textmodel_cnnlstmemb: [Experimental] Convolutional NN + LSTM model fitted to word...

In quanteda/quanteda.classifiers: Models for supervised text classification

[Experimental] Convolutional NN + LSTM model fitted to word embeddings

Description

A function that combines a convolutional neural network layer with a long short-term memory layer. It is designed to incorporate word sequences, represented as sequentially ordered word embeddings, into text classification. The model takes as an input a quanteda tokens object.

Usage

textmodel_cnnlstmemb(
  x,
  y,
  dropout1 = 0.2,
  dropout2 = 0.2,
  dropout3 = 0.2,
  dropout4 = 0.2,
  wordembeddim = 30,
  cnnlayer = TRUE,
  filter = 48,
  kernel_size = 5,
  pool_size = 4,
  units_lstm = 128,
  words = NULL,
  maxsenlen = 100,
  optimizer = "adam",
  loss = "categorical_crossentropy",
  metrics = "categorical_accuracy",
  ...
)

Arguments

x	tokens object
y	vector of training labels associated with each document identified in train. (These will be converted to factors if not already factors.)
dropout1	A floating variable bound between 0 and 1. It determines the rate at which units are dropped for the linear transformation of the inputs for the embedding layer.
dropout2	A floating variable bound between 0 and 1. It determines the rate at which units are dropped for the linear transformation of the inputs for the CNN layer.
dropout3	A floating variable bound between 0 and 1. It determines the rate at which units are dropped for the linear transformation of the inputs for the recurrent layer.
dropout4	A floating variable bound between 0 and 1. It determines the rate at which units are dropped for the linear transformation of the inputs for the recurrent layer.
wordembeddim	The number of word embedding dimensions to be fit
cnnlayer	A logical parameter that allows user to include or exclude a convolutional layer in the neural network model
filter	The number of output filters in the convolution
kernel_size	An integer or list of a single integer, specifying the length of the 1D convolution window
pool_size	Size of the max pooling windows. keras::layer_max_pooling_1d()
units_lstm	The number of nodes of the lstm layer
words	The maximum number of words used to train model. Defaults to the number of features in x
maxsenlen	The maximum sentence length of training data
optimizer	optimizer used to fit model to training data, see keras::compile.keras.engine.training.Model()
loss	objective loss function, see keras::compile.keras.engine.training.Model()
metrics	metric used to train algorithm, see keras::compile.keras.engine.training.Model()
...	additional options passed to keras::fit.keras.engine.training.Model()

See Also

save.textmodel_cnnlstmemb(), load.textmodel_cnnlstmemb()

Examples

## Not run: 
# create dataset with evenly balanced coded & uncoded immigration sentences
corpcoded <- corpus_subset(data_corpus_manifestosentsUK,
                           !is.na(crowd_immigration_label))
corpuncoded <- data_corpus_manifestosentsUK %>%
    corpus_subset(is.na(crowd_immigration_label) & year > 1980) %>%
    corpus_sample(size = ndoc(corpcoded))
corp <- corpcoded + corpuncoded

tok <- tokens(corp)

tmod <- textmodel_cnnlstmemb(tok,
                             y = docvars(tok, "crowd_immigration_label"),
                             epochs = 5, verbose = 1)

newdata = tokens_subset(tok, subset = is.na(crowd_immigration_label))
pred <- predict(tmod, newdata = newdata)
table(pred)
tail(texts(corpuncoded)[pred == "Immigration"], 10)


## End(Not run)

quanteda/quanteda.classifiers documentation built on Oct. 20, 2023, 6:53 a.m.