-
-
Notifications
You must be signed in to change notification settings - Fork 4.4k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Japanese model: add user_dict entries and small refactor #5573
Japanese model: add user_dict entries and small refactor #5573
Conversation
spacy/master
deleting: unidic_tags improve code readability around the token alignment procedure
Thanks for this contribution! In case there are any minor differences between these updates and the current version (which we've already used to train models), I would like to hold off on merging these updates for 2.3.0, but they can be included in the next releases (which would be 2.3.1 and 3.0.0). |
@adrianeboyd Fortunately, there is no difference in the Doc contents except the |
In addition, @polm -san is still improving the SudachiPy. |
Thanks for the ping. My work on Sudachi shouldn't change the output in any way so shouldn't require retraining the model. However, the ellipsis-related issue that came up while we were working on the Japanese model had a fix released around the same time. I think the fix won't change the actual tokens spaCy sees (since it discards empty tokens), but I'm not sure if tokenizations changes are possible. If there were changes, I would assume they would only be with rare/weird tokens. @sorami might know more about what the possible issues with the ellipsis change. What version of SudachiPy was used for training the model? |
The "ellipsis change" was included in SudachiPy v.0.4.5 (released Tuesday last week). There will be no change in the tokenization analysis result; Only the order of empty tokens will change. |
The models were trained with sudachipy v0.4.5 and this is the minimum required version listed in the This refactor is too large for me to feel comfortable including it in spacy v2.3.0. The models have already been trained and released and we're in the final stages of preparing the spacy library release, so we don't want to merge any more significant changes at this point. I would expect the first patch release to be relatively soon as bugs are fixed, so I don't think you'll have to wait too long for v2.3.1. |
@adrianeboyd Okay. I understand where are you coming from and agree to wait until release v2.3.1. Thanks for taking time for releasing the Japanese model! |
@adrianeboyd Thanks a lot! |
* Fix typos and auto-format [ci skip] * Add pkuseg warnings and auto-format [ci skip] * Update Binder URL [ci skip] * Update Binder version [ci skip] * Update alignment example for new gold.align * Update POS in tagging example * Fix numpy.zeros() dtype for Doc.from_array * Change example title to Dr. Change example title to Dr. so the current model does exclude the title in the initial example. * Fix spacy convert argument * Warning for sudachipy 0.4.5 (#5611) * Create myavrum.md (#5612) * Update lex_attrs.py (#5608) * Create mahnerak.md (#5615) * Some changes for Armenian (#5616) * Fixing numericals * We need a Armenian question sign to make the sentence a question * Add Nepali Language (#5622) * added support for nepali lang * added examples and test files * added spacy contributor agreement * Japanese model: add user_dict entries and small refactor (#5573) * user_dict fields: adding inflections, reading_forms, sub_tokens deleting: unidic_tags improve code readability around the token alignment procedure * add test cases, replace fugashi with sudachipy in conftest * move bunsetu.py to spaCy Universe as a pipeline component BunsetuRecognizer * tag is space -> both surface and tag are spaces * consider len(text)==0 * Add warnings example in v2.3 migration guide (#5627) * contribute (#5632) * Fix polarity of Token.is_oov and Lexeme.is_oov (#5634) Fix `Token.is_oov` and `Lexeme.is_oov` so they return `True` when the lexeme does **not** have a vector. * Extend what's new in v2.3 with vocab / is_oov (#5635) * Skip vocab in component config overrides (#5624) * Fix backslashes in warnings config diff (#5640) Fix backslashes in warnings config diff in v2.3 migration section. * Disregard special tag _SP in check for new tag map (#5641) * Skip special tag _SP in check for new tag map In `Tagger.begin_training()` check for new tags aside from `_SP` in the new tag map initialized from the provided gold tuples when determining whether to reinitialize the morphology with the new tag map. * Simplify _SP check Co-authored-by: Ines Montani <[email protected]> Co-authored-by: Marat M. Yavrumyan <[email protected]> Co-authored-by: Karen Hambardzumyan <[email protected]> Co-authored-by: Rameshh <[email protected]> Co-authored-by: Hiroshi Matsuda <[email protected]> Co-authored-by: Richard Liaw <[email protected]>
Add user_dict entries (reading_forms, inflections, sub_tokens) and improving the code readability around the token alignment procedure.
Description
py.test spacy/tests/lang/ja/test_tokenizer.py
and found no error.Types of change
Checklist