Learn Text 'Embeddings' with 'Starspace'.

ruimtehol: R package to Embed All the Things! using StarSpace

This repository contains an R package which wraps the StarSpace C++ library (https://github.com/facebookresearch/StarSpace), allowing the following:

• Text classification
• Learning word, sentence or document level embeddings
• Finding sentence or document similarity
• Ranking web documents
• Content-based recommendation (e.g. recommend text/music based on the content)
• Collaborative filtering based recommendation (e.g. recommend text/music based on interest)
• Identification of entity relationships

Installation

• For regular users, install the package from your local CRAN mirror install.packages("ruimtehol")
• For installing the development version of this package: devtools::install_github("bnosac/ruimtehol", build_vignettes = TRUE)

Look to the vignette and the documentation of the functions

vignette("ground-control-to-ruimtehol", package = "ruimtehol")
help(package = "ruimtehol")

Main functionalities

This R package allows to Build Starspace models on your own text / Get embeddings of words/ngrams/sentences/documents/labels / Get predictions from a model (e.g. classification / ranking) / Get nearest neighbours similarity

The following functions are made available.

Example

Short example showing word embeddings

library(ruimtehol)
set.seed(123456789)

## Get some training data
download.file("https://s3.amazonaws.com/fair-data/starspace/wikipedia_train250k.tgz", "wikipedia_train250k.tgz")
x <- readLines("wikipedia_train250k.tgz", encoding = "UTF-8")
x <- x[-c(1:9)]
x <- x[sample(x = length(x), size = 10000)]
writeLines(text = x, sep = "\n", con = "wikipedia_train10k.txt")

## Train
set.seed(123456789)
model <- starspace(file = "wikipedia_train10k.txt", fileFormat = "labelDoc", dim = 10, trainMode = 3)
model

Object of class textspace
 dimension of the embedding: 10
 training arguments:
      loss: hinge
      margin: 0.05
      similarity: cosine
      epoch: 5
      adagrad: TRUE
      lr: 0.01
      termLr: 1e-09
      norm: 1
      maxNegSamples: 10
      negSearchLimit: 50
      p: 0.5
      shareEmb: TRUE
      ws: 5
      dropoutLHS: 0
      dropoutRHS: 0
      initRandSd: 0.001

embedding <- as.matrix(model)
embedding[c("school", "house"), ]

              [,1]         [,2]        [,3]        [,4]         [,5]        [,6]       [,7]       [,8]         [,9]       [,10]
school 0.008395348  0.002858619 0.004770191 -0.03791502 -0.016193179 0.008368539 -0.0221493 0.01587386 -0.002012054 0.029385706
house  0.005371093 -0.007831781 0.010563998  0.01040361  0.000616577 0.005770847 -0.0097075 0.01678141 -0.004738560 0.009139475
dictionary <- starspace_dictionary(model)

## Save trained model as a binary file or as TSV so that you can inspect the embeddings e.g. with data.table::fread("wikipedia_embeddings.tsv")
starspace_save_model(model, file = "textspace.ruimtehol",      method = "ruimtehol")
starspace_save_model(model, file = "wikipedia_embeddings.tsv", method = "tsv-data.table")
## Load a pre-trained model or pre-trained embeddings
model <- starspace_load_model("textspace.ruimtehol",      method = "ruimtehol")
model <- starspace_load_model("wikipedia_embeddings.tsv", method = "tsv-data.table", trainMode = 3)

## Get the document embedding
starspace_embedding(model, "get the embedding of a full document")

                                          [,1]        [,2]      [,3]       [,4]      [,5]      [,6]       [,7]      [,8]     [,9]     [,10]
get the embedding of a full document 0.1489144 -0.09543591 0.1242385 -0.1080941 0.6971645 0.3131362 -0.3405705 0.3293449 0.231894 -0.281555

The following functionalities do similar things. They see what is the closest word or sentence to a provided sentence.

## What is closest term from the dictionary
starspace_knn(model, "What does this bunch of text look like", k = 10)

## What is closest sentence to vector of sentences
predict(model, newdata = "what does this bunch of text look like", 
        basedoc = c("what does this bunch of text look like", 
                    "word abracadabra was not part of the dictionary", 
                    "give me back my mojo",
                    "cosine distance is what i show"))
                    
## Get cosine distance between 2 sentence vectors
embedding_similarity(
  starspace_embedding(model, "what does this bunch of text look like"),
  starspace_embedding(model, "word abracadabra was not part of the dictionary"), 
  type = "cosine")

Short example showing classification modelling (tagspace)

Below Starspace is used for classification

library(ruimtehol)
data("dekamer", package = "ruimtehol")
dekamer$x <- strsplit(dekamer$question, "\\W")
dekamer$x <- sapply(dekamer$x, FUN = function(x) paste(setdiff(x, ""), collapse = " "))
dekamer$x <- tolower(dekamer$x)
dekamer$y <- strsplit(dekamer$question_theme, split = ",")
dekamer$y <- lapply(dekamer$y, FUN=function(x) gsub(" ", "-", x))

set.seed(123456789)
model <- embed_tagspace(x = dekamer$x, y = dekamer$y,
                        dim = 50, 
                        lr = 0.01, epoch = 40, loss = "softmax", adagrad = TRUE, 
                        similarity = "cosine", negSearchLimit = 50,
                        ngrams = 2, minCount = 2)
plot(model)                        
            
text <- c("de nmbs heeft het treinaanbod uitgebreid via onteigening ...",
          "de migranten komen naar europa de asielcentra ...")                   
predict(model, text, k = 3)  
predict(model, "koning filip", k = 10, type = "knn")
predict(model, "koning filip", k = 10, type = "embedding")

Notes

• Why did you call the package ruimtehol? Because that is the translation of StarSpace in WestVlaams.
• The R wrapper is distributed under the Mozilla Public License 2.0. The package contains a copy of the StarSpace C++ code (namely all code under src/Starspace) which has a BSD license (which is available in file LICENSE.notes) and also has an accompanying PATENTS file which you can inspect here.

Support in text mining

Need support in text mining? Contact BNOSAC: http://www.bnosac.be.