Merge pull request #55 from LuminosoInsight/version2

Version 2, with standalone text pre-processing

CHANGELOG.md

@@ -1,3 +1,66 @@
+## Version 2.0 (2018-03-14)
+
+The big change in this version is that text preprocessing, tokenization, and
+postprocessing to look up words in a list are separate steps.
+
+If all you need is preprocessing to make text more consistent, use
+`wordfreq.preprocess.preprocess_text(text, lang)`. If you need preprocessing
+and tokenization, use `wordfreq.tokenize(text, lang)` as before. If you need
+all three steps, use the new function `wordfreq.lossy_tokenize(text, lang)`.
+
+As a breaking change, this means that the `tokenize` function no longer has
+the `combine_numbers` option, because that's a postprocessing step. For
+the same behavior, use `lossy_tokenize`, which always combines numbers.
+
+Similarly, `tokenize` will no longer replace Chinese characters with their
+Simplified Chinese version, while `lossy_tokenize` will.
+
+Other changes:
+
+- There's a new default wordlist for each language, called "best". This
+  chooses the "large" wordlist for that language, or if that list doesn't
+  exist, it falls back on "small".
+
+- The wordlist formerly named "combined" (this name made sense long ago)
+  is now named "small". "combined" remains as a deprecated alias.
+
+- The "twitter" wordlist has been removed. If you need to compare word
+  frequencies from individual sources, you can work with the separate files in
+  [exquisite-corpus][].
+
+- Tokenizing Chinese will preserve the original characters, no matter whether
+  they are Simplified or Traditional, instead of replacing them all with
+  Simplified characters.
+
+- Different languages require different processing steps, and the decisions
+  about what these steps are now appear in the `wordfreq.language_info` module,
+  replacing a bunch of scattered and inconsistent `if` statements.
+
+- Tokenizing CJK languages while preserving punctuation now has a less confusing
+  implementation.
+
+- The preprocessing step can transliterate Azerbaijani, although we don't yet
+  have wordlists in this language. This is similar to how the tokenizer
+  supports many more languages than the ones with wordlists, making future
+  wordlists possible.
+
+- Speaking of that, the tokenizer will log a warning (once) if you ask to tokenize
+  text written in a script we can't tokenize (such as Thai).
+
+- New source data from [exquisite-corpus][] includes OPUS OpenSubtitles 2018.
+
+Nitty gritty dependency changes:
+
+- Updated the regex dependency to 2018.02.21. (We would love suggestions on
+  how to coexist with other libraries that use other versions of `regex`,
+  without a `>=` requirement that could introduce unexpected data-altering
+  changes.)
+
+- We now depend on `msgpack`, the new name for `msgpack-python`.
+
+[exquisite-corpus]: https://github.com/LuminosoInsight/exquisite-corpus
+
+
 ## Version 1.7.0 (2017-08-25)
 
 - Tokenization will always keep Unicode graphemes together, including

README.md

@@ -7,7 +7,7 @@ Author: Robyn Speer
 ## Installation
 
 wordfreq requires Python 3 and depends on a few other Python modules
-(msgpack, langcodes, and ftfy). You can install it and its dependencies
+(msgpack, langcodes, and regex). You can install it and its dependencies
 in the usual way, either by getting it from pip:
 
     pip3 install wordfreq
@@ -23,20 +23,21 @@ steps that are necessary to get Chinese, Japanese, and Korean word frequencies.
 ## Usage
 
 wordfreq provides access to estimates of the frequency with which a word is
-used, in 27 languages (see *Supported languages* below).
+used, in 35 languages (see *Supported languages* below).
 
-It provides three kinds of pre-built wordlists:
+It provides both 'small' and 'large' wordlists:
 
-- `'combined'` lists, containing words that appear at least once per
-  million words, averaged across all data sources.
-- `'twitter'` lists, containing words that appear at least once per
-  million words on Twitter alone.
-- `'large'` lists, containing words that appear at least once per 100
-  million words, averaged across all data sources.
+- The 'small' lists take up very little memory and cover words that appear at
+  least once per million words.
+- The 'large' lists cover words that appear at least once per 100 million
+  words.
 
-The most straightforward function is:
+The default list is 'best', which uses 'large' if it's available for the
+language, and 'small' otherwise.
 
-    word_frequency(word, lang, wordlist='combined', minimum=0.0)
+The most straightforward function for looking up frequencies is:
+
+    word_frequency(word, lang, wordlist='best', minimum=0.0)
 
 This function looks up a word's frequency in the given language, returning its
 frequency as a decimal between 0 and 1. In these examples, we'll multiply the
@@ -47,10 +48,10 @@ frequencies by a million (1e6) to get more readable numbers:
     11.748975549395302
 
     >>> word_frequency('café', 'en') * 1e6
-    3.981071705534969
+    3.890451449942805
 
     >>> word_frequency('cafe', 'fr') * 1e6
-    1.4125375446227555
+    1.4454397707459279
 
     >>> word_frequency('café', 'fr') * 1e6
     53.70317963702532
@@ -65,25 +66,25 @@ example, and a word with Zipf value 3 appears once per million words.
 
 Reasonable Zipf values are between 0 and 8, but because of the cutoffs
 described above, the minimum Zipf value appearing in these lists is 1.0 for the
-'large' wordlists and 3.0 for all others. We use 0 as the default Zipf value
+'large' wordlists and 3.0 for 'small'. We use 0 as the default Zipf value
 for words that do not appear in the given wordlist, although it should mean
 one occurrence per billion words.
 
     >>> from wordfreq import zipf_frequency
     >>> zipf_frequency('the', 'en')
-    7.75
+    7.77
 
     >>> zipf_frequency('word', 'en')
     5.32
 
     >>> zipf_frequency('frequency', 'en')
-    4.36
+    4.38
 
     >>> zipf_frequency('zipf', 'en')
-    0.0
+    1.32
 
-    >>> zipf_frequency('zipf', 'en', wordlist='large')
-    1.28
+    >>> zipf_frequency('zipf', 'en', wordlist='small')
+    0.0
 
 
 The parameters to `word_frequency` and `zipf_frequency` are:
@@ -95,7 +96,7 @@ The parameters to `word_frequency` and `zipf_frequency` are:
 - `lang`: the BCP 47 or ISO 639 code of the language to use, such as 'en'.
 
 - `wordlist`: which set of word frequencies to use. Current options are
-  'combined', 'twitter', and 'large'.
+  'small', 'large', and 'best'.
 
 - `minimum`: If the word is not in the list or has a frequency lower than
   `minimum`, return `minimum` instead. You may want to set this to the minimum
@@ -108,7 +109,7 @@ Other functions:
 way that the words in wordfreq's data were counted in the first place. See
 *Tokenization*.
 
-`top_n_list(lang, n, wordlist='combined')` returns the most common *n* words in
+`top_n_list(lang, n, wordlist='best')` returns the most common *n* words in
 the list, in descending frequency order.
 
     >>> from wordfreq import top_n_list
@@ -118,18 +119,18 @@ the list, in descending frequency order.
     >>> top_n_list('es', 10)
     ['de', 'la', 'que', 'el', 'en', 'y', 'a', 'los', 'no', 'se']
 
-`iter_wordlist(lang, wordlist='combined')` iterates through all the words in a
+`iter_wordlist(lang, wordlist='best')` iterates through all the words in a
 wordlist, in descending frequency order.
 
-`get_frequency_dict(lang, wordlist='combined')` returns all the frequencies in
+`get_frequency_dict(lang, wordlist='best')` returns all the frequencies in
 a wordlist as a dictionary, for cases where you'll want to look up a lot of
 words and don't need the wrapper that `word_frequency` provides.
 
-`supported_languages(wordlist='combined')` returns a dictionary whose keys are
+`supported_languages(wordlist='best')` returns a dictionary whose keys are
 language codes, and whose values are the data file that will be loaded to
 provide the requested wordlist in each language.
 
-`random_words(lang='en', wordlist='combined', nwords=5, bits_per_word=12)`
+`random_words(lang='en', wordlist='best', nwords=5, bits_per_word=12)`
 returns a selection of random words, separated by spaces. `bits_per_word=n`
 will select each random word from 2^n words.
 
@@ -256,7 +257,7 @@ into multiple tokens:
     >>> zipf_frequency('New York', 'en')
     5.35
     >>> zipf_frequency('北京地铁', 'zh')  # "Beijing Subway"
-    3.55
+    3.54
 
 The word frequencies are combined with the half-harmonic-mean function in order
 to provide an estimate of what their combined frequency would be. In Chinese,

setup.cfg

@@ -0,0 +1,5 @@
+[nosetests]
+verbosity=2
+with-doctest=1
+with-coverage=0
+cover-package=wordfreq

setup.py

@@ -28,7 +28,7 @@ README_contents = open(os.path.join(current_dir, 'README.md'),
                        encoding='utf-8').read()
 doclines = README_contents.split("\n")
 dependencies = [
-    'ftfy >= 5', 'msgpack', 'langcodes >= 1.4', 'regex == 2017.07.28'
+    'msgpack', 'langcodes >= 1.4.1', 'regex == 2018.02.21'
 ]
 if sys.version_info < (3, 4):
     dependencies.append('pathlib')
@@ -36,7 +36,7 @@ if sys.version_info < (3, 4):
 
 setup(
     name="wordfreq",
-    version='1.7.0',
+    version='2.0',
     maintainer='Luminoso Technologies, Inc.',
     maintainer_email='info@luminoso.com',
     url='http://github.com/LuminosoInsight/wordfreq/',

tests/test.py

@@ -1,9 +1,9 @@
 from wordfreq import (
     word_frequency, available_languages, cB_to_freq,
-    top_n_list, random_words, random_ascii_words, tokenize
+    top_n_list, random_words, random_ascii_words, tokenize, lossy_tokenize
 )
 from nose.tools import (
-    eq_, assert_almost_equal, assert_greater, raises
+    eq_, assert_almost_equal, assert_greater, raises, assert_not_equal
 )
 
 
@@ -15,35 +15,29 @@ def test_freq_examples():
     assert_greater(word_frequency('de', 'es'),
                    word_frequency('the', 'es'))
 
-
-# To test the reasonableness of the Twitter list, we want to look up a
-# common word representing laughter in each language. The default for
-# languages not listed here is 'haha'.
-LAUGHTER_WORDS = {
-    'en': 'lol',
-    'hi': 'lol',
-    'cs': 'lol',
-    'ru': 'лол',
-    'zh': '笑',
-    'ja': '笑',
-    'ar': 'ﻪﻬﻬﻬﻫ',
-    'fa': 'خخخخ',
-    'ca': 'jaja',
-    'es': 'jaja',
-    'fr': 'ptdr',
-    'pt': 'kkkk',
-    'he': 'חחח',
-    'bg': 'ахаха',
-    'uk': 'хаха',
-    'bn': 'হা হা',
-    'mk': 'хаха'
-}
+    # We get word frequencies from the 'large' list when available
+    assert_greater(word_frequency('infrequency', 'en'), 0.)
 
 
 def test_languages():
-    # Make sure the number of available languages doesn't decrease
+    # Make sure we get all the languages when looking for the default
+    # 'best' wordlist
     avail = available_languages()
-    assert_greater(len(avail), 26)
+    assert_greater(len(avail), 32)
+
+    # 'small' covers the same languages, but with some different lists
+    avail_small = available_languages('small')
+    eq_(len(avail_small), len(avail))
+    assert_not_equal(avail_small, avail)
+
+    # 'combined' is the same as 'small'
+    avail_old_name = available_languages('combined')
+    eq_(avail_old_name, avail_small)
+
+    # 'large' covers fewer languages
+    avail_large = available_languages('large')
+    assert_greater(len(avail_large), 12)
+    assert_greater(len(avail), len(avail_large))
 
     # Look up the digit '2' in the main word list for each language
     for lang in avail:
@@ -55,17 +49,6 @@ def test_languages():
         assert_greater(word_frequency('2', new_lang_code), 0, new_lang_code)
 
 
-def test_twitter():
-    avail = available_languages('twitter')
-    assert_greater(len(avail), 15)
-
-    for lang in avail:
-        assert_greater(word_frequency('rt', lang, 'twitter'),
-                       word_frequency('rt', lang, 'combined'))
-        text = LAUGHTER_WORDS.get(lang, 'haha')
-        assert_greater(word_frequency(text, lang, wordlist='twitter'), 0, (text, lang))
-
-
 def test_minimums():
     eq_(word_frequency('esquivalience', 'en'), 0)
     eq_(word_frequency('esquivalience', 'en', minimum=1e-6), 1e-6)
@@ -164,13 +147,13 @@ def test_casefolding():
 def test_number_smashing():
     eq_(tokenize('"715 - CRΣΣKS" by Bon Iver', 'en'),
         ['715', 'crσσks', 'by', 'bon', 'iver'])
-    eq_(tokenize('"715 - CRΣΣKS" by Bon Iver', 'en', combine_numbers=True),
+    eq_(lossy_tokenize('"715 - CRΣΣKS" by Bon Iver', 'en'),
         ['000', 'crσσks', 'by', 'bon', 'iver'])
-    eq_(tokenize('"715 - CRΣΣKS" by Bon Iver', 'en', combine_numbers=True, include_punctuation=True),
+    eq_(lossy_tokenize('"715 - CRΣΣKS" by Bon Iver', 'en', include_punctuation=True),
         ['"', '000', '-', 'crσσks', '"', 'by', 'bon', 'iver'])
-    eq_(tokenize('1', 'en', combine_numbers=True), ['1'])
-    eq_(tokenize('3.14', 'en', combine_numbers=True), ['0.00'])
-    eq_(tokenize('24601', 'en', combine_numbers=True), ['00000'])
+    eq_(lossy_tokenize('1', 'en'), ['1'])
+    eq_(lossy_tokenize('3.14', 'en'), ['0.00'])
+    eq_(lossy_tokenize('24601', 'en'), ['00000'])
     eq_(word_frequency('24601', 'en'), word_frequency('90210', 'en'))
 
 
@@ -231,6 +214,7 @@ def test_ideographic_fallback():
         ['ひらがな', 'カタカナ', 'romaji']
     )
 
+
 def test_other_languages():
     # Test that we leave Thai letters stuck together. If we had better Thai support,
     # we would actually split this into a three-word phrase.

tests/test_chinese.py

@@ -55,10 +55,19 @@ def test_tokens():
         ]
     )
 
-    # You match the same tokens if you look it up in Traditional Chinese.
-    eq_(tokenize(fact_simplified, 'zh'), tokenize(fact_traditional, 'zh'))
+    # Check that Traditional Chinese works at all
     assert_greater(word_frequency(fact_traditional, 'zh'), 0)
 
+    # You get the same token lengths if you look it up in Traditional Chinese,
+    # but the words are different
+    simp_tokens = tokenize(fact_simplified, 'zh', include_punctuation=True)
+    trad_tokens = tokenize(fact_traditional, 'zh', include_punctuation=True)
+    eq_(''.join(simp_tokens), fact_simplified)
+    eq_(''.join(trad_tokens), fact_traditional)
+    simp_lengths = [len(token) for token in simp_tokens]
+    trad_lengths = [len(token) for token in trad_tokens]
+    eq_(simp_lengths, trad_lengths)
+
 
 def test_combination():
     xiexie_freq = word_frequency('谢谢', 'zh')   # "Thanks"
@@ -83,5 +92,3 @@ def test_alternate_codes():
     # Separate codes for Mandarin and Cantonese
     eq_(tokenize('谢谢谢谢', 'cmn'), tokens)
     eq_(tokenize('谢谢谢谢', 'yue'), tokens)
-
-

tests/test_transliteration.py

@@ -1,5 +1,6 @@
 from nose.tools import eq_
 from wordfreq import tokenize
+from wordfreq.preprocess import preprocess_text
 
 
 def test_transliteration():
@@ -10,6 +11,21 @@ def test_transliteration():
     eq_(tokenize("Pa, ima tu mnogo stvari koje ne shvataš.", 'sr'),
         ['pa', 'ima', 'tu', 'mnogo', 'stvari', 'koje', 'ne', 'shvataš'])
 
+    # I don't have examples of complete sentences in Azerbaijani that are
+    # naturally in Cyrillic, because it turns out everyone writes Azerbaijani
+    # in Latin letters on the Internet, _except_ sometimes for Wiktionary.
+    # So here are some individual words.
+
+    # 'library' in Azerbaijani Cyrillic
+    eq_(preprocess_text('китабхана', 'az'), 'kitabxana')
+    eq_(preprocess_text('КИТАБХАНА', 'az'), 'kitabxana')
+    eq_(preprocess_text('KİTABXANA', 'az'), 'kitabxana')
+
+    # 'scream' in Azerbaijani Cyrillic
+    eq_(preprocess_text('бағырты', 'az'), 'bağırtı')
+    eq_(preprocess_text('БАҒЫРТЫ', 'az'), 'bağırtı')
+    eq_(preprocess_text('BAĞIRTI', 'az'), 'bağırtı')
+
 
 def test_actually_russian():
     # This looks mostly like Serbian, but was probably actually Russian.

wordfreq/__init__.py

@@ -1,4 +1,3 @@
-from wordfreq.tokens import tokenize, simple_tokenize
 from pkg_resources import resource_filename
 from functools import lru_cache
 import langcodes
@@ -10,18 +9,15 @@ import random
 import logging
 import math
 
+from .tokens import tokenize, simple_tokenize, lossy_tokenize
+from .language_info import get_language_info
+
 logger = logging.getLogger(__name__)
 
 
 CACHE_SIZE = 100000
 DATA_PATH = pathlib.Path(resource_filename('wordfreq', 'data'))
 
-# Chinese and Japanese are written without spaces. In Chinese, in particular,
-# we have to infer word boundaries from the frequencies of the words they
-# would create. When this happens, we should adjust the resulting frequency
-# to avoid creating a bias toward improbable word combinations.
-INFERRED_SPACE_LANGUAGES = {'zh'}
-
 # We'll divide the frequency by 10 for each token boundary that was inferred.
 # (We determined the factor of 10 empirically by looking at words in the
 # Chinese wordlist that weren't common enough to be identified by the
@@ -30,8 +26,9 @@ INFERRED_SPACE_LANGUAGES = {'zh'}
 # frequency.)
 INFERRED_SPACE_FACTOR = 10.0
 
-# simple_tokenize is imported so that other things can import it from here.
-# Suppress the pyflakes warning.
+# tokenize and simple_tokenize are imported so that other things can import
+# them from here. Suppress the pyflakes warning.
+tokenize = tokenize
 simple_tokenize = simple_tokenize
 
 
@@ -87,11 +84,21 @@ def read_cBpack(filename):
     return data[1:]
 
 
-def available_languages(wordlist='combined'):
+def available_languages(wordlist='best'):
     """
-    List the languages (as language-code strings) that the wordlist of a given
-    name is available in.
+    Given a wordlist name, return a dictionary of language codes to filenames,
+    representing all the languages in which that wordlist is available.
     """
+    if wordlist == 'best':
+        available = available_languages('small')
+        available.update(available_languages('large'))
+        return available
+    elif wordlist == 'combined':
+        logger.warning(
+            "The 'combined' wordlists have been renamed to 'small'."
+        )
+        wordlist = 'small'
+
     available = {}
     for path in DATA_PATH.glob('*.msgpack.gz'):
         if not path.name.startswith('_'):
@@ -103,7 +110,7 @@ def available_languages(wordlist='combined'):
 
 
 @lru_cache(maxsize=None)
-def get_frequency_list(lang, wordlist='combined', match_cutoff=30):
+def get_frequency_list(lang, wordlist='best', match_cutoff=30):
     """
     Read the raw data from a wordlist file, returning it as a list of
     lists. (See `read_cBpack` for what this represents.)
@@ -117,7 +124,8 @@ def get_frequency_list(lang, wordlist='combined', match_cutoff=30):
     best, score = langcodes.best_match(lang, list(available),
                                        min_score=match_cutoff)
     if score == 0:
-        raise LookupError("No wordlist available for language %r" % lang)
+        raise LookupError("No wordlist %r available for language %r"
+                          % (wordlist, lang))
 
     if best != lang:
         logger.warning(
@@ -184,7 +192,7 @@ def freq_to_zipf(freq):
 
 
 @lru_cache(maxsize=None)
-def get_frequency_dict(lang, wordlist='combined', match_cutoff=30):
+def get_frequency_dict(lang, wordlist='best', match_cutoff=30):
     """
     Get a word frequency list as a dictionary, mapping tokens to
     frequencies as floating-point probabilities.
@@ -198,7 +206,7 @@ def get_frequency_dict(lang, wordlist='combined', match_cutoff=30):
     return freqs
 
 
-def iter_wordlist(lang, wordlist='combined'):
+def iter_wordlist(lang, wordlist='best'):
     """
     Yield the words in a wordlist in approximate descending order of
     frequency.
@@ -215,8 +223,9 @@ def iter_wordlist(lang, wordlist='combined'):
 # it takes to look up frequencies from scratch, so something faster is needed.
 _wf_cache = {}
 
+
 def _word_frequency(word, lang, wordlist, minimum):
-    tokens = tokenize(word, lang, combine_numbers=True)
+    tokens = lossy_tokenize(word, lang)
     if not tokens:
         return minimum
 
@@ -234,39 +243,31 @@ def _word_frequency(word, lang, wordlist, minimum):
 
     freq = 1.0 / one_over_result
 
-    if lang in INFERRED_SPACE_LANGUAGES:
+    if get_language_info(lang)['tokenizer'] == 'jieba':
+        # If we used the Jieba tokenizer, we could tokenize anything to match
+        # our wordlist, even nonsense. To counteract this, we multiply by a
+        # probability for each word break that was inferred.
         freq /= INFERRED_SPACE_FACTOR ** (len(tokens) - 1)
 
     return max(freq, minimum)
 
 
-def word_frequency(word, lang, wordlist='combined', minimum=0.):
+def word_frequency(word, lang, wordlist='best', minimum=0.):
     """
     Get the frequency of `word` in the language with code `lang`, from the
-    specified `wordlist`. The default wordlist is 'combined', built from
-    whichever of these five sources have sufficient data for the language:
+    specified `wordlist`.
 
-      - Full text of Wikipedia
-      - A sample of 72 million tweets collected from Twitter in 2014,
-        divided roughly into languages using automatic language detection
-      - Frequencies extracted from OpenSubtitles
-      - The Leeds Internet Corpus
-      - Google Books Syntactic Ngrams 2013
+    These wordlists can be specified:
 
-    Another available wordlist is 'twitter', which uses only the data from
-    Twitter.
-
-    Words that we believe occur at least once per million tokens, based on
-    the average of these lists, will appear in the word frequency list.
+    - 'large': a wordlist built from at least 5 sources, containing word
+      frequencies of 10^-8 and higher
+    - 'small': a wordlist built from at least 3 sources, containing word
+      frquencies of 10^-6 and higher
+    - 'best': uses 'large' if available, and 'small' otherwise
 
     The value returned will always be at least as large as `minimum`.
-
-    If a word decomposes into multiple tokens, we'll return a smoothed estimate
-    of the word frequency that is no greater than the frequency of any of its
-    individual tokens.
-
-    It should be noted that the current tokenizer does not support
-    multi-word Chinese phrases.
+    You could set this value to 10^-8, for example, to return 10^-8 for
+    unknown words in the 'large' list instead of 0, avoiding a discontinuity.
     """
     args = (word, lang, wordlist, minimum)
     try:
@@ -278,7 +279,7 @@ def word_frequency(word, lang, wordlist='combined', minimum=0.):
         return _wf_cache[args]
 
 
-def zipf_frequency(word, lang, wordlist='combined', minimum=0.):
+def zipf_frequency(word, lang, wordlist='best', minimum=0.):
     """
     Get the frequency of `word`, in the language with code `lang`, on the Zipf
     scale.
@@ -306,7 +307,7 @@ def zipf_frequency(word, lang, wordlist='combined', minimum=0.):
 
 
 @lru_cache(maxsize=100)
-def top_n_list(lang, n, wordlist='combined', ascii_only=False):
+def top_n_list(lang, n, wordlist='best', ascii_only=False):
     """
     Return a frequency list of length `n` in descending order of frequency.
     This list contains words from `wordlist`, of the given language.
@@ -321,7 +322,7 @@ def top_n_list(lang, n, wordlist='combined', ascii_only=False):
     return results
 
 
-def random_words(lang='en', wordlist='combined', nwords=5, bits_per_word=12,
+def random_words(lang='en', wordlist='best', nwords=5, bits_per_word=12,
                  ascii_only=False):
     """
     Returns a string of random, space separated words.
@@ -346,7 +347,7 @@ def random_words(lang='en', wordlist='combined', nwords=5, bits_per_word=12,
     return ' '.join([random.choice(choices) for i in range(nwords)])
 
 
-def random_ascii_words(lang='en', wordlist='combined', nwords=5,
+def random_ascii_words(lang='en', wordlist='best', nwords=5,
                        bits_per_word=12):
     """
     Returns a string of random, space separated, ASCII words.

wordfreq/chinese.py

@@ -49,4 +49,11 @@ def jieba_tokenize(text, external_wordlist=False):
     else:
         if jieba_tokenizer is None:
             jieba_tokenizer = jieba.Tokenizer(dictionary=DICT_FILENAME)
-        return jieba_tokenizer.lcut(simplify_chinese(text), HMM=False)
+
+        # Tokenize the Simplified Chinese version of the text, but return
+        # those spans from the original text, even if it's in Traditional
+        # Chinese
+        tokens = []
+        for _token, start, end in jieba_tokenizer.tokenize(simplify_chinese(text), HMM=False):
+            tokens.append(text[start:end])
+        return tokens