python -m utoken.utokenize [-h] [-i INPUT-FILENAME] [-o OUTPUT-FILENAME] [-a ANNOTATION-FILENAME] 
                           [--annotation_format ANNOTATION_FORMAT] [-p PROFILE-FILENAME] 
                           [--profile_scope PROFILE_SCOPE] [-d DATA_DIRECTORY] [--lc LANGUAGE-CODE] 
                           [-f] [-v] [-c] [--simple] [--version]

or simply

utokenize [-h] [-i INPUT-FILENAME] [-o OUTPUT-FILENAME] [-a ANNOTATION-FILENAME] 
          [--annotation_format ANNOTATION_FORMAT] [-p PROFILE-FILENAME] 
          [--profile_scope PROFILE_SCOPE] [-d DATA_DIRECTORY] [--lc LANGUAGE-CODE] 
          [-f] [-v] [-c] [--simple] [--version]  

optional arguments:
  -h, --help            show this help message and exit
  -i INPUT-FILENAME, --input INPUT-FILENAME
                        (default: STDIN)
  -o OUTPUT-FILENAME, --output OUTPUT-FILENAME
                        (default: STDOUT)
  -a ANNOTATION-FILENAME, --annotation_file ANNOTATION-FILENAME
                        (optional output)
  --annotation_format ANNOTATION_FORMAT
                        (default: 'json'; alternative: 'double-colon')
  -p PROFILE-FILENAME, --profile PROFILE-FILENAME
                        (optional output for performance analysis)
  --profile_scope PROFILE_SCOPE
                        (optional scope for performance analysis)
  -d DATA_DIRECTORY, --data_directory DATA_DIRECTORY
                        (default: standard data directory)
  --lc LANGUAGE-CODE    ISO 639-3, e.g. 'fas' for Persian
  -f, --first_token_is_line_id
                        First token is line ID (and will be exempt from any tokenization)
  -v, --verbose         write change log etc. to STDERR
  -c, --chart           build annotation chart, even without annotation output
  --simple              prevent MT-style output (e.g. @-@). Note: can degrade any detokenization
  --version             show program's version number and exit

Note: Please make sure that your $PYTHONPATH includes the directory in which this README file resides.

Example:

echo "Wait,dont tell me." | utokenize --lc eng

Output:

Wait , do n't tell me .

python -m utoken.detokenize [-h] [-i INPUT-FILENAME] [-o OUTPUT-FILENAME] [-d DATA_DIRECTORY] 
                            [--lc LANGUAGE-CODE] [-f] [-v] [--version]

or simply

detokenize [-h] [-i INPUT-FILENAME] [-o OUTPUT-FILENAME] [-d DATA_DIRECTORY] 
           [--lc LANGUAGE-CODE] [-f] [-v] [--version]

optional arguments:
  -h, --help            show this help message and exit
  -i INPUT-FILENAME, --input INPUT-FILENAME
                        (default: STDIN)
  -o OUTPUT-FILENAME, --output OUTPUT-FILENAME
                        (default: STDOUT)
  -d DATA_DIRECTORY, --data_directory DATA_DIRECTORY
                        (default: standard data directory)
  --lc LANGUAGE-CODE    ISO 639-3, e.g. 'fas' for Persian
  -f, --first_token_is_line_id
                        First token is line ID (and will be exempt from any tokenization)
  -v, --verbose         write change log etc. to STDERR
  --version             show program's version number and exit

Note: Please make sure that your $PYTHONPATH includes the directory in which this README file resides.

Example:

echo "Wait , do n't tell me ." | detokenize --lc eng

Output:

Wait, don't tell me.

from utoken import utokenize

tok = utokenize.Tokenizer(lang_code='eng')  # Initialize tokenizer, load resources
print(tok.utokenize_string("Dont worry!"))
print(tok.utokenize_string("Sold,for $9,999.99 on ebay.com."))

Output:

Do n't worry !
Sold , for $ 9,999.99 on ebay.com .

Note: Please make sure that your $PYTHONPATH includes the directory in which this README file resides.

from utoken import detokenize

detok = detokenize.Detokenizer(lang_code='eng')  # Initialize detokenizer, load resources
print(detok.detokenize_string("Do n't worry !"))
print(detok.detokenize_string("Sold , for $ 9,999.99 on ebay.com ."))

Output:

Don't worry!
Sold, for $9,999.99 on ebay.com.

Note: Please make sure that your $PYTHONPATH includes the directory in which this README file resides.

pip install utoken

or

git clone https://github.com/uhermjakob/utoken.git

What gets split

• Contractions: John's → John 's; we've → we 've; can't → can n't; won't → will n't
• Quantities into number and unit: 5,000km² → 5,000 km²
• Ordinal numbers into number and ordinal particle: 350th → 350 th
• Non-lexical hyphenated expressions: peace-loving → peace @-@ loving
• Name initials: J.S.Bach → J. S. Bach

What stays together

• XML tags: <a href="http://www.hollywoodbowl.com">
• URLs: https://www.youtube.com/watch?v=AaZ_RSt0KP8
• Email addresses: а.almukhanov@energo.gov.kz
• Filenames: Оперплан_каз2015.doc
• Numbers: -12,345,678.90   १,२३,४५,६७८.९०
• Abbreviations: Mr.   e.g.   w/o
• Lexicon entries with dashes: T-shirt   father-in-law   so-called   Port-au-Prince
• Lexicon entries with apostrophe: Xi’an   's-Gravenhage
• Hashtags, handles: #global_warming   #2   @GermanBeer
• Groups of related punctuation: ???
• Groups of emojis and other symbols: ⚽👍🎉
• Words with an internal zero width non-joiner: e.g. Farsi می‌خواهم

Mark-up of certain punctuation (e.g. @-@)

For many application such as machine translation, tokenization is important, but should be reversed when producing the final output. In some cases, this is relatively straight forward, so . and , typically attach to the word on the left and ( attaches to the word on the right. In other cases, it can generally be very hard to decide how to detokenize, so we add a special tag such as @ during tokenization in order to guide later dekonization. A @ on one or both sides of punctuation indicates that in the original text, the punctuation and neighboring word were together. To look at it in another way, the tokenizer basically upgrades the non-directional " to an open "@ or close @" delimiter.

Example: ("Hello,world!")   Tokenized: ( "@ Hello , world ! @" )   Detokenized: ("Hello, world!")

If later detokenization is not import and you want to suppress any markup with @, call utokenizer.py with the option --simple

Example: ("Hello,world!")   Tokenized (simple): ( " Hello , world ! " )   Detokenized: (" Hello, world! ")

Option --first_token_is_line_id

In some applications, the text to be tokenized is preceded by a sentence ID at the beginning of each line and tokenization should not be applied to those sentence IDs.
Option --first_token_is_line_id, or -f for short, suppresses tokenization of those sentence IDs.

• Example input: GEN:1:1 In the beginning, God created the heavens and the earth.
• utokenize.pl tokenization: GEN @:@ 1 @:@ 1 In the beginning , God created the heavens and the earth .
• utokenize.pl -f tokenization: GEN:1:1 In the beginning , God created the heavens and the earth .

Why is tokenization hard?

Tokenization is more then just splitting a sentence along spaces, as a lot of punctuation such as commas and periods are attached to adjacent words. But we can't just blindly split off commas and periods, as this would break numbers such as 12,345.60, abbreviations such as Mr. or URLs such as www.usc.edu.

• There are many special types of entities that need to be preserved in tokenization, e.g.
  • XML tags: <a href="http://www.hollywoodbowl.com">
  • URLs: https://www.youtube.com/watch?v=AaZ_RSt0KP8
  • Email addresses: а.almukhanov@energo.gov.kz
  • Filenames: Оперплан_каз2015.doc
  • Numbers: -12,345,678.90   १,२३,४५,६७८.९०
  • Hashtags, handles: #global_warming   #2   @GermanBeer
• Abbreviations can be hard to determine in many languages, as a period might indicate an abbreviation or the end of a sentence.
  • Abbreviations: Mr.   e.g.   w/o
• Apostrophes are normal letters in some languages, e.g. Somali su'aal (question). Apostrophes can appear in foreign names (e.g. Xi'an and 's-Gravenhage). In some languages, an apostrophe is used for contractions, such as John's and we'll in English. Additionally, an apostrophe can be used as a quote around a word or phrase such as 'Good job!'. All these cases have to be treated differently.
• Hyphens can join independent words such as in peace-loving (which should be split). But they also occur inside lexical phrases such as T-shirt that should not be split.
• Many applications need to map a tokenized sentence back to 'normal' untokenized text. To support such a detokenizer, the tokenizer's output must facilitate future detokenization. For example, by default, the tokenizer adds attachment tags such as '@' to punctuation to indicate to which side(s) they should attach after detokenization. For more on this topic, please see topic Mark-up of certain punctuation above.
• Other challenges: symbols, variation selectors, non-standard whitespaces, special characters such as zero width non-joiner.
• In general, it is hard to make a tokenizer work universally, for a wide range of languages, scripts and conventions.
• utoken uses a combination of general patterns and lists of specific tokens to solve many of the challenges above. (See more under topic Tokenization data files.)
• Example for a language-specific challenge: In Modern Hebrew, acronyms are marked by placing a gershayim between the last two characters, e.g. ארה״ב (USA). In practice, the gershayim is often replaced by the more readily available quotation mark ("). However, quotation marks are also used for quotations, e.g. ה"סתום" (the "valve"), so care has to be taken to do justice to both acronymns (preserve as a single token) and quotes (separate into multiple tokens).

Tokenization data files

utokenize includes a number of data files to supports its operation:

• tok-resource.txt includes language-independent tokenization resource entries, especially for punctuation, abbreviations (e.g. km²) and names (especially those with hyphens, spaces and other non-alpha characters)
• tok-resource-eng-global.txt contains tokenization resource entries for English that are also loaded for other languages. This is helpful as foreign texts often code-switch to English.
• tok-resource-eng.txt contains tokenization resource entries for English that are not shared, including those that would not work in other languages. For example, in English, dont in a non-standard version of don't and is tokenized into do n't, but in French, dont (of which) is a regular word that should be left alone.
• detok-resource.txt includes resources for detokenization. The file is also used by the tokenizer to mark up certain punctuation with attachment tags such as @-@.
• There are numerous other tok-resource-xxx.txt files for other languages, some larger than others. Some languages such as Farsi just don't use contractions and abbreviations with periods that much, so there are few entries. Others files might benefit from additional contributions.
• top-level-domain-codes.txt contains a list of suffixes such as .com, .org, .uk, .tv to support tokenization of URLs and email address.

Exmaples of resource entries:

::punct-split ! ::side end ::group True ::comment multiple !!! remain grouped as a single token
::contraction can't ::target can n't ::lcode eng
::repair wo n't ::target will n't ::lcode eng ::problem previous tokenizer
::abbrev No. ::exp number ::lcode eng ::sem-class corpus-component ::case-sensitive True ::right-context \s*\d
::lexical T-shirt ::lcode eng ::plural +s
::misspelling accomodate ::target accommodate ::lcode eng ::suffix-variations e/ed;es;ing;ion;ions

::markup-attach - ::group True ::comment hyphen-minus ::example the hyphen in _peace-loving_ will be marked up as ```@-@```
::auto-attach th ::side left ::left-context \d ::lcode eng ::example 20th

Speed

210,000 characters per second (real time) on a 39k sentence English AMR corpus on a 2021 MacBook Pro using a single CPU. Parallelization is trivial as sentences are tokenized independent of each other.

Testing

utoken has been tested on 71 corpora in 60 languages and 18 scripts (as of Oct. 18, 2021). Tests include

• Manual review of tokenization
• Comparison to other tokenizers: Sacremoses and ulf-tokenizer
• Tokenization analysis scripts:
  • wildebeest (text normalization and cleaning; analysis of types of characters used, encoding issues)
  • aux/tok-analysis.py (looks for a number of potential problems such as tokens with mixed letters/digits, mixed letters/punctuation, potential abbreviations separated from period)
• Comparisons to previous versions of all test corpora before release.

Related software

• Universal romanizer uroman, written by Ulf Hermjakob (same author)

Future work — Feedback and contributions welcome

Plans include

• Building resources, testing and fine-tuning of additional languages such as Hausa and Serbian.
• Adding new special entity types such as IPA pronunciations, geographic coordinates, complex IDs such as 403(k).
• Semi-supervised learning of lexical and abbreviation resources from large corpora.