While developing v2.1, I ran a bunch of hyper-parameter search
experiments to find settings that performed well for spaCy's NER and
parser. I ended up changing the default Adam settings from beta1=0.9,
beta2=0.999, eps=1e-8 to beta1=0.8, beta2=0.8, eps=1e-5. This was giving
a small improvement in accuracy (like, 0.4%).
Months later, I run the models with Prodigy, which uses beam-search
decoding even when the model has been trained with a greedy objective.
The new models performed terribly...So, wtf? After a couple of days
debugging, I figured out that the new optimizer settings was causing the
model to converge to solutions where the top-scoring class often had
a score of like, -80. The variance on the weights had gone up
enormously. I guess I needed to update the L2 regularisation as well?
Anyway. Let's just revert the change --- if the optimizer is finding
such extreme solutions, that seems bad, and not nearly worth the small
improvement in accuracy.
Currently training a slate of models, to verify the accuracy change is minimal.
Once the training is complete, we can merge this.
<!--- Provide a general summary of your changes in the title. -->
## Description
<!--- Use this section to describe your changes. If your changes required
testing, include information about the testing environment and the tests you
ran. If your test fixes a bug reported in an issue, don't forget to include the
issue number. If your PR is still a work in progress, that's totally fine – just
include a note to let us know. -->
### Types of change
<!-- What type of change does your PR cover? Is it a bug fix, an enhancement
or new feature, or a change to the documentation? -->
## Checklist
<!--- Before you submit the PR, go over this checklist and make sure you can
tick off all the boxes. [] -> [x] -->
- [x] I have submitted the spaCy Contributor Agreement.
- [x] I ran the tests, and all new and existing tests passed.
- [x] My changes don't require a change to the documentation, or if they do, I've added all required information.
Add and document CLI options for batch size, max doc length, min doc length for `spacy pretrain`.
Also improve CLI output.
Closes#3216
## Checklist
<!--- Before you submit the PR, go over this checklist and make sure you can
tick off all the boxes. [] -> [x] -->
- [x] I have submitted the spaCy Contributor Agreement.
- [x] I ran the tests, and all new and existing tests passed.
- [x] My changes don't require a change to the documentation, or if they do, I've added all required information.
* merging conllu/conll and conllubio scripts
* tabs to spaces
* removing conllubio2json from converters/__init__.py
* Move not-really-CLI tests to misc
* Add converter test using no-ud data
* Fix test I broke
* removing include_biluo parameter
* fixing read_conllx
* remove include_biluo from convert.py
* label in span not writable anymore
* more explicit unit test and error message for readonly label
* bit more explanation (view)
* error msg tailored to specific case
* fix None case
Closes#2091.
## Description
With the new `vocab.writing_system` property introduced in #3390 (exposed via the language defaults), I was able to finally fix this (I think!). Based on the `Doc`, dispaCy now detects whether it's a RTL or LTR language and adjusts the visualization accordingly. Wherever possible, I've also added `direction` and `lang` attributes.
Entity visualization now looks like this:
<img width="318" alt="Screenshot 2019-03-11 at 16 06 51" src="https://user-images.githubusercontent.com/13643239/54136866-d97afd80-441c-11e9-8c27-3d46994cc833.png">
And dependencies like this (ignore the most likely incorrect tags and dependencies):
<img width="621" alt="Screenshot 2019-03-11 at 16 51 59" src="https://user-images.githubusercontent.com/13643239/54137771-8b66f980-441e-11e9-8460-0682b95eef2a.png">
### Types of change
enhancement, bug fix
## Checklist
<!--- Before you submit the PR, go over this checklist and make sure you can
tick off all the boxes. [] -> [x] -->
- [x] I have submitted the spaCy Contributor Agreement.
- [x] I ran the tests, and all new and existing tests passed.
- [x] My changes don't require a change to the documentation, or if they do, I've added all required information.
* Add xfail test for vocab.writing_system
* Add vocab.writing_system property
* Set Language.Defaults.writing_system
* Set default writing system
* Remove xfail on test_vocab_writing_system
Closes#2203. Closes#3268.
Lemmas set from outside the `Morphology` class were being overwritten. The result was especially confusing when deserialising, as it meant some lemmas could change when storing and retrieving a `Doc` object.
This PR applies two fixes:
1) When we go to set the lemma in the `Morphology` class, first check whether a lemma is already set. If so, don't overwrite.
2) When we load with `doc.from_array()`, take care to apply the `TAG` field first. This allows other fields to overwrite the `TAG` implied properties, if they're provided explicitly (e.g. the `LEMMA`).
## Checklist
<!--- Before you submit the PR, go over this checklist and make sure you can
tick off all the boxes. [] -> [x] -->
- [x] I have submitted the spaCy Contributor Agreement.
- [x] I ran the tests, and all new and existing tests passed.
- [x] My changes don't require a change to the documentation, or if they do, I've added all required information.
* Add component_cfg kwarg to begin_training
* Document component_cfg arg to begin_training
* Update docs and auto-format
* Support component_cfg across Language
* Format
* Update docs and docstrings [ci skip]
* Fix begin_training
* Make serialization methods consistent
exclude keyword argument instead of random named keyword arguments and deprecation handling
* Update docs and add section on serialization fields
* Use default return instead of else
* Add Doc.is_nered to indicate if entities have been set
* Add properties in Doc.to_json if they were set, not if they're available
This way, if a processed Doc exports "pos": None, it means that the tag was explicitly unset. If it exports "ents": [], it means that entity annotations are available but that this document doesn't contain any entities. Before, this would have been unclear and problematic for training.
<!--- Provide a general summary of your changes in the title. -->
## Description
* tidy up and adjust Cython code to code style
* improve docstrings and make calling `help()` nicer
* add URLs to new docs pages to docstrings wherever possible, mostly to user-facing objects
* fix various typos and inconsistencies in docs
### Types of change
enhancement, docs
## Checklist
<!--- Before you submit the PR, go over this checklist and make sure you can
tick off all the boxes. [] -> [x] -->
- [x] I have submitted the spaCy Contributor Agreement.
- [x] I ran the tests, and all new and existing tests passed.
- [x] My changes don't require a change to the documentation, or if they do, I've added all required information.
* Improve handling of missing NER tags
GoldParse can accept missing NER tags, if entities is provided
in BILUO format (rather than as spans). Missing tags can be provided
as None values.
Fix bug that occurred when first tag was a None value. Closes#2603.
* Document specification of missing NER tags.
* Classes for Ukrainian; small fix in Russian.
* Contributor agreement
* pymorphy2 initialization split for ru and uk (#3327)
* stop-words fixed
* Unit-tests updated
<!--- Provide a general summary of your changes in the title. -->
## Description
This PR adds the abilility to override custom extension attributes during merging. This will only work for attributes that are writable, i.e. attributes registered with a default value like `default=False` or attribute that have both a getter *and* a setter implemented.
```python
Token.set_extension('is_musician', default=False)
doc = nlp("I like David Bowie.")
with doc.retokenize() as retokenizer:
attrs = {"LEMMA": "David Bowie", "_": {"is_musician": True}}
retokenizer.merge(doc[2:4], attrs=attrs)
assert doc[2].text == "David Bowie"
assert doc[2].lemma_ == "David Bowie"
assert doc[2]._.is_musician
```
### Types of change
enhancement
## Checklist
<!--- Before you submit the PR, go over this checklist and make sure you can
tick off all the boxes. [] -> [x] -->
- [x] I have submitted the spaCy Contributor Agreement.
- [x] I ran the tests, and all new and existing tests passed.
- [x] My changes don't require a change to the documentation, or if they do, I've added all required information.
* Keep TextCategorizer default model same as v2.0
* Add option 'architecture' that allows "simple_cnn" to switch to
simpler model.
* Add option exclusive_classes, defaulting to False. If set to True,
the model treats classes as mutually exclusive, i.e. only one class can
be true per instance.
* splitting up latin unicode interval
* removing hyphen as infix for French
* adding failing test for issue 1235
* test for issue #3002 which now works
* partial fix for issue #2070
* keep the hyphen as infix for French (as it was)
* restore french expressions with hyphen as infix (as it was)
* added succeeding unit test for Issue #2656
* Fix issue #2822 with custom Italian exception
* Fix issue #2926 by allowing numbers right before infix /
* splitting up latin unicode interval
* removing hyphen as infix for French
* adding failing test for issue 1235
* test for issue #3002 which now works
* partial fix for issue #2070
* keep the hyphen as infix for French (as it was)
* restore french expressions with hyphen as infix (as it was)
* added succeeding unit test for Issue #2656
* Fix issue #2822 with custom Italian exception
* Fix issue #2926 by allowing numbers right before infix /
* remove duplicate
* remove xfail for Issue #2179 fixed by Matt
* adjust documentation and remove reference to regex lib
* Fix matching on extension attrs and predicates
* Fix detection of match_id when using extension attributes. The match
ID is stored as the last entry in the pattern. We were checking for this
with nr_attr == 0, which didn't account for extension attributes.
* Fix handling of predicates. The wrong count was being passed through,
so even patterns that didn't have a predicate were being checked.
* Fix regex pattern
* Fix matcher set value test
* Change retokenize.split() API for heads
* Pass lists as values for attrs in split
* Fix test_doc_split filename
* Add error for mismatched tokens after split
* Raise error if new tokens don't match text
* Fix doc test
* Fix error
* Move deps under attrs
* Fix split tests
* Fix retokenize.split
* Add base classes for more languages
* Add test for language class initialization
Make sure language can be initialize – otherwise, it's difficult to catch serious errors in the test suite, because languages are lazy-loaded
* Add split one token into several (resolves#2838)
* Improve error message for token splitting
* Make retokenizer.split() tests use a Token object
Change retokenizer.split() to use a Token object, instead of an index.
* Pass Token into retokenize.split()
Tweak retokenize.split() API so that we pass the `Token` object, not the index.
* Fix token.idx in retokenize.split()
* Test that token.idx is correct after split
* Fix token.idx for split tokens
* Fix retokenize.split()
* Fix retokenize.split
* Fix retokenize.split() test
Otherwise, the true error that happens within a Language subclass is swallowed, because if it's imported lazily like that, it'll always be an ImportError
* Add custom MatchPatternError
* Improve validators and add validation option to Matcher
* Adjust formatting
* Never validate in Matcher within PhraseMatcher
If we do decide to make validate default to True, the PhraseMatcher's Matcher shouldn't ever validate. Here, we create the patterns automatically anyways (and it's currently unclear whether the validation has performance impacts at a very large scale).
In most cases, the PhraseMatcher will match on the verbatim token text or as of v2.1, sometimes the lowercase text. This means that we only need a tokenized Doc, without any other attributes.
If phrase patterns are created by processing large terminology lists with the full `nlp` object, this easily can make things a lot slower, because all components will be applied, even if we don't actually need the attributes they set (like part-of-speech tags, dependency labels).
The warning message also includes a suggestion to use nlp.make_doc or nlp.tokenizer.pipe for even faster processing. For now, the validation has to be enabled explicitly by setting validate=True.
* Improved stop words list
* Removed some wrong stop words form list
* Improved stop words list
* Removed some wrong stop words form list
* Improved Polish Tokenizer (#38)
* Add tests for polish tokenizer
* Add polish tokenizer exceptions
* Don't split any words containing hyphens
* Fix test case with wrong model answer
* Remove commented out line of code until better solution is found
* Add source srx' license
* Rename exception_list.py to match spaCy conventionality
* Add a brief explanation of where the exception list comes from
* Add newline after reach exception
* Rename COPYING.txt to LICENSE
* Delete old files
* Add header to the license
* Agreements signed
* Stanisław Giziński agreement
* Krzysztof Kowalczyk - signed agreement
* Mateusz Olko agreement
* Add DoomCoder's contributor agreement
* Improve like number checking in polish lang
* like num tests added
* all from SI system added
* Final licence and removed splitting exceptions
* Added polish stop words to LEX_ATTRA
* Add encoding info to pl tokenizer exceptions
## Description
1. Added the same infix rule as in French (`d'une`, `j'ai`) for Italian (`c'è`, `l'ha`), bringing F-score on `it_isdt-ud-train.txt` from 96% to 99%. Added unit test to check this behaviour.
2. Added specific Urdu punctuation character as suffix, improving F-score on `ur_udtb-ud-train.txt` from 94% to 100%. Added unit test to check this behaviour.
### Types of change
Enhancement of Italian & Urdu tokenization
## Checklist
- [x] I have submitted the spaCy Contributor Agreement.
- [x] I ran the tests, and all new and existing tests passed.
- [x] My changes don't require a change to the documentation, or if they do, I've added all required information.
