ruimtehol (version 0.3.2)

embed_entityrelationspace: Build a Starspace model for entity relationship completion

Description

Build a Starspace model for entity relationship completion (graphspace).

Usage

embed_entityrelationspace(
  x,
  model = "graphspace.bin",
  early_stopping = 0.75,
  useBytes = FALSE,
  ...
)

Value

an object of class textspace as returned by starspace.

Arguments

x

a data.frame with columns entity_head, entity_tail and relation indicating the relation between the head and tail entity

model

name of the model which will be saved, passed on to starspace

early_stopping

the percentage of the data that will be used as training data. If set to a value smaller than 1, 1-early_stopping percentage of the data which will be used as the validation set and early stopping will be executed. Defaults to 0.75.

useBytes

set to TRUE to avoid re-encoding when writing out train and/or test files. See writeLines for details

...

further arguments passed on to starspace except file, trainMode and fileFormat

Examples

## Example on Freebase - download the data
filename <- paste(
  "https://raw.githubusercontent.com/bnosac-dev/GraphEmbeddings/master/",
  "diffbot_data/FB15k/freebase_mtr100_mte100-train.txt",
  sep = "")
tmpfile <- tempfile(pattern = "freebase_mtr100_mte100_", fileext = "txt")
ok <- suppressWarnings(try(
  download.file(url = filename, destfile = tmpfile),
  silent = TRUE))
if(!inherits(ok, "try-error") && ok == 0){
  ## Build the model on the downloaded data
  x <- read.delim(tmpfile, header = FALSE, nrows = 1000,
                  col.names = c("entity_head", "relation", "entity_tail"),
                  stringsAsFactors = FALSE)
  head(x)

  set.seed(123456789)
  model <- embed_entityrelationspace(x, dim = 50)
  plot(model)

  predict(model, "/m/027rn /location/country/form_of_government")

  ## Also add reverse relation
  x_reverse <- x
  colnames(x_reverse) <- c("entity_tail", "relation", "entity_head")
  x_reverse$relation <- sprintf("REVERSE_%s", x_reverse$relation)

  relations <- rbind(x, x_reverse)
  set.seed(123456789)
  model <- embed_entityrelationspace(relations, dim = 50)
  predict(model, "/m/027rn /location/country/form_of_government")
  predict(model, "/m/06cx9 REVERSE_/location/country/form_of_government")
}

## cleanup for cran
if(file.exists(tmpfile)) file.remove(tmpfile)