* Detect more empty matches in tokenizer.explain()
* Include a few languages in explain non-slow tests
Mark a few languages in tokenizer.explain() tests as not slow so they're
run by default.
* Expose tokenizer rules as a property
Expose the tokenizer rules property in the same way as the other core
properties. (The cache resetting is overkill, but consistent with
`from_bytes` for now.)
Add tests and update Tokenizer API docs.
* Update Hungarian punctuation to remove empty string
Update Hungarian punctuation definitions so that `_units` does not match
an empty string.
* Use _load_special_tokenization consistently
Use `_load_special_tokenization()` and have it to handle `None` checks.
* Fix precedence of `token_match` vs. special cases
Remove `token_match` check from `_split_affixes()` so that special cases
have precedence over `token_match`. `token_match` is checked only before
infixes are split.
* Add `make_debug_doc()` to the Tokenizer
Add `make_debug_doc()` to the Tokenizer as a working implementation of
the pseudo-code in the docs.
Add a test (marked as slow) that checks that `nlp.tokenizer()` and
`nlp.tokenizer.make_debug_doc()` return the same non-whitespace tokens
for all languages that have `examples.sentences` that can be imported.
* Update tokenization usage docs
Update pseudo-code and algorithm description to correspond to
`nlp.tokenizer.make_debug_doc()` with example debugging usage.
Add more examples for customizing tokenizers while preserving the
existing defaults.
Minor edits / clarifications.
* Revert "Update Hungarian punctuation to remove empty string"
This reverts commit f0a577f7a5.
* Rework `make_debug_doc()` as `explain()`
Rework `make_debug_doc()` as `explain()`, which returns a list of
`(pattern_string, token_string)` tuples rather than a non-standard
`Doc`. Update docs and tests accordingly, leaving the visualization for
future work.
* Handle cases with bad tokenizer patterns
Detect when tokenizer patterns match empty prefixes and suffixes so that
`explain()` does not hang on bad patterns.
* Remove unused displacy image
* Add tokenizer.explain() to usage docs
* Rework Chinese language initialization
* Create a `ChineseTokenizer` class
* Modify jieba post-processing to handle whitespace correctly
* Modify non-jieba character tokenization to handle whitespace correctly
* Add a `create_tokenizer()` method to `ChineseDefaults`
* Load lexical attributes
* Update Chinese tag_map for UD v2
* Add very basic Chinese tests
* Test tokenization with and without jieba
* Test `like_num` attribute
* Fix try_jieba_import()
* Fix zh code formatting
The model registry refactor of the Tok2Vec function broke loading models
trained with the previous function, because the model tree was slightly
different. Specifically, the new function wrote:
concatenate(norm, prefix, suffix, shape)
To build the embedding layer. In the previous implementation, I had used
the operator overloading shortcut:
( norm | prefix | suffix | shape )
This actually gets mapped to a binary association, giving something
like:
concatenate(norm, concatenate(prefix, concatenate(suffix, shape)))
This is a different tree, so the layers iterate differently and we
loaded the weights wrongly.
* Xfail new tokenization test
* Put new alignment behind feature flag
* Move USE_ALIGN to top of the file [ci skip]
Co-authored-by: Ines Montani <ines@ines.io>
The `Matcher` in `merge_subtokens()` returns all possible subsequences
of `subtok`, so for sequences of two or more subtoks it's necessary to
filter the matches so that the retokenizer is only merging the longest
matches with no overlapping spans.
* Add arch for MishWindowEncoder
* Support mish in tok2vec and conv window >=2
* Pass new tok2vec settings from parser
* Syntax error
* Fix tok2vec setting
* Fix registration of MishWindowEncoder
* Fix receptive field setting
* Fix mish arch
* Pass more options from parser
* Support more tok2vec options in pretrain
* Require thinc 7.3
* Add docs [ci skip]
* Require thinc 7.3.0.dev0 to run CI
* Run black
* Fix typo
* Update Thinc version
Co-authored-by: Ines Montani <ines@ines.io>
