TRUE) that caused an error when x contained no acronyms.
Creating a corpus from a tm VCorpus now works if a "document" is a vector of texts rather than a single text
Fixed a bug in texts(x, groups = MORE THAN ONE DOCVAR) that now groups correctly on combinations of multiple groups
trim() now accepts proportions in addition to integer thresholds. Also accepts a new sparsity argument, which works like tm's removeSparseTerms(x, sparse = ) (for those who really want to think of sparsity this way).
[i] and [i, j] indexing of corpus objects is now possible, for extracting texts or docvars using convenient notation. See ?corpus Details.
ngrams() and skipgrams() now use the same underlying function, with skip replacing the previous
window argument (where a skip = window - 1). For efficiency, both are now implemented in C++.
tokenize() has a new argument, removeHyphens, that controls the treatment of intra-word hyphens.
Added new measures from readability for mean syllables per word and mean words per sentence directly.
wordstem now works on ngrams (tokenizedTexts and dfm objects).
Enhanced operation of kwic(), including the definition of a kwic class object, and a plot method for this object (produces a dispersion plot).
Lots more error checking of arguments passed to ... (and potentially misspecified or misspelled). Addresses Issue #62.
Almost all methods are now methods defined for objects, from a generic.
texts(x, groups = ) now allows groups to be factors, not just document variable labels. There is a new method for texts.character(x, groups = ) which is useful for supplying a factor to concatenate character objects by group.
corrected inaccurate printing of valuetype in verbose note of selectFeatures.dfm(). (Did not affect functionality.)
fixed broken quanteda.R demo, expanded demonstration code.
removeFeatures.dfm(x, stopwords), selectFeatures.dfm(x, features), and dfm(x, ignoredFeatures) now work on objects created with ngrams. (Any ngram containing a stopword is removed.) Performance on these functions is already good but will be improved further soon.
selectFeatures(x, features =
head.dfm() and tail.dfm() methods added.
kwic() has new formals and new functionality, including a completely flexible set of matching for phrases, as well as control over how the texts and matching keyword(s) are tokenized.
segment(x, what = "sentence"), and changeunits(x, to = "sentences") now uses tokenize(x, what = "sentence"). Annoying warning messages now gone.
smoother() and weight() formal "smooth" now changed to "smoothing" to avoid clashes with stats::smooth().
Updated corpus.VCorpus() to work with recent updates to the tm package.
added print method for tokenizedTexts
fixed signature error message caused by weight(x, "relFreq") and weight(x, "tfidf"). Both now correctly produce objects of class dfmSparse.
fixed bug in dfm(, keptFeatures = "whatever") that passed it through as a glob rather than a regex to selectFeatures(). Now takes a regex, as per the manual description.
fixed textfeatures() for type json, where now it can call jsonlite::fromJSON() on a file directly.
dictionary(x, format = "LIWC") now expanded to 25 categories by default, and handles entries that are listed on multiple lines in .dic files, such as those distributed with the LIWC.
ngrams() rewritten to accept fully vectorized arguments for n and for window, thus implementing "skip-grams".
Separate function skipgrams() behaves in the standard "skipgram" fashion.
bigrams(), deprecated since 0.7, has been removed from the namespace.
corpus() no longer checks all documents for text encoding; rather, this is now based on a random sample of max()
wordstem.dfm() both faster and more robust when working with large objects.
toLower.NULL() now allows toLower() to work on texts with no words (returns NULL for NULL input)
textfile() now works on zip archives of *.txt files, although this may not be entirely portable.
fixed bug in selectFeatures() / removeFeatures() that returned zero features if no features were found matching removal pattern
corpus() previously removed document names, now fixed
non-portable \donttest{} examples now removed completely from all documentation
0.8.2-1: Changed R version dependency to 3.2.0 so that Mac binary would build on CRAN.
0.8.2-1: sample.corpus() now samples documents from a corpus, and sample.dfm() samples documents
or features from a dfm. trim() method for with nsample argument now calls sample.dfm().
sample.corpus() now samples documents from a corpus, and sample.dfm() samples documents
or features from a dfm. trim() method for with nsample argument now calls sample.dfm().
tokenize improvements for what = "sentence": more robust to specifying options, and does not split sentences after common abbreviations such as "Dr.", "Prof.", etc.
corpus() no longer automatically converts encodings detected as non-UTF-8, as this detection is too imprecise.
new function scrabble() computes English Scrabble word values for any text, applying any summary numerical function.
dfm() now 2x faster, replacing previous data.table matching with direct construction of sparse matrix from match().
Code is also much simpler, based on using three new functions that are also available directly:
subset.corpus() related to environments that sometimes caused the method to break if
nested in function environments.clean() is no more.addto option removed from dfm()ignoredFeatures and removeFeatures() applied to ngrams; change behaviour of stem = TRUE applied to ngrams (in dfm())ngrams.tokenizedTexts() method, replacing current ngrams(), bigrams()The workflow is now more logical and more streamlined, with a new workflow vignette as well as a design vignette explaining the principles behind the workflow and the commands that encourage this workflow. The document also details the development plans and things remaining to be done on the project.
Newly rewritten command encoding() detects encoding for character, corpus, and corpusSource objects (created by textfile). When creating a corpus using corpus(), detection is automatic to UTF-8 if an encoding other than UTF-8, ASCII, or ISO-8859-1 is detected.
The tokenization, cleaning, lower-casing, and dfm construction functions now use the stringi package, based on the ICU library. This results not only in substantial speed improvements,
but also more correctly handles Unicode characters and strings.
tokenize() and clean() now using stringi, resulting in much faster performance and more consistent behaviour across platforms.
tokenize() now works on sentences
summary.corpus() and summary.character() now use the new tokenization functions for counting tokens
dfm(x, dictionary = mydict) now uses stringi and is both more reliable and many many times faster.
phrasetotoken() now using stringi.
removeFeatures() now using stringi and fixed binary matches on tokenized texts
textfile has a new option, cache = FALSE, for not writing the data to a temporary file, but rather storing the object in memory if that is preferred.
language() is removed. (See Encoding... section above for changes to encoding().)
new object encodedTexts contains some encoded character objects for testing.
ie2010Corpus now has UTF-8 encoded texts (previously was Unicode escaped for non-ASCII characters)
texts() and docvars() methods added for corpusSource objects.
new methods for tokenizedTexts objects: dfm(), removeFeatures(), and syllables()
syllables() is now much faster, using matching through stringi and merging using data.table.
added readability() to compute (fast!) readability indexes on a text or corpus
tokenize() now creates ngrams of any length, with two new arguments: ngrams = concatenator = "_". The new arguments to tokenize() can be passed through from dfm().
fixed a problem in textfile() causing it to fail on Windows machines when loading *.txt
nsentence() was not counting sentences correctly if the text was lower-cased - now issues an error if no upper-case characters are detected. This was also causing readability() to fail.
added an ntoken() method for dfm objects.
fixed a bug wherein convert(anydfm, to = "tm") created a DocumentTermMatrix, not a
TermDocumentMatrix. Now correctly creates a TermDocumentMatrix. (Both worked
previously in topicmodels::LDA() so many users may not notice the change.)
phrasetotokens works with dictionaries and collocations, to transform multi-word expressions into single tokens in texts or corpora
dictionaries now redefined as S4 classes
improvements to collocations(), now does not include tokens that are separated by punctuation
created tokenizeOnly*() functions, for testing tokenizing separately from cleaning, and a cleanC(), where both new separate functions are implemented in C
tokenize() now has a new option, cpp=TRUE, to use a C++ tokenizer and cleaner, resulting in much faster text tokenization and cleaning, including that used in dfm()
textmodel_wordfish now implemented entirely in C for speed. No std errors yet but coming soon. No predict method currently working either.
ie2010Corpus, and exampleString now moved into quanteda (formerly were only in quantedaData because of non-ASCII characters in each - solved with native2ascii and \uXXXX encodings).
All dependencies, even conditional, to the quantedaData and austin packages have been removed.
Many major changes to the syntax in this version.
trimdfm, flatten.dictionary, the textfile functions, dictionary converters are all gone from the NAMESPACE
formals changed a bit in clean(), kwic().
compoundWords() -> phrasetotoken()
Cleaned up minor issues in documentation.
countSyllables data object renamed to englishSyllables.Rdata, and function renamed to syllables().
stopwordsGet() changed to stopwords(). stopwordsRemove() changed to removeFeatures().
new dictionary() constructor function that also does import and conversion, replacing old readWStatdict and readLIWCdict functions.
one function to read in text files, called textsource, that does the
work for different file types based on the filename extension, and works
also for wildcard expressions (that can link to directories for example)
dfm now sparse by default, implemented as subclasses of the Matrix package. Option dfm(..., matrixType="sparse") is now the default, although matrixType="dense" will still produce the old S3-class dfm based on a regular matrix, and all dfm methods will still work with this object.
Improvements to: weight(), print() for dfms.
New methods for dfms: docfreq(), weight(), summary(), as.matrix(), as.data.frame.
No more depends, all done through imports. Passes clean check. The start of our reliance more on the master branch rather than having merges from dev to master happen only once in a blue moon.
bigrams in dfm() when bigrams=TRUE and ignoredFeatures=
stopwordsRemove() now defined for sparse dfms and for collocations.
stopwordsRemove() now requires an explicit stopwords=
New engine for dfm now implemented as standard, using data.table and Matrix for fast, efficient (sparse) matrixes.
Added trigram collocations (n=3) to collocations().
Improvements to clean(): Minor fixes to clean() so that removeDigits=TRUE removes "€10bn" entirely and not just the "€10". clean() now removes http and https URLs by default, although does not preserve them (yet). clean also handles numbers better, to remove 1,000,000 and 3.14159 if removeDigits=TRUE but not crazy8 or 4sure.
dfm works for documents that contain no features, including for dictionary counts. Thanks to Kevin Munger for catching this.
first cut at REST APIs for Twitter and Facebook
some minor improvements to sentence segmentation
improvements to package dependencies and imports - but this is ongoing!
Added more functions to dfms, getting there...
Added the ability to segment a corpus on tags (e.g. ##TAG1 text text, ##TAG2) and have the document split using the tags as a delimiter and the tag then added to the corpus as a docvar.
added textmodel_lda support, including LDA, CTM, and STM. Added a converter dfm2stmformat() between dfm and stm's input format.
as.dfm works now for data.frame objects
added Arabic to list of stopwords. (Still working on a stemmer for Arabic.)
The first appearance of dfms(), to create a sparse Matrix using the Matrix package. Eventually this will become the default format for all but small dfms. Not only is this far more efficient, it is also much faster.
Minor speed gains for clean() -- but still much more work to be done with clean().
started textmodel_wordfish, textmodel_ca. textmodel_wordfish takes an mcmc argument that calls JAGS wordfish.
now depends on ca, austin rather than importing them
dfm subsetting with [,] now works
docnames()[], []<-, docvars()[] and []<- now work correctly
Added textmodel for scaling and prediction methods, including for starters, wordscores and naivebayes class models. LIKELY TO BE BUGGY AND QUIRKY FOR A WHILE.
Added smoothdfm() and weight() methods for dfms.
Fixed a bug in segmentSentence().
added compoundWords() to turn space-delimited phrases into single "tokens". Works with dfm(, dictionary=) if the text has been pre-processed with compoundWords() and the dictionary joins phrases with the connector ("_"). May add this functionality to be more automatic in future versions.
new keep argument for trimdfm() now takes a regular expression for which feature labels to retain. New defaults for minDoc and minCount (1 each).
added nfeature() method for dfm objects.
thesaurus: works to record equivalency classes as lists of words or regular expressions for a given key/label.
keep: regular expression pattern match for features to keep
added readLIWCdict() to read LIWC-formatted dictionaries
fixed a "bug"/feature in readWStatDict() that eliminated wildcards (and all other punctuation marks) - now only converts to lower.
improved clean() functions to better handle Twitter, punctuation, and removing extra whitespace
fixed broken dictionary option in dfm()
fixed a bug in dfm() that was preventing clean() options from being passed through
added Dice and point-wise mutual information as association measures for collocations()
added: similarity() to implement similarity measures for documents or features as vector representations
begun: implementing dfm resample methods, but this will need more time to work.
(Solution: a three way table where the third dim is the resampled text.)
added is.resample() for dfm and corpus objects
added Twitter functions: getTweets() performs a REST search through twitteR, corpus.twitter creates a corpus object with test and docvars form each tweet (operational but needs work)
added various resample functions, including making dfm a multi-dimensional object when created from a resampled corpus and dfm(, bootstrap=TRUE).
modified the print.dfm() method.
updated corpus.directory to allow specification of the file extension mask
updated docvars<- and metadoc<- to take the docvar names from the assigned data.frame if field is omitted.
added field to docvars()
enc argument in corpus() methods now actually converts from enc to "UTF-8"
started working on clean to give it exceptions for @ # _ for twitter text and to allow preservation of underscores used in bigrams/collocations.
Added: a + method for corpus objects, to combine a corpus using this operator.
Changed and fixed: collocations(), which was not only fatally slow and inefficient, but also wrong. Now is much faster and O(n) because it uses data.table and vector operations only.
Added: resample() for corpus texts.
added statLexdiv() to compute the lexical diversity of texts from a dfm.
minor bug fixes; update to print.corpus() output messages.
added a wrapper function for SnowballC::wordStem, called wordstem(), so that this can be imported without loading the whole package.
Added a corpus constructor method for the VCorpus class object from the tm package.
added zipfiles() to unzip a directory of text files from disk or a URL, for easy import into a corpus using corpus.directory(zipfiles())
Fixed all the remaining issues causing warnings in R CMD CHECK, now all are fixed.
Mostly these related to documentation.
Fixed corpus.directory to better implementing naming of docvars, if found.
Moved twitter.R to the R_NEEDFIXING until it can be made to pass tests. Apparently setup_twitter_oauth() is deprecated in the latest version of the twitteR package.
plot.dfm method for producing word clouds from dfm objects
print.dfm, print.corpus, and summary.corpus methods now defined
new accessor functions defined, such as docnames(), settings(), docvars(), metadoc(), metacorpus(), encoding(), and language()
replacement functions defined that correspond to most of the above accessor functions, e.g. encoding(mycorpus) <- "UTF-8"
segment(x, to=c("tokens", "sentences", "paragraphs", "other", ...) now provides an easy and powerful method for segmenting a corpus by units other than just tokens
a settings() function has been added to manage settings that would commonly govern how texts are converted for processing, so that these can be preserved in a corpus and applied to operations that are relevant. These settings also propagate to a dfm for both replication purposes and to govern operations for which they would be relevant, when applied to a dfm.
better ways now exist to manage corpus internals, such as through the accessor functions, rather than trying to access the internal structure of the corpus directly.
basic functions such as tokenize(), clean(), etc are now faster, neater, and operate generally on vectors and return consistent object types
the corpus object has been redesigned with more flexible components, including a settings list, better corpus-level metadata, and smarter implementation of document-level attributes including user-defined variables (docvars) and document- level meta-data (metadoc)
the dfm now has a proper class definition, including additional attributes that hold the settings used to produce the dfm.
all important functions are now defined as methods for classes of built-in (e.g. character) objects, or quanteda objects such as a corpus or dfm. Lots of functions operate on both, for instance dfm.corpus(x) and dfm.character(x).
all functions are now documented and have working examples
quanteda.pdf provides a pdf version of the function documentation in one easy-to-access document