* Tokenize by graphemes, not codepoints
* Add more documentation to TOKEN_RE
* Remove extra line break
* Update docstring - Brahmic scripts are no longer an exception
* approve using version 2017.07.28 of regex
This changes the version from 1.4.2 to 1.5. Things done in this update include:
* include Common Crawl; support 11 more languages
* new frequency-merging strategy
* New sources: Chinese from Wikipedia (mostly Trad.), Dutch big list
* Remove kinda bad sources, i.e. Greek Twitter (too often kaomoji are detected as Greek) and Ukrainian Common Crawl. This results in dropping Ukrainian as an available language, and causing Greek to not be a 'large' language after all.
* Add Korean tokenization, and include MeCab files in data
* Remove marks from more languages
* Deal with commas and cedillas in Turkish and Romanian
Former-commit-id: e6a8f028e3
* Remove marks from more languages
* Add Korean tokenization, and include MeCab files in data
* add a Hebrew tokenization test
* fix terminology in docstrings about abjad scripts
* combine Japanese and Korean tokenization into the same function
Former-commit-id: fec6eddcc3
Our regex already has a special case to leave Chinese and Japanese alone
when an appropriate tokenizer for the language isn't being used, as
Unicode's default segmentation would make every character into its own
token.
The same thing happens in Thai, and we don't even *have* an appropriate
tokenizer for Thai, so I've added a similar fallback.
Former-commit-id: 07f16e6f03