Please check the latest news (change log) and keep this package updated.
⚠️ All users should update the package to version ≥ 0.3.0. Old versions (≤ 0.2.0) may run slowly, and some old functions have been deprecated.
[ method for embed, see new
examples in as_embed().unique() method to delete duplicate words.str() method to print the data structure and
attributes.pattern() function designed for S3 [
method of embed: Users can directly use regular expression
like embed[pattern("^for")] to extract a subset of
embedding matrix.plot_network() function: Visualize a (partial
correlation) network graph of words. Very useful for identifying
potential semantic clusters from a list of words and even useful for
disentangling antonyms from synonyms.targets argument of text_unmask():
Return specific fill-mask results for certain target words (rather than
the top n results).tab_similarity(),
most_similar(), dict_expand(),
dict_reliability(), test_WEAT(),
test_RND().print() method for embed and
wordvec.pair_similarity() has been improved by using matrix
operation tcrossprod(embed, embed) to compute cosine
similarity, with embed normalized.data_wordvec_load() has got two wrapper functions
load_wordvec() and load_embed() for faster
use.data_wordvec_normalize() (deprecated) has been renamed
to normalize().get_wordvecs() (deprecated) has been integrated into
get_wordvec().tab_similarity_cross() (deprecated) has been integrated
into tab_similarity().test_WEAT() and test_RND(): Warning if
T1 and T2 or A1 and
A2 have duplicate values.embed or wordvec, and too many words to be
printed to console. Now all related functions have been substantially
improved so that they would not take unnecessarily long time.embed (an extended
class of matrix) rather than wordvec in order to enhance
the speed!text_* functions for contextualized word
embeddings! Based on the R package text (and using the R
package reticulate to call functions from the Python module
transformers), a series of new functions have been
developed to (1) download HuggingFace Transformers
pre-trained language models (PLM; thousands of options such as
GPT, BERT, RoBERTa, DeBERTa, DistilBERT, etc.), (2) extract
contextualized token (roughly word) embeddings and text embeddings, and
(3) fill in the blank mask(s) in a query (e.g., “Beijing is the [MASK]
of China.”).
text_init(): set up a Python environment for PLMtext_model_download(): download PLMs from HuggingFace to local “.cache”
foldertext_model_remove(): remove PLMs from local “.cache”
foldertext_to_vec(): extract contextualized token and text
embeddingstext_unmask(): fill in the blank mask(s) in a
queryorth_procrustes() function: Orthogonal Procrustes
matrix alignment. Users can input either two matrices of word embeddings
or two wordvec objects as loaded by
data_wordvec_load() or transformed from matrices by
as_wordvec().dict_expand() function: Expand a dictionary from
the most similar words, based on most_similar().dict_reliability() function: Reliability analysis
(Cronbach’s α) and Principal Component Analysis (PCA) of a dictionary.
Note that Cronbach’s α may be misleading when the number of items/words
is large.sum_wordvec() function: Calculate the sum vector of
multiple words.plot_similarity() function: Visualize cosine
similarities between word pairs in a style of correlation matrix
plot.tab_similarity_cross() function: A wrapper of
tab_similarity() to tabulate cosine similarities for only
n1 * n2 word pairs from two sets of words (arguments:
words1, words2).print.wordvec(),
print.embed(), rbind.wordvec(),
rbind.embed(), subset.wordvec(),
subset.embed()as_matrix() has been renamed to
as_embed(): Now PsychWordVec supports two
classes of data objects – wordvec (data.table) and
embed (matrix). Most functions now use embed
(or transform wordvec to embed) internally so
as to enhance the speed. Matrix is much faster!data_wordvec_reshape(): Now use
as_wordvec() and as_embed().data_wordvec_subset(),
get_wordvecs(), tab_similarity(), and
plot_similarity(): If neither words nor
pattern are specified (NULL), then all words
in data will be extracted.print.weat() and
print.rnd().test_WEAT() and test_RND(): Users can specify
the number of permutation samples and choose to calculate either
one-sided or two-sided p value. It can well reproduce the
results in Caliskan et al.’s (2017) article.pooled.sd argument for
test_WEAT(): Users can choose the method used to calculate
the pooled SD for effect size estimate in WEAT. However, the
original approach proposed by Caliskan et al. (2017) is the default and
highly suggested.as_matrix() and
as_wordvec() for data_wordvec_reshape(), which
can make it easier to reshape word embeddings data from
matrix to “wordvec” data.table or vice
versa.test_WEAT() and test_RND() now have
changed the element names and S3 print method of their returned objects
(of new class weat and rnd, respectively): The
elements $eff.raw, $eff.size, and
$eff.sum are now deprecated and replaced by
$eff, which is a data.table containing the
overall raw/standardized effects and permutation p value. The
new S3 print methods print.weat() and
print.rnd() can make a tidy report of the test results when
you directly type and print the returned object (see code
examples).cli
package.library(PsychWordVec).wordvec as the primary class of word vectors
data: Now the data classes contain wordvec,
data.table, and data.frame, which actually
perform as a data.table.train_wordvec() function: Train word vectors using
the Word2Vec, GloVe, or FastText algorithm
with multi-threading.tokenize() function: Tokenize raw texts for
training word vectors.data_wordvec_reshape() function: Reshape word
vectors data from dense (a data.table of new classs
wordvec with two variables word and
vec) to plain (a matrix of word vectors) or
vice versa.test_RND() function, and tab_WEAT() is
renamed to test_WEAT(): These two functions serve as
convenient tools of word semantic similarity analysis and conceptual
association test.plot_wordvec_tSNE() function: Visualize 2-D or 3-D
word vectors with dimensionality reduced using the t-Distributed
Stochastic Neighbor Embedding (t-SNE) method.data_wordvec_subset() function.unique argument for
tab_similarity().test_WEAT().