Multilingual Stopwords

Silvie Cinkova

Maciej Eder

Functions and arguments

The tidystopwords package gives you potential stopwords in more than 100 languages. Its main function is generate_stoplist. Its argument language accepts atomic strings and character vectors of language names or language abbreviations corresponding to those listed by the helping function list_supported_languages.

The list_supported_languages function comes with three numbered output options.

Definition of stopwords

The list_supported_languages output is based on multilingual_stoplist - a data frame that was automatically extracted from the Universal Dependencies treebanks (henceforth UD). Universal Dependencies is a framework for cross-linguistically consistent grammatical annotation. The tidystopwords package uses their lemmatization, universal parts of speech, and universal features to derive an inventory of stop classes:

In terms of the Universal Dependencies, the stop classes are defined as follows:

Vocabulary coverage

Each version of this package uses the latest UD release available to generate the multilingual_stoplist data frame. Therefore multilingual_stoplist can differ from version to version. Typically, a new UD release brings bigger annotated corpora and emerging corpora of new languages.

All stopword lists in tidystopwords have been generated automatically from the data available at the moment. Hence their quality depends on the size of the underlying corpora as well as the morphological richness of the given language.

To allow the user to assess the reliability of the stopword list for the given language, the multilingual_stoplist contains relevant frequency information for each word form in three columns: n_formlemma, n_uposlemma, and
n_stopclasses.

The n_formlemma column gives the absolute frequency of the given word form with the given lemma. The n_uposformlemma column gives the absolute frequency of the given word form with the given lemma and upos.

The n_stopclasses column says in how many stop classes the given word form with the given lemma occurs. For instance that occurs as determiner_quantifier (that pie tastes good), pronominal (don’t mention that), and conjunction_subordinator (say that you will do it).

Noise control

Even high-quality reference corpora such as the UD treebanks contain tagging errors and typos. A two step frequency filter minimizes the noise: 1) a word form must occur more than three times with a given lemma; 2) if a word form with a given lemma (rendered by n_formlemma) occurs in
several different upos combinations (n_uposlemma), only combinations that represent more than 20% of n_formlemma remain listed.