HuggingFace Transformers in R: Word Embeddings Defaults and Specifications

A word embedding comprises values that represent the latent meaning of a word. The numbers may be seen as coordinates in a space that comprises several hundred dimensions. The more similar two words’ embeddings are, the closer positioned they are in this embedding space, and thus, the more similar the words are in meaning. Hence, embeddings reflect the relationships among words, where proximity in the embedding space represents similarity in latent meaning. The text-package enables you to use already existing Transformers (language models (from Hugging Face) to map text data to high quality word embeddings.

To represent several words, sentences and paragraphs, word embeddings of single words may be combined or aggregated into one word embedding. This can be achieved by taking the mean, minimum or maximum value of each dimension of the embeddings.

This tutorial focuses on how to retrieve layers and how to aggregate them to receive word embeddings in text. The focus will be on the actual functions.
For more detailed information about word embeddings and the language models in regard to text please see text: An R-package for Analyzing and Visualizing Human Language Using Natural Language Processing and Deep Learning; and for more comprehensive information about the inner workings of the language models, for example see Illustrated BERT or the references given in Table 1.

Table 1 show some of the more common language models; for more detailed information see HuggingFace

Models References Layers Dimensions Language
‘bert-base-uncased’ Devlin et al. 2019 12 768 English
‘roberta-base’ Liu et al. 2019 12 768 English
‘distilbert-base-cased’ Sahn et al., 2019 6? 768? English
‘bert-base-multilingual-cased’ Devlin et al.2019 12 768 104 top languages at Wikipedia
‘xlm-roberta-large’ Liu et al 24 1024 100 language

textEmbed: Reflecting standards and state-of-the-arts

The text-package has 3 functions for mapping text to word embeddings. The textEmbed() is the high-level function, which encompasses textEmbedRawLayers() and textEmbedLayerAggregation(). textEmbedRawLayers() retrieves layers and hidden states from a given language model; and textEmbedLayerAggregation() aggregates these layers in order to form word embeddings.

textEmbed()

textEmbed() selects character variables in a given dataset (a dataframe/tibble) and transforms these to word embeddings. It can output contextualized (and decontextualized) embeddings for both tokens and texts.

The Language model

Set the language language model that you want using the model parameter. The text-package automatically downloads the model from HuggingFace, the first time it is being called.

The Layers

The layers parameter controls the layer(s) to extract (default is the second to last layer). The textEmbed() function also provides parameters to aggregate the layers in various ways. The aggregation_from_layers_to_tokens parameter controls how to aggregate layers representing the same token (default is “concatenate”). The aggregation_from_tokens_to_texts parameter controls how embeddings from different tokens should be aggregated to represent a text (default = “mean”). There is also an optional setting aggregation_from_tokens_to_word_types that controls how the word types embeddings are aggregated.

library(text)
# Example text 
texts <- c("I feel great")

# Transform the text to BERT word embeddings
wordembeddings <- textEmbed(texts = texts,
                            model = 'bert-base-uncased',
                            layers = 11:12,
                            aggregation_from_layers_to_tokens = "concatenate",
                            aggregation_from_tokens_to_texts = "mean"
                            )

wordembeddings

Note that it is also possible to submit an entire dataset to textEmbed() – as well as only retrieving text-level and word-type level embeddings. This is achieved by setting keep_token_embeddings to FALSE, and aggregation_from_tokens_to_word_types to, for example, “mean”. Word type-level embeddings can be used for plotting words in the embedding space.

library(text)

# Transform the text data to BERT word embeddings
wordembeddings <- textEmbed(texts = Language_based_assessment_data_8[1:2],
                            aggregation_from_tokens_to_word_types = "mean",
                            keep_token_embeddings = FALSE)

# See how word embeddings are structured
wordembeddings

The textEmbed() function is suitable when you are just interested in getting good word embeddings to test some research hypothesis with. That is, the defaults are based on general experience of what works. Under the hood textEmbed uses one function for retrieving the layers (textEmbedRawLayers) and another function for aggregating them (textEmbedLayerAggreation). So, if you are interested in examining different layers and different aggregation methods it is better to split up the work flow so that you first retrieve all layers (which takes most time) and then test different aggregation methods.

textEmbedRawLayers: Get tokens and all the layers

The textEmbedRawLayers function is used to retrieve the layers of hidden states.

library(text)

#Transform the text data to BERT word embeddings

# Example test 
texts <- c("I feel great")
 
wordembeddings_tokens_layers <- textEmbedRawLayers(
  texts = texts,
  layers = 10:12)
wordembeddings_tokens_layers

textEmbedLayerAggreation: Testing different layers

The textEmbedLayerAggreation() function gives you the possibility to aggregate the layers in different ways (without having to retrieve them from the language model several times). In textEmbedLayerAggreation(), you can select any combination of the layers that you want to aggregate; and then you can aggregate them using the mean of the dimensions, the minimum or maximum value.

library(text)

# Aggregating layer 11 and 12 by taking the mean of each dimension. 
we_11_12_mean <- textEmbedLayerAggregation(
  word_embeddings_layers = wordembeddings_tokens_layers$context_tokens$texts,
  layers = 11:12,
  aggregation_from_layers_to_tokens = "concatenate",
  aggregation_from_tokens_to_texts = "mean")
we_11_12_mean
# Aggregating layer 11 and 12 by taking the minimum of each dimension accross the two layers.
we_10_11_min <- textEmbedLayerAggregation(
  word_embeddings_layers = wordembeddings_tokens_layers$context_tokens$texts,
  layers = 10:11,
  aggregation_from_layers_to_tokens = "concatenate",
  aggregation_from_tokens_to_texts = "min")
we_10_11_min
# Aggregating layer 1 to 12 by taking the max value of each dimension across the 12 layers.
we_11_max <- textEmbedLayerAggregation(
  word_embeddings_layers = wordembeddings_tokens_layers$context_tokens$texts,
  layers = 11,
  aggregation_from_tokens_to_texts = "max")
we_11_max

Now the word embeddings are ready to be used in down stream tasks such as predicting numeric variables or be plotted according to different dimensions.