`tokenization`

This library provides extensible tokenization and sentence splitting for atoms, character lists, and character code lists. A representation-aware facade delegates canonical character-code lists to a language provider. The included English provider uses portable deterministic rules for words, contractions, hyphenated words, abbreviations, numbers, email addresses, punctuation, and sentence boundaries. URL and email address recognition delegates to the url library for parsing and validation.

API documentation

Open the ../../apis/library_index.html#tokenization link in a web browser.

Loading

To load the library and the English provider:

| ?- logtalk_load(tokenization(loader)).

Testing

To test this library predicates:

| ?- logtalk_load(tokenization(tester)).

Usage

The tokenizer(Representation, Language) object accepts atom, chars, and codes representations. The language parameter must be an object implementing the tokenizer_language_protocol protocol.

| ?- tokenizer(atom, english_tokenizer)::tokenize(
         'Mr. Smith bought a new car for $12,500. Isn\'t that great?',
         Tokens
     ).
Tokens = ['Mr.', 'Smith', bought, a, new, car, for, '$', '12,500', '.', 'Isn\'t', that, great, '?']
yes

| ?- tokenizer(atom, english_tokenizer)::tokenize(
         'Hello, world!', Tokens, [keep_punctuation(false)]
     ).
Tokens = ['Hello', world]
yes

Sentence splitting is available independently or combined with tokenization:

| ?- tokenizer(atom, english_tokenizer)::split_sentences(
         'Dr. Brown arrived at 3.14 p.m. She was late. Really?', Sentences
     ).
Sentences = ['Dr. Brown arrived at 3.14 p.m.', 'She was late.', 'Really?']
yes

| ?- tokenizer(atom, english_tokenizer)::tokenize_sentences(
         'She was late. Really?', Sentences
     ).
Sentences = [['She', was, late, '.'], ['Really', '?']]
yes

Options

The tokenize/3, split_sentences/3, and tokenize_sentences/3 predicates accept these options:

keep_punctuation(Boolean) controls emission of standalone punctuation tokens. The default is true. Punctuation internal to abbreviations, numbers, URLs, email addresses, contractions, and hyphenated words is kept.
lowercase(Boolean) controls lowercasing of emitted tokens. The default is false.
normalize_quotes(Boolean) maps common Unicode single and double quotation marks to ASCII quotes before processing. The default is true.
normalize_dashes(Boolean) maps common Unicode dash characters to the ASCII hyphen before processing. The default is true.

Pipeline integration

Token lists can be passed directly to the other text and NLP libraries:

text_to_trigrams(Text, Trigrams) :-
    tokenizer(atom, english_tokenizer)::tokenize(
        Text, Tokens0, [keep_punctuation(false), lowercase(true)]
    ),
    stop_words(atom, stopwords_en)::exclude(Tokens0, Tokens1),
    lemmatizer(atom, english_lemmatizer)::lemmas(Tokens1, Tokens2),
    n_grams(atom)::trigrams(Tokens2, Trigrams).

Adding providers

Providers operate on character-code lists and receive validated, merged options. A provider can implement tokenizer_language_protocol directly:

:- object(example_tokenizer,
    implements(tokenizer_language_protocol)).

    tokenize_codes(Codes, [Codes], _).
    split_sentence_codes(Codes, [Codes], _).

:- end_object.

Rule-based language providers can instead import the tokenizer_rules category and define its protected abbreviation and internal punctuation hooks, as illustrated by the english_tokenizer object. The provider protocol does not otherwise constrain the implementation; finite-state, statistical, cached, dynamically loaded, and external tokenizers can use the same facade.

English coverage

The English provider includes common honorifics and ranks, month names, academic and professional credentials, organization and editorial forms, geographic initialisms, time abbreviations, and common Latin abbreviations. Titles and ranks are treated as non-terminal when followed by more text; other abbreviations use the following token and surrounding context to decide whether a period ends a sentence.

Limitations

The included rules are practical heuristics, not a complete linguistic model. They do not implement Unicode normalization or Unicode text segmentation, emoji grapheme clustering, social-media conventions, typed tokens, or multi-word expression recognition. Non-ASCII characters not recognized as supported punctuation are conservatively treated as word characters.

tokenization