Package 'GMTM'

Convert a dictionary to a seed word matrix

Description

Convert a dictionary to a seed word matrix

Usage

as.seedwords(x, model, residual = 0, levels = 1)

Arguments

x	a quanteda::dictionary of seed words.
model	a wordvector::textmodel_word2vec object.
residual	the number of unseeded topics.
levels	integers specifying the levels of entries in a hierarchical dictionary.

Details

Word vectors of the dictionary values within the same key are all averaged as a single seed vector. If dictionary values are multi-word expressions, they are matched against individual and compounded words in the model.

Unseeded topics are labeled "other", but it can be changed via options("GMTM.residual.name").

Value

Returns a seed word matrix.

Examples

library(quanteda)
library(wordvector)
options(wordvector_threads = 2)

corp <- head(wordvector::data_corpus_news2014, 1000)
toks <- tokens(corp, remove_punct = TRUE,
               remove_symbols = TRUE, remove_numbers = TRUE) %>%
        tokens_remove(stopwords("en"), min_nchar = 2)
wov <- textmodel_word2vec(toks, dim = 50)

dict <- dictionary(list(eco = "econom*", sec = "securit*", spo = "sport*",
                        pol = "politi*", cri = "crime"))
seed <- as.seedwords(dict, wov, residual = 2)

---

Extract the probabilities for topics

Description

Extract the probabilities for topics

Usage

## S3 method for class 'textmodel_gmm'
probability(x, group = NULL, ...)

Arguments

x	a fitted model.
group	a factor to group documents and average their probability for topics. Ignored if x is a textmodel_kmeans object.
...	not used.

Value

Returns the probabilities of topics as a matrix.

---

Extract lost likely topic terms from documents

Description

Extract lost likely topic terms from documents

Usage

terms(x, data, n = 10, filter = NULL, ...)

Arguments

x	a fitted model or a factor from GMTM::topics().
data	a quanteda::dfm or quanteda::tokens from which words are extracted for each topic.
n	the number of topic words.
filter	a character vector of words to be included in the output.
...	not used.

Details

To identify topic terms, data must be provided along with a fitted model because the information about the frequency of individual words are lost in document vectors.

When x is a textmodel_gmm, data is weighted by the documents' probabilities for topics to extract most likely terms. When x is textmodel_kmeans or a factor, data is grouped by topic and weighted by TF-IDF to select the most distinctive words for each topic (known as c-TF-IDF).

Value

Returns a character matrix with the most distinctive words for each topic.

---

Topic analysis using Gaussian mixture models

Description

Perform topic analysis of document vectors using Gaussian mixture models.

Usage

textmodel_gmm(
  x,
  k = 10,
  model = NULL,
  seeds = NULL,
  omit = NULL,
  verbose = quanteda_options("verbose"),
  ...
)

Arguments

x	a wordvector::textmodel_doc2vec or a dense matrix of document vectors in the rows.
k	the number of topics to identify.
model	a fitted model from which initial centroids are extracted.
seeds	a matrix created using as.seedwords.
omit	indices of singular values of x to be zero. See the details.
verbose	print the progress if TRUE.
...	passed to the underlying function.

Details

multi-threading

Users can change the number of threads for the parallel computing via options(GMTM.threads) or OMP_THREAD_LIMIT in the environmental variable. To reproduce results, set options(GMTM.threads = 1) and call set.seed() immediately before textmodel_gmm() or textmodel_kmeans().

On MacOS, only one thread is used regardless of GMTM.threads because CRAN's toolchain for the platform does not support OpenMP.

noise reduction

omit is used to reduce the noise in the x by applying base::svd before clustering. If it is not NULL, singular values corresponding to omit are set to zero, removing their variance in x. See Chan et al. (2020) https://doi.org/10.1080/19312458.2020.1812555 for the methodology.

additional arguments

The number of iterations in k-means (iter_km) and expectation maximization (iter_em) stages can be set via ....

Value

Returns a fitted textmodel_gmm object.

Examples

library(quanteda)
library(wordvector)
options(wordvector_threads = 2)

corp <- head(wordvector::data_corpus_news2014, 1000)
toks <- tokens(corp, remove_punct = TRUE,
               remove_symbols = TRUE, remove_numbers = TRUE) %>%
        tokens_remove(stopwords("en"), min_nchar = 2)
wov <- textmodel_word2vec(toks, dim = 50)
dov <- as.textmodel_doc2vec(dfm(toks), wov)

gmm <- textmodel_gmm(dov, k = 10)
table(topics(gmm))

---

Topic analysis using k-means

Description

Perform topic analysis of document vectors using k-means.

Usage

textmodel_kmeans(
  x,
  k = 10,
  model = NULL,
  seeds = NULL,
  verbose = quanteda_options("verbose"),
  ...
)

Arguments

x	a wordvector::textmodel_doc2vec or a dense matrix of document vectors in the rows.
k	the number of topics to identify.
model	a fitted model from which initial centroids are extracted.
seeds	a matrix created using as.seedwords.
verbose	print the progress if TRUE.
...	passed to the underlying function.

Value

Returns a fitted textmodel_kmeans object.

Examples

library(quanteda)
library(wordvector)
options(wordvector_threads = 2)

corp <- head(wordvector::data_corpus_news2014, 1000)
toks <- tokens(corp, remove_punct = TRUE,
               remove_symbols = TRUE, remove_numbers = TRUE) %>%
        tokens_remove(stopwords("en"), min_nchar = 2)
wov <- textmodel_word2vec(toks, dim = 50)
dov <- as.textmodel_doc2vec(dfm(toks), wov)

km <- textmodel_kmeans(dov, k = 10)
table(topics(km))

---

Extract the most likely topics of documents

Description

Extract the most likely topics of documents

Usage

topics(x, ...)

Arguments

x	a fitted model.
...	passed to probability.textmodel_gmm().

Value

Returns predicted topics as a vector.

---