Граница слова

What is a "Word"?

English speakers intuitively know where words start and end: spaces and punctuation serve as dividers. But for a computer processing multilingual text, "word" is a surprisingly complex concept. Chinese, Japanese, and Thai use no spaces between words. German compounds like "Donaudampfschifffahrtsgesellschaft" are single orthographic words. English contractions like "don't" and "I've" may be one or two tokens depending on the application.

UAX #29 Word Boundary rules provide an algorithmic definition of word boundaries that works reasonably across scripts, suitable for applications that need to tokenize text, implement double-click selection, count words, or process natural language.

Word Boundary Rules Overview

The UAX #29 word boundary algorithm assigns each character to a word break property and applies a table of rules to determine if a boundary exists between adjacent characters. Key properties:

Notable rules: - Contractions: don't is ONE word token because apostrophe (MidLetter) between two Letter characters is not a boundary. - Numbers: 3.14 and 1,000 are single tokens because . and , between digits are MidNum characters. - Identifiers: my_variable is one token because _ is ExtendNumLet. - Email/URL: UAX #29 has special rules to keep [email protected] as tokens.

CJK and Scripts Without Spaces

For Chinese, Japanese, and Thai, UAX #29 uses a simplified approach: every character is its own "word" at the UAX #29 level. Real word segmentation for these scripts requires language-specific processing (statistical models, dictionaries):

# UAX #29 treats each CJK character as a separate word token
# For real Japanese segmentation, use MeCab or SudachiPy
# For real Chinese, use jieba or pkuseg
# For real Thai, use PyThaiNLP

import jieba  # pip install jieba
tokens = list(jieba.cut("我喜欢学习自然语言处理"))
print(tokens)  # ['我', '喜欢', '学习', '自然语言处理']

Python and ICU Word Segmentation

from icu import BreakIterator, Locale, RuleBasedBreakIterator

text = "Don't stop, it's 3.14 o'clock. [email protected]"
bi = BreakIterator.createWordInstance(Locale("en_US"))
bi.setText(text)

start = 0
for end in bi:
    token = text[start:end]
    rule_status = bi.getRuleStatus()
    # rule_status == 200-299: word (letter-based)
    # rule_status == 100-199: number
    # rule_status == 0: non-word (space/punctuation)
    if rule_status != 0:
        print(f"word: {repr(token)}")
    start = end

Quick Facts