nns_ratio: Computes the ratio of cosine similarities for two embeddings...

In conText: 'a la Carte' on Text (ConText) Embedding Regression

Computes the ratio of cosine similarities for two embeddings over the union of their respective top N nearest neighbors.

Description

Computes the ratio of cosine similarities between group embeddings and features –that is, for any given feature it first computes the similarity between that feature and each group embedding, and then takes the ratio of these two similarities. This ratio captures how "discriminant" a feature is of a given group. Values larger (smaller) than 1 mean the feature is more (less) discriminant of the group in the numerator (denominator).

Usage

nns_ratio(
  x,
  N = 10,
  numerator = NULL,
  candidates = character(0),
  pre_trained,
  stem = FALSE,
  language = "porter",
  verbose = TRUE,
  show_language = TRUE
)

Arguments

x	a (quanteda) dem-class or fem-class object.
N	(numeric) number of nearest neighbors to return. Nearest neighbors consist of the union of the top N nearest neighbors of the embeddings in x. If these overlap, then resulting N will be smaller than 2*N.
numerator	(character) defines which group is the nuemerator in the ratio
candidates	(character) vector of features to consider as candidates to be nearest neighbor You may for example want to only consider features that meet a certian count threshold or exclude stop words etc. To do so you can simply identify the set of features you want to consider and supply these as a character vector in the candidates argument.
pre_trained	(numeric) a F x D matrix corresponding to pretrained embeddings. F = number of features and D = embedding dimensions. rownames(pre_trained) = set of features for which there is a pre-trained embedding.
stem	(logical) - whether to stem candidates when evaluating nns. Default is FALSE. If TRUE, candidate stems are ranked by their average cosine similarity to the target. We recommend you remove misspelled words from candidate set candidates as these can significantly influence the average.
language	the name of a recognized language, as returned by getStemLanguages, or a two- or three-letter ISO-639 code corresponding to one of these languages (see references for the list of codes).
verbose	report which group is the numerator and which group is the denominator.
show_language	(logical) if TRUE print out message with language used for stemming.

Value

a data.frame with following columns:

feature

(character) features in candidates (or all features if candidates not defined), one instance for each embedding in x.

value

(numeric) ratio of cosine similarities.

Examples

library(quanteda)

# tokenize corpus
toks <- tokens(cr_sample_corpus)

# build a tokenized corpus of contexts sorrounding a target term
immig_toks <- tokens_context(x = toks, pattern = "immigr*", window = 6L)

# build document-feature matrix
immig_dfm <- dfm(immig_toks)

# construct document-embedding-matrix
immig_dem <- dem(immig_dfm, pre_trained = cr_glove_subset,
transform = TRUE, transform_matrix = cr_transform, verbose = FALSE)

# to get group-specific embeddings, average within party
immig_wv_party <- dem_group(immig_dem, groups = immig_dem@docvars$party)

# compute the cosine similarity between each party's
# embedding and a specific set of features
nns_ratio(x = immig_wv_party, N = 10, numerator = "R",
candidates = immig_wv_party@features,
pre_trained = cr_glove_subset, verbose = FALSE)

# with stemming
nns_ratio(x = immig_wv_party, N = 10, numerator = "R",
candidates = immig_wv_party@features,
pre_trained = cr_glove_subset, stem = TRUE, verbose = FALSE)

conText documentation built on Feb. 16, 2023, 7:32 p.m.