Average Embeddings

Calculate the (weighted) average of multiple embeddings.

Usage

average_embedding(x, w = NULL, method = "mean", ...)

Arguments

x

an embeddings object or list of embeddings objects

w

optional weighting for rows in x. This can be an unnamed numeric vector with one item per row of x, a named numeric vector of any length with names that match the row names of x, "trillion_word" (125,000 English word frequencies from Peter Norvig's compilation, derived from the Google Web Trillion Word Corpus), or "trillion_word_sif" for smooth inverse frequencies (SIF) calculated using the same list.

method

method to use for averaging. "mean" (the default) is the standard arithmetic mean. "median" is the geometric median (also called spatial median or L1-median), computed using Gmedian::Gmedian() or, if weights are provided, Gmedian::Weiszfeld(). "sum" is the (weighted) sum.

...

additional arguments to be passed to the averaging function

Details

For w = "trillion_word" or w = "trillion_word_sif", tokens that do not appear in the word frequency list are treated as if they appeared as often as the least frequent word in the list. If w is a named vector, rows that do not match any items in the vector will be assigned the minimum value of that vector.

Value

A named numeric vector. If x is a list, the function will be called recursively and output a list of the same length.

Examples

happy_dict <- c("happy", "joy", "smile", "enjoy")
happy_dict_embeddings <- predict(glove_twitter_25d, happy_dict)

happy_dict_vec <- average_embedding(happy_dict_embeddings)
happy_dict_vec_weighted <- average_embedding(happy_dict_embeddings, w = "trillion_word")

happy_dict_list <- find_nearest(glove_twitter_25d, happy_dict, each = TRUE)
happy_dict_vec_list <- average_embedding(happy_dict_list)