text2vec (version 0.5.0)

RelaxedWordMoversDistance: Creates model which can be used for calculation of "relaxed word movers distance".

Description

Relaxed word movers distance tries to measure distance between documents by calculating how hard is to transform words from first document into words from second document and vice versa. For more detail see original article: http://mkusner.github.io/publications/WMD.pdf.

Usage

RelaxedWordMoversDistance
RWMD

Format

R6Class object.

Fields

progressbar

logical = TRUE whether to display progressbar

Usage

For usage details see Methods, Arguments and Examples sections.

rwmd = RelaxedWordMoversDistance$new(wv, method = c("cosine", "euclidean"), normalize = TRUE, progressbar = interactive())
rwmd$dist2(x, y)
rwmd$pdist2(x, y)

Methods

$new(wv, method = c("cosine", "euclidean"))

Constructor for RWMD model For description of arguments see Arguments section

$dist2(x, y)

Computes distance between each row of sparse matrix x and each row of sparse matrix y

$pdist2(x, y)

Computes "parallel" distance between rows of sparse matrix x and corresponding rows of the sparse matrix y

Arguments

rwmd

RWMD object

x

x sparse document term matrix

y

y = NULL sparse document term matrix. If y = NULL (as by default), we will assume y = x

wv

word vectors. Numeric matrix which contains word embeddings. Rows - words, columns - corresponding vectors. Rows should have word names.

method

name of the distance for measuring similarity between two word vectors. In original paper authors use "euclidean", however we use "cosine" by default (better from our experience). This means distance = 1 - cosine_angle_betwen_wv

Examples

# NOT RUN {
data("movie_review")
tokens = movie_review$review %>%
  tolower %>%
  word_tokenizer
v = create_vocabulary(itoken(tokens)) %>%
  prune_vocabulary(term_count_min = 5, doc_proportion_max = 0.5)
it = itoken(tokens)
vectorizer = vocab_vectorizer(v)
dtm = create_dtm(it, vectorizer)
tcm = create_tcm(it, vectorizer, skip_grams_window = 5)
glove_model = GloVe$new(word_vectors_size = 50, vocabulary = v, x_max = 10)
wv = glove_model$fit_transform(tcm, n_iter = 10)
# get average of main and context vectors as proposed in GloVe paper
wv = wv + t(glove_model$components)
rwmd_model = RWMD$new(wv)
rwmd_dist = dist2(dtm[1:100, ], dtm[1:10, ], method = rwmd_model, norm = 'none')
head(rwmd_dist)
# }