textstat_keyness {quanteda.textstats}

Calculate keyness statistics

Description

Calculate "keyness", a score for features that occur differentially across different categories. Here, the categories are defined by reference to a "target" document index in the dfm, with the reference group consisting of all other documents.

Usage

textstat_keyness(
  x,
  target = 1L,
  measure = c("chi2", "exact", "lr", "pmi"),
  sort = TRUE,
  correction = c("default", "yates", "williams", "none"),
  ...
)

Arguments

x

a dfm containing the features to be examined for keyness

target

the document index (numeric, character or logical) identifying the document forming the "target" for computing keyness; all other documents' feature frequencies will be combined for use as a reference

measure

(signed) association measure to be used for computing keyness. Currently available: "chi2"; "exact" (Fisher's exact test); "lr" for the likelihood ratio; "pmi" for pointwise mutual information. Note that the "exact" test is very computationally intensive and therefore much slower than the other methods.

sort

logical; if TRUE sort features scored in descending order of the measure, otherwise leave in original feature order

correction

if "default", Yates correction is applied to "chi2"; William's correction is applied to "lr"; and no correction is applied for the "exact" and "pmi" measures. Specifying a value other than the default can be used to override the defaults, for instance to apply the Williams correction to the chi2 measure. Specifying a correction for the "exact" and "pmi" measures has no effect and produces a warning.

...

not used

Value

a data.frame of computed statistics and associated p-values, where the features scored name each row, and the number of occurrences for both the target and reference groups. For measure = "chi2" this is the chi-squared value, signed positively if the observed value in the target exceeds its expected value; for measure = "exact" this is the estimate of the odds ratio; for measure = "lr" this is the likelihood ratio G2 statistic; for "pmi" this is the pointwise mutual information statistics.

textstat_keyness returns a data.frame of features and their keyness scores and frequency counts.

References

Bondi, M. & Scott, M. (eds) (2010). Keyness in Texts. Amsterdam, Philadelphia: John Benjamins.

Stubbs, M. (2010). Three Concepts of Keywords. In Keyness in Texts, Bondi, M. & Scott, M. (eds): 1–42. Amsterdam, Philadelphia: John Benjamins.

Scott, M. & Tribble, C. (2006). Textual Patterns: Keyword and Corpus Analysis in Language Education. Amsterdam: Benjamins: 55.

Dunning, T. (1993). Accurate Methods for the Statistics of Surprise and Coincidence. Computational Linguistics, 19(1): 61–74.

Examples

library("quanteda")

# compare pre- v. post-war terms using grouping
period <- ifelse(docvars(data_corpus_inaugural, "Year") < 1945, "pre-war", "post-war")
dfmat1 <- tokens(data_corpus_inaugural) %>%
    dfm() %>%
    dfm_group(groups = period)
head(dfmat1) # make sure 'post-war' is in the first row
head(tstat1 <- textstat_keyness(dfmat1), 10)
tail(tstat1, 10)

# compare pre- v. post-war terms using logical vector
dfmat2 <- dfm(tokens(data_corpus_inaugural))
head(textstat_keyness(dfmat2, docvars(data_corpus_inaugural, "Year") >= 1945), 10)

# compare Trump 2017 to other post-war preseidents
dfmat3 <- dfm(tokens(corpus_subset(data_corpus_inaugural, period == "post-war")))
head(textstat_keyness(dfmat3, target = "2017-Trump"), 10)

# using the likelihood ratio method
head(textstat_keyness(dfm_smooth(dfmat3), measure = "lr", target = "2017-Trump"), 10)