tokenizer 2.0.3

A tokenizer for Icelandic text

Overview

Tokenization is a necessary first step in many natural language processing tasks, such as word counting, parsing, spell checking, corpus generation, and statistical analysis of text.

Tokenizer is a compact pure-Python (2 and 3) executable program and module for tokenizing Icelandic text. It converts input text to streams of tokens, where each token is a separate word, punctuation sign, number/amount, date, e-mail, URL/URI, etc. It also segments the token stream into sentences, considering corner cases such as abbreviations and dates in the middle of sentences.

The package contains a dictionary of common Icelandic abbreviations, in the file src/tokenizer/Abbrev.conf.

Tokenizer is an independent spinoff from the Greynir project (GitHub repository here), by the same authors. The Greynir natural language parser for Icelandic uses Tokenizer on its input.

Note that Tokenizer is licensed under the MIT license while Greynir is licensed under GPLv3.

Deep vs. shallow tokenization

Tokenizer can do both deep and shallow tokenization.

Shallow tokenization simply returns each sentence as a string (or as a line of text in an output file), where the individual tokens are separated by spaces.

Deep tokenization returns token objects that have been annotated with the token type and further information extracted from the token, for example a (year, month, day) tuple in the case of date tokens.

In shallow tokenization, tokens are in most cases kept intact, although consecutive white space is always coalesced. The input strings "800 MW", "21. janúar" and "800 7000" thus become two tokens each, output with a single space between them.

In deep tokenization, the same strings are represented by single token objects, of type TOK.MEASUREMENT, TOK.DATEREL and TOK.TELNO, respectively. The text associated with a single token object may contain one or more spaces, although consecutive space is always coalesced.

By default, the command line tool performs shallow tokenization. If you want deep tokenization with the command line tool, use the --json or --csv switches.

From Python code, call split_into_sentences() for shallow tokenization, or tokenize() for deep tokenization. These functions are documented with examples below.

Installation

To install:

$ pip install tokenizer

Command line tool

After installation, the tokenizer can be invoked directly from the command line:

$ tokenize input.txt output.txt

Input and output files are encoded in UTF-8. If the files are not given explicitly, stdin and stdout are used for input and output, respectively.

Empty lines in the input are treated as sentence boundaries.

By default, the output consists of one sentence per line, where each line ends with a single newline character (ASCII LF, chr(10), "\n"). Within each line, tokens are separated by spaces.

The following (mutually exclusive) options can be specified on the command line:

--csv	Deep tokenization. Output token objects in CSV format, one per line. Sentences are separated by lines containing 0,"",""
--json	Deep tokenization. Output token objects in JSON format, one per line.
--normalize	Normalize punctuation, causing e.g. quotes to be output in Icelandic form and hyphens to be regularized. This option is only applicable to shallow tokenization.

Type tokenize -h or tokenize --help to get a short help message.

Example

$ echo "3.janúar sl. keypti   ég 64kWst rafbíl. Hann kostaði € 30.000." | tokenize
3. janúar sl. keypti ég 64kWst rafbíl .
Hann kostaði €30.000 .

$ echo "3.janúar sl. keypti   ég 64kWst rafbíl. Hann kostaði € 30.000." | tokenize --csv
19,"3. janúar","0|1|3"
6,"sl.","síðastliðinn"
6,"keypti",""
6,"ég",""
22,"64kWst","J|230400000.0"
6,"rafbíl",""
1,".","."
0,"",""
6,"Hann",""
6,"kostaði",""
13,"€30.000","30000|EUR"
1,".","."
0,"",""

$ echo "3.janúar sl. keypti   ég 64kWst rafbíl. Hann kostaði € 30.000." | tokenize --json
{"k":"BEGIN SENT"}
{"k":"DATEREL","t":"3. janúar","v":[0,1,3]}
{"k":"WORD","t":"sl.","v":["síðastliðinn"]}
{"k":"WORD","t":"keypti"}
{"k":"WORD","t":"ég"}
{"k":"MEASUREMENT","t":"64kWst","v":["J",230400000.0]}
{"k":"WORD","t":"rafbíl"}
{"k":"PUNCTUATION","t":".","v":"."}
{"k":"END SENT"}
{"k":"BEGIN SENT"}
{"k":"WORD","t":"Hann"}
{"k":"WORD","t":"kostaði"}
{"k":"AMOUNT","t":"€30.000","v":[30000,"EUR"]}
{"k":"PUNCTUATION","t":".","v":"."}
{"k":"END SENT"}

Python module

Shallow tokenization example

An example of shallow tokenization from Python code goes something like this:

from __future__ import print_function
# The following import is optional but convenient under Python 2.7
from __future__ import unicode_literals

from tokenizer import split_into_sentences

# A string to be tokenized, containing two sentences
s = "3.janúar sl. keypti   ég 64kWst rafbíl. Hann kostaði € 30.000."

# Obtain a generator of sentence strings
g = split_into_sentences(s)

# Loop through the sentences
for sentence in g:

    # Obtain the individual token strings
    tokens = sentence.split()

    # Print the tokens, comma-separated
    print(", ".join(tokens))

The program outputs:

3., janúar, sl., keypti, ég, 64kWst, rafbíl, .
Hann, kostaði, €30.000, .

Deep tokenization example

To do deep tokenization from within Python code:

# The following import is optional but convenient under Python 2.7
from __future__ import unicode_literals
from tokenizer import tokenize, TOK

text = ("Málinu var vísað til stjórnskipunar- og eftirlitsnefndar "
    "skv. 3. gr. XVII. kafla laga nr. 10/2007 þann 3. janúar 2010.")

for token in tokenize(text):

    print("{0}: '{1}' {2}".format(
        TOK.descr[token.kind],
        token.txt or "-",
        token.val or ""))

Output:

BEGIN SENT: '-' (0, None)
WORD: 'Málinu'
WORD: 'var'
WORD: 'vísað'
WORD: 'til'
WORD: 'stjórnskipunar- og eftirlitsnefndar'
WORD: 'skv.' [('samkvæmt', 0, 'fs', 'skst', 'skv.', '-')]
ORDINAL: '3.' 3
WORD: 'gr.' [('grein', 0, 'kvk', 'skst', 'gr.', '-')]
ORDINAL: 'XVII.' 17
WORD: 'kafla'
WORD: 'laga'
WORD: 'nr.' [('númer', 0, 'hk', 'skst', 'nr.', '-')]
NUMBER: '10' (10, None, None)
PUNCTUATION: '/' (4, '/')
YEAR: '2007' 2007
WORD: 'þann'
DATEABS: '3. janúar 2010' (2010, 1, 3)
PUNCTUATION: '.' (3, '.')
END SENT: '-'

Note the following:

• Sentences are delimited by TOK.S_BEGIN and TOK.S_END tokens.

• Composite words, such as stjórnskipunar- og eftirlitsnefndar, are coalesced into one token.

• Well-known abbreviations are recognized and their full expansion is available in the token.val field.

• Ordinal numbers (3., XVII.) are recognized and their value (3, 17) is available in the token.val field.

• Dates, years and times, both absolute and relative, are recognized and the respective year, month, day, hour, minute and second values are included as a tuple in token.val.

• Numbers, both integer and real, are recognized and their value is available in the token.val field.

• Further details of how Tokenizer processes text can be inferred from the test module in the project’s GitHub repository.

The tokenize() function

To deep-tokenize a text string, call tokenizer.tokenize(text, **options). The text parameter can be a string, or an iterable that yields strings (such as a text file object).

The function returns a Python generator of token objects. Each token object is a simple namedtuple with three fields: (kind, txt, val) (further documented below).

The tokenizer.tokenize() function is typically called in a for loop:

import tokenizer
for token in tokenizer.tokenize(mystring):
    kind, txt, val = token
    if kind == tokenizer.TOK.WORD:
        # Do something with word tokens
        pass
    else:
        # Do something else
        pass

Alternatively, create a token list from the returned generator:

token_list = list(tokenizer.tokenize(mystring))

In Python 2.7, you can pass either unicode strings or str byte strings to tokenizer.tokenize(). In the latter case, the byte string is assumed to be encoded in UTF-8.

The split_into_sentences() function

To shallow-tokenize a text string, call tokenizer.split_into_sentences(text_or_gen, **options). The text_or_gen parameter can be a string, or an iterable that yields strings (such as a text file object).

This function returns a Python generator of strings, yielding a string for each sentence in the input. Within a sentence, the tokens are separated by spaces.

You can pass the option normalize=True to the function if you want the normalized form of punctuation tokens. Normalization outputs Icelandic single and double quotes („these“) instead of English-style ones (“these”), converts three-dot ellipsis … to single character ellipsis …, and casts en-dashes – and em-dashes — to regular hyphens.

The tokenizer.split_into_sentences() function is typically called in a for loop:

import tokenizer
with open("example.txt", "r", encoding="utf-8") as f:
    # You can pass a file object directly to split_into_sentences()
    for sentence in tokenizer.split_into_sentences(f):
        # sentence is a string of space-separated tokens
        tokens = sentence.split()
        # Now, tokens is a list of strings, one for each token
        for t in tokens:
            # Do something with the token t
            pass

The correct_spaces() function

The tokenizer.correct_spaces(text) function returns a string after splitting it up and re-joining it with correct whitespace around punctuation tokens. Example:

>>> import tokenizer
>>> tokenizer.correct_spaces(
... "Frétt \n  dagsins:Jón\t ,Friðgeir og Páll ! 100  /  2  =   50"
... )
'Frétt dagsins: Jón, Friðgeir og Páll! 100/2 = 50'

The detokenize() function

The tokenizer.detokenize(tokens, normalize=False) function takes an iterable of token objects and returns a corresponding, correctly spaced text string, composed from the tokens’ text. If the normalize parameter is set to True, the function uses the normalized form of any punctuation tokens, such as proper Icelandic single and double quotes instead of English-type quotes. Example:

>>> import tokenizer
>>> toklist = list(tokenizer.tokenize("Hann sagði: „Þú ert ágæt!“."))
>>> tokenizer.detokenize(toklist, normalize=True)
'Hann sagði: „Þú ert ágæt!“.'

The normalized_text() function

The tokenizer.normalized_text(token) function returns the normalized text for a token. This means that the original token text is returned except for certain punctuation tokens, where a normalized form is returned instead. Specifically, English-type quotes are converted to Icelandic ones, and en- and em-dashes are converted to regular hyphens.

The text_from_tokens() function

The tokenizer.text_from_tokens(tokens) function returns a concatenation of the text contents of the given token list, with spaces between tokens. Example:

>>> import tokenizer
>>> toklist = list(tokenizer.tokenize("Hann sagði: \"Þú ert ágæt!\"."))
>>> tokenizer.text_from_tokens(toklist)
'Hann sagði : " Þú ert ágæt ! " .'

The normalized_text_from_tokens() function

The tokenizer.normalized_text_from_tokens(tokens) function returns a concatenation of the normalized text contents of the given token list, with spaces between tokens. Example (note the double quotes):

>>> import tokenizer
>>> toklist = list(tokenizer.tokenize("Hann sagði: \"Þú ert ágæt!\"."))
>>> tokenizer.normalized_text_from_tokens(toklist)
'Hann sagði : „ Þú ert ágæt ! “ .'

Tokenization options

You can optionally pass one or more of the following options as keyword parameters to the tokenize() and split_into_sentences() functions:

• convert_numbers=[bool]

  Setting this option to True causes the tokenizer to convert numbers and amounts with English-style decimal points (.) and thousands separators (,) to Icelandic format, where the decimal separator is a comma (,) and the thousands separator is a period (.). $1,234.56 is thus converted to a token whose text is $1.234,56.

  The default value for the convert_numbers option is False.

  Note that in versions of Tokenizer prior to 1.4, convert_numbers was True.

• handle_kludgy_ordinals=[value]

  This options controls the way Tokenizer handles ‘kludgy’ ordinals, such as 1sti, 4ðu, or 2ja. By default, such ordinals are returned unmodified (‘passed through’) as word tokens (TOK.WORD). However, this can be modified as follows:

  • tokenizer.KLUDGY_ORDINALS_MODIFY: Kludgy ordinals are corrected to become ‘proper’ word tokens, i.e. 1sti becomes fyrsti and 2ja becomes tveggja.

  • tokenizer.KLUDGY_ORDINALS_TRANSLATE: Kludgy ordinals that represent proper ordinal numbers are translated to ordinal tokens (TOK.ORDINAL), with their original text and their ordinal value. 1sti thus becomes a TOK.ORDINAL token with a value of 1, and 3ja becomes a TOK.ORDINAL with a value of 3.

  • tokenizer.KLUDGY_ORDINALS_PASS_THROUGH is the default value of the option. It causes kludgy ordinals to be returned unmodified as word tokens.

  Note that versions of Tokenizer prior to 1.4 behaved as if handle_kludgy_ordinals were set to tokenizer.KLUDGY_ORDINALS_TRANSLATE.

The token object

Each token is represented by a namedtuple with three fields: (kind, txt, val).

The kind field

The kind field contains one of the following integer constants, defined within the TOK class:

Constant	Value	Explanation	Examples
PUNCTUATION	1	Punctuation	. ! ; % &
TIME	2	Time (h, m, s)	11:35:40 kl. 7:05 klukkan 23:35
DATE *	3	Date (y, m, d)	[Unused, see DATEABS and DATEREL]
YEAR	4	Year	árið 874 e.Kr. 1965 44 f.Kr.
NUMBER	5	Number	100 1.965 1.965,34 1,965.34 2⅞
WORD	6	Word	kattaeftirlit hunda- og kattaeftirlit
TELNO	7	Telephone number	5254764 699-4244 410 4000
PERCENT	8	Percentage	78%
URL	9	URL	https://greynir.is http://tiny.cc/28695y
ORDINAL	10	Ordinal number	30. XVIII.
TIMESTAMP *	11	Timestamp	[Unused, see TIMESTAMPABS and TIMESTAMPREL]
CURRENCY *	12	Currency name	[Unused]
AMOUNT	13	Amount	€2.345,67 750 þús.kr. 2,7 mrð. USD kr. 9.900 EUR 200
PERSON *	14	Person name	[Unused]
EMAIL	15	E-mail	[e-mail]
ENTITY *	16	Named entity	[Unused]
UNKNOWN	17	Unknown token
DATEABS	18	Absolute date	30. desember 1965 30/12/1965 1965-12-30 1965/12/30
DATEREL	19	Relative date	15. mars 15/3 15.3. mars 1911
TIMESTAMPABS	20	Absolute timestamp	30. desember 1965 11:34 1965-12-30 kl. 13:00
TIMESTAMPREL	21	Relative timestamp	30. desember kl. 13:00
MEASUREMENT	22	Value with a measurement unit	690 MW 1.010 hPa 220 m² 80° C
NUMWLETTER	23	Number followed by a single letter	14a 7B
DOMAIN	24	Domain name	greynir.is Reddit.com www.wikipedia.org
HASHTAG	25	Hashtag	#MeToo #12stig
MOLECULE	26	Molecular formula	H2SO4 CO2
SSN	27	Social security number (kennitala)	591213-1480
USERNAME	28	Twitter user handle	@username_123
SERIALNUMBER	29	Serial number	394-5388 12-345-6789
S_BEGIN	11001	Start of sentence
S_END	11002	End of sentence

(*) The token types marked with an asterisk are reserved for the Reynir package and not currently returned by the tokenizer.

To obtain a descriptive text for a token kind, use TOK.descr[token.kind] (see example above).

The txt field

The txt field contains the original source text for the token, with the following exceptions:

• All contiguous whitespace (spaces, tabs, newlines) is coalesced into single spaces (" ") within the txt field. A date token that is parsed from a source text of "29. \n janúar" thus has a txt of "29. janúar".

• Tokenizer automatically merges Unicode COMBINING ACUTE ACCENT (code point 769) and COMBINING DIAERESIS (code point 776) with vowels to form single code points for the Icelandic letters á, é, í, ó, ú, ý and ö, in both lower and upper case.

• If the appropriate options are specified (see above), it converts kludgy ordinals (3ja) to proper ones (þriðja), and English-style thousand and decimal separators to Icelandic ones (10,345.67 becomes 10.345,67).

In the case of abbreviations that end a sentence, the final period "." is a separate token, and it is consequently omitted from the abbreviation token’s txt field. A sentence ending in o.s.frv. will thus end with two tokens, the next-to-last one being the tuple (TOK.WORD, "o.s.frv", "og svo framvegis") - note the omitted period in the txt field - and the last one being (TOK.PUNCTUATION, ".", (3, ".")) (this tuple form is further explained below).

The val field

The val field contains auxiliary information, corresponding to the token kind, as follows:

• For TOK.PUNCTUATION, the val field contains a tuple with two items: (whitespace, normalform). The first item (token.val[0]) specifies the whitespace normally found around the symbol in question, as an integer:

  TP_LEFT = 1   # Whitespace to the left
  TP_CENTER = 2 # Whitespace to the left and right
  TP_RIGHT = 3  # Whitespace to the right
  TP_NONE = 4   # No whitespace

  The second item (token.val[1]) contains a normalized representation of the punctuation. For instance, various forms of single and double quotes are represented as Icelandic ones (i.e. „these“ or ‚these‘) in normalized form, and ellipsis (”…”) are represented as the single character “…”.

• For TOK.TIME, the val field contains an (hour, minute, second) tuple.

• For TOK.DATEABS, the val field contains a (year, month, day) tuple (all 1-based).

• For TOK.DATEREL, the val field contains a (year, month, day) tuple (all 1-based), except that a least one of the tuple fields is missing and set to 0. Example: 3. júní becomes TOK.DATEREL with the fields (0, 6, 3) as the year is missing.

• For TOK.YEAR, the val field contains the year as an integer. A negative number indicates that the year is BCE (fyrir Krist), specified with the suffix f.Kr. (e.g. árið 33 f.Kr.).

• For TOK.NUMBER, the val field contains a tuple (number, None, None). (The two empty fields are included for compatibility with Greynir.)

• For TOK.WORD, the val field contains the full expansion of an abbreviation, as a list containing a single tuple, or None if the word is not abbreviated.

• For TOK.PERCENT, the val field contains a tuple of (percentage, None, None).

• For TOK.ORDINAL, the val field contains the ordinal value as an integer. The original ordinal may be a decimal number or a Roman numeral.

• For TOK.TIMESTAMP, the val field contains a (year, month, day, hour, minute, second) tuple.

• For TOK.AMOUNT, the val field contains an (amount, currency, None, None) tuple. The amount is a float, and the currency is an ISO currency code, e.g. USD for dollars ($ sign), EUR for euros (€ sign) or ISK for Icelandic króna (kr. abbreviation). (The two empty fields are included for compatibility with Greynir.)

• For TOK.MEASUREMENT, the val field contains a (unit, value) tuple, where unit is a base SI unit (such as g, m, m², s, W, Hz, K for temperature in Kelvin).

• For TOK.TELNO, the val field contains a tuple: (number, cc) where the first item is the phone number in a normalized NNN-NNNN format, i.e. always including a hyphen, and the second item is the country code, eventually prefixed by +. The country code defaults to 354 (Iceland).

Abbreviations

Abbreviations recognized by Tokenizer are defined in the Abbrev.conf file, found in the src/tokenizer/ directory. This is a text file with abbreviations, their definitions and explanatory comments.

When an abbreviation is encountered, it is recognized as a word token (i.e. having its kind field equal to TOK.WORD). Its expansion(s) are included in the token’s val field as a list containing tuples of the format (ordmynd, utg, ordfl, fl, stofn, beyging). An example is o.s.frv., which results in a val field equal to [('og svo framvegis', 0, 'ao', 'frasi', 'o.s.frv.', '-')].

The tuple format is designed to be compatible with the Database of Modern Icelandic Inflection (DMII), Beygingarlýsing íslensks nútímamáls.

Development installation

To install Tokenizer in development mode, where you can easily modify the source files (assuming you have git available):

$ git clone https://github.com/mideind/Tokenizer
$ cd Tokenizer
$ # [ Activate your virtualenv here, if you have one ]
$ pip install -e .

Test suite

Tokenizer comes with a large test suite. The file test/test_tokenizer.py contains built-in tests that run under pytest.

To run the built-in tests, install pytest, cd to your Tokenizer subdirectory (and optionally activate your virtualenv), then run:

$ python -m pytest

The file test/toktest_large.txt contains a test set of 13,075 lines. The lines test sentence detection, token detection and token classification. For analysis, test/toktest_large_gold_perfect.txt contains the expected output of a perfect shallow tokenization, and test/toktest_large_gold_acceptable.txt contains the current output of the shallow tokenization.

The file test/Overview.txt (only in Icelandic) contains a description of the test set, including line numbers for each part in both test/toktest_large.txt and test/toktest_large_gold_acceptable.txt, and a tag describing what is being tested in each part.

It also contains a description of a perfect shallow tokenization for each part, acceptable tokenization and the current behaviour. As such, the description is an analysis of which edge cases the tokenizer can handle and which it can not.

To test the tokenizer on the large test set the following needs to be typed in the command line:

$ tokenize test/toktest_large.txt test/toktest_large_out.txt

To compare it to the acceptable behaviour:

$ diff test/toktest_large_out.txt test/toktest_large_gold_acceptable.txt > diff.txt

The file test/toktest_normal.txt contains a running text from recent news articles, containing no edge cases. The gold standard for that file can be found in the file test/toktest_normal_gold_expected.txt.

Changelog

• Version 2.0.3: Fixed bug in detokenize() where abbreviations, domains and e-mails containing periods were wrongly split

• Version 2.0.2: Spelled-out day ordinals are no longer included as a part of TOK.DATEREL tokens. Thus, þriðji júní is now a TOK.WORD followed by a TOK.DATEREL. 3. júní continues to be parsed as a single TOK.DATEREL.

• Version 2.0.1: Order of abbreviation meanings within the token.val field made deterministic; fixed bug in measurement unit handling

• Version 2.0.0: Added command line tool; added split_into_sentences() and detokenize() functions; removed convert_telno option; splitting of coalesced tokens made more robust; added TOK.SSN, TOK.MOLECULE, TOK.USERNAME and TOK.SERIALNUMBER token kinds; abbreviations can now have multiple meanings

• Version 1.4.0: Added the **options parameter to the tokenize() function, giving control over the handling of numbers, telephone numbers, and ‘kludgy’ ordinals

• Version 1.3.0: Added TOK.DOMAIN and TOK.HASHTAG token types; improved handling of capitalized month name Ágúst, which is now recognized when following an ordinal number; improved recognition of telephone numbers; added abbreviations

• Version 1.2.3: Added abbreviations; updated GitHub URLs

• Version 1.2.2: Added support for composites with more than two parts, i.e. „dómsmála-, ferðamála-, iðnaðar- og nýsköpunarráðherra“; added support for ± sign; added several abbreviations

• Version 1.2.1: Fixed bug where the name Ágúst was recognized as a month name; Unicode nonbreaking and invisible space characters are now removed before tokenization

• Version 1.2.0: Added support for Unicode fraction characters; enhanced handing of degrees (°, °C, °F); fixed bug in cubic meter measurement unit; more abbreviations

• Version 1.1.2: Fixed bug in liter (l and ltr) measurement units

• Version 1.1.1: Added mark_paragraphs() function

• Version 1.1.0: All abbreviations in Abbrev.conf are now returned with their meaning in a tuple in token.val; handling of ‘mbl.is’ fixed

• Version 1.0.9: Added abbreviation ‘MAST’; harmonized copyright headers

• Version 1.0.8: Bug fixes in DATEREL, MEASUREMENT and NUMWLETTER token handling; added ‘kWst’ and ‘MWst’ measurement units; blackened

• Version 1.0.7: Added TOK.NUMWLETTER token type

• Version 1.0.6: Automatic merging of Unicode COMBINING ACUTE ACCENT and COMBINING DIAERESIS code points with vowels

• Version 1.0.5: Date/time and amount tokens coalesced to a further extent

• Version 1.0.4: Added TOK.DATEABS, TOK.TIMESTAMPABS, TOK.MEASUREMENT