Select and aggregate layers of hidden states to form a word embeddings.

Source: R/1_1_textEmbed.R

Select and aggregate layers of hidden states to form a word embeddings.

textLayerAggregation(
  word_embeddings_layers,
  layers = 11:12,
  aggregate_layers = "concatenate",
  aggregate_tokens = "mean",
  tokens_select = NULL,
  tokens_deselect = NULL
)

Arguments

word_embeddings_layers

Layers outputted from textHuggingFace.
layers

The numbers of the layers to be aggregated (e.g., c(11:12) to aggregate the eleventh and twelfth). Note that layer 0 is the input embedding to the transformer, and should normally not be used. Selecting 'all' thus removes layer 0.
aggregate_layers

Method to carry out the aggregation among the layers for each word/token, including "min", "max" and "mean" which takes the minimum, maximum or mean across each column; or "concatenate", which links together each layer of the word embedding to one long row. Default is "concatenate"
aggregate_tokens

Method to carry out the aggregation among the word embeddings for the words/tokens, including "min", "max" and "mean" which takes the minimum, maximum or mean across each column; or "concatenate", which links together each layer of the word embedding to one long row.
tokens_select

Option to only select embeddings linked to specific tokens such as "[CLS]" and "[SEP]" (default NULL).
tokens_deselect

Option to deselect embeddings linked to specific tokens such as "[CLS]" and "[SEP]" (default NULL).

Value

A tibble with word embeddings. Note that layer 0 is the input embedding to the transformer, which is normally not used.

See also

see textHuggingFace and textEmbed

Examples

if (FALSE) {
wordembeddings <- textLayerAggregation(word_embeddings_layers)
}