0.7.0

Fixed

• we don't add regexes to the RegexIntentParser anymore when there are enough queries for a given intent (> 50). In this case we rely on the ProbabilisticIntentParser to parse correclty. Before that if you added 1000 queries per intent then 1000 regexes were compiled and kept in memory per intent. This fix will both improve memory footprint and serialization

Changed

• rename the NLUEngine.add_pretrained_model to NLUEngine.add_fitted_tagger

Added

• added a NLUEngine.get_fitted_tagger(dataset, intent) that return a trained CRFGTagger for a given intent. This tagger can then be added to the NLUEngine, like this: tagger = engine.get_fitted_tagger(dataset, "my_intent") and engine.add_fitted_tagger(intent="my_intent", model_data=tagger.to_dict()). Now when you train the NLUEngine make sure that you do not retrain the "my_intent" intent. What you can do is passing the list of the intent you want to train when calling: engine.fit(dataset, intents=[<list the intent you need to retrain>])

0.6.0

Removed

• Removed the NLUENgine.tag method. Tagging is now handled by this library: https://github.com/snipsco/snips-nlu-tagging

Fixed

• Fixed the serialiaztion bug of the SnipsIntentClassfier

0.5.0

Added

• Tests to ensure that the whole pipeline is robust to naughty strings
• Added a NLUEngine.tag(text, intent) method for autotagging. When an intent as less than x queries, the tag method return (by order of priority in case of overlap): the previously seen entities in the NLUEngine intents, the bulitin entities, the result of the intent model trained with x queries. When there are more than x queries, then the autotagging output is the output of the model trained with x queries
• Added the ability to add pretrained intent to an NLUEngine with the NLUEngine.add_pretrained_model(intent, model_data) method
• Added the ability to only train particular intents when fitting a dataset with the intents argument of the NLUEngine.fit(dataset, intents=None) method

Changed

• Changed the serialization of the NLUEngine and most other objects of the lib
• Improved some feature performances by adding caching
• Improved builtin entities handling in regex generation
• snips_nlu_version key is now mandatory in the input dataset

Fixed

• Fixed bug with synonyms in the dataset, see: #224
• Fixed bug with unseen CRF labels at inference time, if some labels were not seen during training, we could ask the CRF probabilities of unseens labels when post processing builtin entities
• Fixed the bug happening with the intent classification feature extraction when the input queries were empty or only contained stop words, leading to an empty vocab for the Featurize.count_vectorizer

Removed

• the force_builtin_entities flag in the NLUEngine.parse method, autotagging is now handle by the NLUEngine.tag method
• removed deep intent support with the rust library, builtin intent are now light intents added from the registry

0.4.1

Fixed

• KeyError with builtins in the dataset

0.4.0

Added

• snips_nlu can now expose the version of the builtin_entities_ontology that it's using through snips_nlu.__builtin_entities_version__
• added test to check that the members of the BuiltInEntities enum class are exactly the members of the builtin_entities_ontology

Changed

• BuiltIn entities in the dataset do not require the following fields anymore: data, automatically_extensible, use_synonyms

0.3.4

Added

• Post processing that enrich parsed slots with additional builtin entities
• New SnipsNLUEngine API that allows to parse and give the intent of an input when it is known, thus skipping the intent classification
• New SnipsNLUEngine API that force the use of builtin entities even when not declared by the user (useful for automatic tagging)

Changed

• Load resources for all languages during the import of the snips_nlu package instead of doing it on-the-fly

Fixed

• Fixed a bug that made the engine crash when used with builtin intents only
• The regex parser now handles correctly intents with no slots

0.3.2

[0.3.2] - 2017-04-28

Added

• expose version number through: snips_nlu.__version__

Changed

• improved tokenization to support keep special tokens like $
• improved CI

Fixed

• Tagging issue when entities contained special characters
• Can now fit an empty dataset. Before when the user add only built in intent, the nlu engine was crashing

0.3.1

[0.3.1] - 2017-04-18

Added

• Korean support
• New duckling entities as model features

Changed

• Improved German and English models
• Improved dataset validation
• Improved the intent classification strategy, so that built-in intents are not mixed up with custom intents

Fixed

• Fixed tagging scheme related issues
• Heavy serialization

Removed

• the internal utterance_text field from the dataset