I wrote a small script to read the UD English training data and check
that our tag map and morph rules were resulting in the best POS map.
This hadn't been done for some time, and there have been various changes
to the UD schema since it has been done. After these changes we should
see much better agreement between our POS assignments and the UD POS
tags.
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.
* replace unicode categories with raw list of code points
* simplifying ranges
* fixing variable length quotes
* removing redundant regular expression
* small cleanup of regexp notations
* quotes and alpha as ranges instead of alterations
* removed most regexp dependencies and features
* exponential backtracking - unit tests
* rewrote expression with pathological backtracking
* disabling double hyphen tests for now
* test additional variants of repeating punctuation
* remove regex and redundant backslashes from load_reddit script
* small typo fixes
* disable double punctuation test for russian
* clean up old comments
* format block code
* final cleanup
* naming consistency
* french strings as unicode for python 2 support
* french regular expression case insensitive
* modifying FR lookup to remove ambiguity and adding lookup vocab to FR files
* modifying FR lookup to remove ambiguity and adding lookup vocab to FR files
* updating the contributor agreement for amperinet
Resolves#3208.
Prevent interactions with other libraries (pandas) that also access `get_ipython().config` and its parameters. See #3208 for details. I don't fully understand why this happens, but in spaCy, we can at least make sure we avoid calling into this method.
<!--- 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.
* running UD eval
* printing timing of tokenizer: tokens per second
* timing of default English model
* structured output and parameterization to compare different runs
* additional flag to allow evaluation without parsing info
* printing verbose log of errors for manual inspection
* printing over- and undersegmented cases (and combo's)
* add under and oversegmented numbers to Score and structured output
* print high-freq over/under segmented words and word shapes
* printing examples as part of the structured output
* print the results to file
* batch run of different models and treebanks per language
* cleaning up code
* commandline script to process all languages in spaCy & UD
* heuristic to remove blinded corpora and option to run one single best per language
* pathlib instead of os for file paths
* Update matcher engine for regex and extensions
Add support for matching over arbitrary Python predicate functions, and
arbitrary Python attribute getters. This will allow matching over regex
patterns, and allow supporting extension attributes.
The results of the Python predicate functions are cached, so that we don't
call the same predicate function twice for the same token. The extension
attributes are fetched into an array for each token in the doc. This
should minimise the performance impact of the new features.
We still need to wire up these features to the patterns, and test it
all.
* Work on wiring up extra attributes in matcher
* Work on tests for extra matcher attrs
* Add support for extension attrs to matcher
* Test extension attribute matching
* Work on implementing predicate-based match patterns
* Get predicates working for set membership
* Add test for set membership
* Make extensions+predicates work
* Test matcher extensions
* Cache predicate results better in Matcher
* Remove print statement in matcher test
* Use srsly to get key for predicates
* Added the same punctuation rules as danish language.
* Added abbreviations and also the possibility to have capitalized abbreviations on some. Added a few specific cases too
* Added test for long texts in swedish
* Added morph rules, infixes and suffixes to __init__.py for swedish
* Added some tests for prefixes, infixes and suffixes
* Added tests for lemma
* Renamed files to follow convention
* [sv] Removed ambigious abbreviations
* Added more tests for tokenizer exceptions
* Added test for problem with punctuation in issue #2578
* Contributor agreement
* Removed faulty lemmatization of 'jag' ('I') as it was lemmatized to 'jaga' ('hunt')
Tamil language support to spaCy
Description
Hereby, creating new PR to add support for Tamil language in spaCy
added stop words, examples and numerical attributes
<--Working on other language data-->
Types of change
Enhancement
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.
* adding adverbs and irregular cases for empty words
* adding adverbs and irregular cases for empty words
* adding adverbs and irregular cases for empty words
* updating contributor agreement for amperinet
* modifying French lookup that contained wrong lemmas
* correcting wrong line breaks on hyphen
* adding contributor agreement for amperinet@
* correcting a typo
This PR adds a test for an untested case of `Span.get_lca_matrix`, and fixes a bug for that scenario, which I introduced in [this PR](https://github.com/explosion/spaCy/pull/3089) (sorry!).
## Description
The previous implementation of get_lca_matrix was failing for the case `doc[j:k].get_lca_matrix()` where `j > 0`. A test has been added for this case and the bug has been fixed.
### Types of change
Bug fix
## 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.
If doc.from_array() was called with say, only entity information, this
would cause doc.is_tagged to be set to False, even if tags were set.
This caused tags to be dropped from serialisation. The same was true for
doc.is_parsed.
Closes#3012.
Initially span.as_doc() was designed to return a view of the span's contents, as a Doc object. This was a nice idea, but it fails due to the token.idx property, which refers to the character offset within the string. In a span, the idx of the first token might not be 0. Because this data is different, we can't have a view --- it'll be inconsistent.
This patch changes span.as_doc() to instead return a copy. The docs are updated accordingly. Closes#1537
* Update test for span.as_doc()
* Make span.as_doc() return a copy. Closes#1537
* Document change to Span.as_doc()
The doc.retokenize() context manager wasn't resizing doc.tensor, leading to a mismatch between the number of tokens in the doc and the number of rows in the tensor. We fix this by deleting rows from the tensor. Merged spans are represented by the vector of their last token.
* Add test for resizing doc.tensor when merging
* Add test for resizing doc.tensor when merging. Closes#1963
* Update get_lca_matrix test for develop
* Fix retokenize if tensor unset
<!--- Provide a general summary of your changes in the title. -->
## Description
See #3079. Here I'm merging into `develop` instead of `master`.
### 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? -->
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.
* Test on #2396: bug in Doc.get_lca_matrix()
* reimplementation of Doc.get_lca_matrix(), (closes#2396)
* reimplement Span.get_lca_matrix(), and call it from Doc.get_lca_matrix()
* tests Span.get_lca_matrix() as well as Doc.get_lca_matrix()
* implement _get_lca_matrix as a helper function in doc.pyx; call it from Doc.get_lca_matrix and Span.get_lca_matrix
* use memory view instead of np.ndarray in _get_lca_matrix (faster)
* fix bug when calling Span.get_lca_matrix; return lca matrix as np.array instead of memoryview
* cleaner conditional, add comment
* Test on #2396: bug in Doc.get_lca_matrix()
* reimplementation of Doc.get_lca_matrix(), (closes#2396)
* reimplement Span.get_lca_matrix(), and call it from Doc.get_lca_matrix()
* tests Span.get_lca_matrix() as well as Doc.get_lca_matrix()
* implement _get_lca_matrix as a helper function in doc.pyx; call it from Doc.get_lca_matrix and Span.get_lca_matrix
* use memory view instead of np.ndarray in _get_lca_matrix (faster)
* fix bug when calling Span.get_lca_matrix; return lca matrix as np.array instead of memoryview
* cleaner conditional, add comment
* Add failing test for matcher bug #3009
* Deduplicate matches from Matcher
* Update matcher ? quantifier test
* Fix bug with ? quantifier in Matcher
The ? quantifier indicates a token may occur zero or one times. If the
token pattern fit, the matcher would fail to consider valid matches
where the token pattern did not fit. Consider a simple regex like:
.?b
If we have the string 'b', the .? part will fit --- but then the 'b' in
the pattern will not fit, leaving us with no match. The same bug left us
with too few matches in some cases. For instance, consider:
.?.?
If we have a string of length two, like 'ab', we actually have three
possible matches here: [a, b, ab]. We were only recovering 'ab'. This
should now be fixed. Note that the fix also uncovered another bug, where
we weren't deduplicating the matches. There are actually two ways we
might match 'a' and two ways we might match 'b': as the second token of the pattern,
or as the first token of the pattern. This ambiguity is spurious, so we
need to deduplicate.
Closes#2464 and #3009
* Fix Python2
* Remove check for overwritten factory
This needs to be handled differently – on first initialization, a new factory will be added and any subsequent initializations will trigger this warning, even if it's a new entry point that doesn't overwrite a built-in.
* Add helper to only load specific entry point
Useful for loading languages via entry points, so that they can be lazy-loaded. Otherwise, all entry point languages would have to be loaded upfront.
* Check entry points for custom languages
## Description
- [x] fix auto-detection of Jupyter notebooks (even if `jupyter=True` isn't set)
- [x] add `displacy.set_render_wrapper` method to define a custom function called around the HTML markup generated in all calls to `displacy.render` (can be used to allow custom integrations, callbacks and page formatting)
- [x] add option to customise host for web server
- [x] show warning if `displacy.serve` is called from within Jupyter notebooks
- [x] move error message to `spacy.errors.Errors`.
### 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.
The output weights often return negative scores for classes, especially
via the bias terms. This means that when we add a new class, we can't
rely on just zeroing the weights, or we'll end up with positive
predictions for those labels.
To solve this, we use nan values as the initial weights for new labels.
This prevents them from ever coming out on top. During backprop, we
replace the nan values with the minimum assigned score, so that we're
still able to learn these classes.
After creating a component, the `.model` attribute is left with the value `True`, to indicate it should be created later during `from_disk()`, `from_bytes()` or `begin_training()`. This had led to confusing errors if you try to use the component without initializing the model.
To fix this, we add a method `require_model()` to the `Pipe` base class. The `require_model()` method needs to be called at the start of the `.predict()` and `.update()` methods of the components. It raises a `ValueError` if the model is not initialized. An error message has been added to `spacy.errors`.
* issue #3012: add test
* add contributor aggreement
* Make test work without models and fix typos
ten.pos_ instead of ten.orth_ and comparison against "10" instead of integer 10
I have added alpha support for the Tagalog language from the Philippines. It is the basis for the country's national language Filipino. I have heavily based the format to the EN and ES languages.
I have provided several words in the lemmatizer lookup table, added stop words from a source, translated numeric words to its Tagalog counterpart, added some tokenizer exceptions, and kept the tag map the same as the English language.
While the alpha language passed the preliminary testing that you provided, I think it needs more data to be useful for most cases.
* Added alpha support for Tagalog language
* Edited contributor template
* Included SCA; Reverted templates
* Fixed SCA template
* Fixed changes in SCA template
* Try to implement cosine loss
This one seems to be correct? Still unsure, but it performs okay
* Try to implement the von Mises-Fisher loss
This one's definitely not right yet.
The new spacy pretrain command implemented BERT/ULMFit/etc-like transfer learning, using our Language Modelling with Approximate Outputs version of BERT's cloze task. Pretraining is convenient, but in some ways it's a bit of a strange solution. All we're doing is initialising the weights. At the same time, we're putting a lot of work into our optimisation so that it's less sensitive to initial conditions, and more likely to find good optima. I discuss this a bit in the pseudo-rehearsal blog post: https://explosion.ai/blog/pseudo-rehearsal-catastrophic-forgetting
Support semi-supervised learning in spacy train
One obvious way to improve these pretraining methods is to do multi-task learning, instead of just transfer learning. This has been shown to work very well: https://arxiv.org/pdf/1809.08370.pdf . This patch makes it easy to do this sort of thing.
Add a new argument to spacy train, --raw-text. This takes a jsonl file with unlabelled data that can be used in arbitrary ways to do semi-supervised learning.
Add a new method to the Language class and to pipeline components, .rehearse(). This is like .update(), but doesn't expect GoldParse objects. It takes a batch of Doc objects, and performs an update on some semi-supervised objective.
Move the BERT-LMAO objective out from spacy/cli/pretrain.py into spacy/_ml.py, so we can create a new pipeline component, ClozeMultitask. This can be specified as a parser or NER multitask in the spacy train command. Example usage:
python -m spacy train en ./tmp ~/data/en-core-web/train/nw.json ~/data/en-core-web/dev/nw.json --pipeline parser --raw-textt ~/data/unlabelled/reddit-100k.jsonl --vectors en_vectors_web_lg --parser-multitasks cloze
Implement rehearsal methods for pipeline components
The new --raw-text argument and nlp.rehearse() method also gives us a good place to implement the the idea in the pseudo-rehearsal blog post in the parser. This works as follows:
Add a new nlp.resume_training() method. This allocates copies of pre-trained models in the pipeline, setting things up for the rehearsal updates. It also returns an optimizer object. This also greatly reduces confusion around the nlp.begin_training() method, which randomises the weights, making it not suitable for adding new labels or otherwise fine-tuning a pre-trained model.
Implement rehearsal updates on the Parser class, making it available for the dependency parser and NER. During rehearsal, the initial model is used to supervise the model being trained. The current model is asked to match the predictions of the initial model on some data. This minimises catastrophic forgetting, by keeping the model's predictions close to the original. See the blog post for details.
Implement rehearsal updates for tagger
Implement rehearsal updates for text categoriz
Currently the TextCategorizer defaults to a fairly complicated model, designed partly around the active learning requirements of Prodigy. The model's a bit slow, and not very GPU-friendly.
This patch implements a straightforward CNN model that still performs pretty well. The replacement model also makes it easy to use the LMAO pretraining, since most of the parameters are in the CNN.
The replacement model has a flag to specify whether labels are mutually exclusive, which defaults to True. This has been a common problem with the text classifier. We'll also now be able to support adding labels to pretrained models again.
Resolves#2934, #2756, #1798, #1748.
