text2vec (version 0.3.0)

glove: Fit a GloVe word-embedded model

Description

This function trains a GloVe word-embeddings model via fully asynchronous and parallel AdaGrad.

Usage

glove(tcm, vocabulary_size, word_vectors_size, x_max, num_iters, shuffle_seed = NA_integer_, learning_rate = 0.05, verbose = TRUE, convergence_threshold = 0, grain_size = 100000L, max_cost = 10, alpha = 0.75, ...)
"glove"(tcm, vocabulary_size = nrow(tcm), word_vectors_size, x_max, num_iters, shuffle_seed = NA_integer_, learning_rate = 0.05, verbose = TRUE, convergence_threshold = -1, grain_size = 100000L, max_cost = 10, alpha = 0.75, ...)
"glove"(tcm, ...)

Arguments

tcm

an object which represents a term-co-occurrence matrix, which is used in training. At the moment only dgTMatrix or objects coercible to a dgTMatrix) are supported. In future releases we will add support for out-of-core learning and streaming a TCM from disk.

vocabulary_size

number of words in in the term-co-occurrence matrix

word_vectors_size

desired dimenson for word vectors

x_max

maximum number of co-occurrences to use in the weighting function. See the GloVe paper for details: http://nlp.stanford.edu/pubs/glove.pdf.

num_iters

number of AdaGrad epochs

shuffle_seed

integer seed. Use NA_integer_ to turn shuffling off. A seed defines shuffling before each SGD iteration. Generelly shuffling is a good idea for stochastic-gradient descent, but from my experience in this particular case it does not improve convergence. By default there is no shuffling. Please report if you find that shuffling improves your score.

learning_rate

learning rate for SGD. I do not recommend that you modify this parameter, since AdaGrad will quickly adjust it to optimal.

verbose

logical whether to display training inforamtion

convergence_threshold

defines early stopping strategy. We stop fitting when one of two following conditions will be satisfied: (a) we have used all iterations, or (b) cost_previous_iter / cost_current_iter - 1 < convergence_threshold.

grain_size

I do not recommend adjusting this paramenter. This is the grain_size for RcppParallel::parallelReduce. For details, see http://rcppcore.github.io/RcppParallel/#grain-size.

max_cost

the maximum absolute value of calculated gradient for any single co-occurrence pair. Try to set this to a smaller value if you have problems with numerical stability.

alpha

the alpha in weighting function formula : $f(x) = 1 if x > x_max; else (x/x_max)^alpha$

...

arguments passed to other methods (not used at the moment).

Methods (by class)

• dgTMatrix: fits GloVe model on a dgTMatrix, a sparse matrix in triplet form
• Matrix: Fits a GloVe model on a Matrix input

See Also

http://nlp.stanford.edu/projects/glove/

Examples

## Not run: 
# library(readr)
# temp <- tempfile()
# download.file('http://mattmahoney.net/dc/text8.zip', temp)
# text8 <- read_lines(unz(temp, "text8"))
# it <- itoken(text8, preprocess_function = identity,
#              tokenizer = function(x) strsplit(x, " ", TRUE))
# vocab <- vocabulary(it) %>%
#  prune_vocabulary(term_count_min = 5)
# 
# it <- itoken(text8, preprocess_function = identity,
#              tokenizer = function(x) strsplit(x, " ", TRUE))
# 
# tcm <- create_tcm(it, vocab_vectorizer(vocab, grow_dtm = FALSE, skip_grams_window = 5L))
# 
# # use the following command to manually set number of threads (if you want)
# # by default glove will use all available CPU cores
# # RcppParallel::setThreadOptions(numThreads = 8)
# fit <- glove(tcm = tcm, shuffle_seed = 1L, word_vectors_size = 50,
#               x_max = 10, learning_rate = 0.2,
#               num_iters = 50, grain_size = 1e5,
#               max_cost = 100, convergence_threshold = 0.005)
# word_vectors <- fit$word_vectors[[1]] + fit$word_vectors[[2]]
# rownames(word_vectors) <- rownames(tcm)
# qlst <- prepare_analogy_questions('./questions-words.txt', rownames(word_vectors))
# res <- check_analogy_accuracy(questions_lst = qlst, m_word_vectors = word_vectors)
# ## End(Not run)

Run the code above in your browser using DataCamp Workspace