0.8.11

Fixed

• Fixed CRF serialization bug

0.8.10

Fixed

Bug with CRF serialization

0.8.9

Fixed

• Fixed the training failure when calling SnipsNLUEngine.get_fitter_tagger(dataset, intent)

0.8.8

Added

• Added integration tests to ensure that parsing performances are kept at a minimum required level
• Support for Korean

Fixed

• Fixed a bug on builtin entity labeling
• Fixed the bug where Ruslting features were not used in the CRF tagging

0.8.7

Added

There is now a __model_version__ package attribute that allows to check model compatibility

0.8.6

Changed

• Improved validation and formatting of the input dataset
• Updated the computation of shape feature to match the Rust NLU

0.8.5

Fixed

bugs in features computation

0.8.4

Fixed

Bug in regular expression patterns that resulted in inconsistent matches.

0.8.2

Fixed

• fixed the RegexIntentParser to be robust to ignored characters (white spaces, commas, etc...) in the beginning and at the end of the sentence. Currently such characters were ignored but only when in the middle of the sentence

0.8.0

Added

• Improved the robustness of the CRF parsing by smartly capitalising entities. Only entities that have been seen as capitalised in the training data will be randomly capitalised.

Changed

• Replaced duckling by rustling
• Updated builtin entities ontology version. Moved to the builtin_entities_ontology==0.2.3 version of the python library which corresponds to the 0.1.3 version of the entity ontology
• Changed the serialisation of the NLU engine to remove pickle from it

Fixed

• Fixed regex escaping
• Lower tokens before stemming
• Fixed the SnipsNLUEngine._parse(text, intent=None) method to call all the parsers when the intent parameter is given. Previously, only the RegexIntentParser was called if the intent was given
• Made the RegexIntentParser robust to some small variations like double spaces, punctuation, etc...

Removed

• Removed the __eq__ and __neq__ method of the `CRFTagger``